Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

You’ve supercharged your research process with ACS and Mendeley!

STEP 1:
Click to create an ACS ID

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

MENDELEY PAIRING EXPIRED
Your Mendeley pairing has expired. Please reconnect
ACS Publications. Most Trusted. Most Cited. Most Read
Graph Comparison of Molecular Crystals in Band Gap Prediction Using Neural Networks
More

OR SEARCH CITATIONS

My Activity
Publications

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:ACS Omega 2023, 8, 42, 39481-39489
  • Open Access
Article

Graph Comparison of Molecular Crystals in Band Gap Prediction Using Neural Networks
Click to copy article linkArticle link copied!

  • Takuya Taniguchi*
    Takuya Taniguchi
    Center for Data Science, Waseda University, 1-6-1 Nishiwaseda, Shinjuku-ku, Tokyo 169-8050, Japan
    *Email: [email protected]
    More by Takuya Taniguchi
    Orcidhttps://orcid.org/0000-0002-7885-2962
  • Mayuko Hosokawa
    Mayuko Hosokawa
    Department of Advanced Science and Engineering, Graduate School of Advanced Science and Engineering, Waseda University, 3-4-1 Okubo, Shinjuku-Ku, Tokyo 169-8555, Japan
    More by Mayuko Hosokawa
  • Toru Asahi
    Toru Asahi
    Department of Advanced Science and Engineering, Graduate School of Advanced Science and Engineering, Waseda University, 3-4-1 Okubo, Shinjuku-Ku, Tokyo 169-8555, Japan
    More by Toru Asahi
Open PDFSupporting Information (1)

ACS Omega

Cite this: ACS Omega 2023, 8, 42, 39481–39489
Click to copy citationCitation copied!
https://doi.org/10.1021/acsomega.3c05224
Published October 16, 2023

Copyright © 2023 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Abstract

Click to copy section linkSection link copied!
High Resolution Image
Download MS PowerPoint Slide

In material informatics, the representation of the material structure is fundamentally essential to obtaining better prediction results, and graph representation has attracted much attention in recent years. Molecular crystals can be graphically represented in molecular and crystal representations, but a comparison of which representation is more effective has not been examined. In this study, we compared the prediction accuracy between molecular and crystal graphs for band gap prediction. The results showed that the prediction accuracies using crystal graphs were better than those obtained using molecular graphs. While this result is not surprising, error analysis quantitatively evaluated that the error of the crystal graph was 0.4 times that of the molecular graph with moderate correlation. The novelty of this study lies in the comparison of molecular crystal representations and in the quantitative evaluation of the contribution of crystal structures to the band gap.

This publication is licensed under

CC-BY-NC-ND 4.0 .
  • cc licence
  • by licence
  • nc licence
  • nd licence
Copyright © 2023 The Authors. Published by American Chemical Society

Subjects

what are subjects

Introduction

Click to copy section linkSection link copied!
Material informatics (MI) constitutes a compelling research subject, attracting significant attention across academic and industrial landscapes. (1−3) The application of MI is becoming widespread in various fields, typically in polymers and inorganics. In most cases of MI, a set of structural information and target properties is required as prior knowledge for model supervision. While a target variable can be defined relatively straightforwardly depending on the purposes, structural information can be represented in various ways. For example, conventional representations of molecules are fingerprint vectors, such as extended connectivity fingerprints. (4) Many types of descriptors have been developed, (5) and the choice of the descriptor influences the predictive performance of a task to be solved.
In recent years, graph representation has received a great deal of attention. Molecular structures can be represented as graph data consisting of nodes and edges, and graph neural networks (GNNs) can handle regression and classification tasks. (6−9) Recent studies on inorganic crystals also have successfully used graphs for inorganic crystal structures. (10−14) Graph-based approaches have been applied not only to molecules and bulk inorganic materials but also to metal–organic frameworks, two-dimensional materials, and even molecular dynamics simulations. (15−17) Furthermore, model architectures such as ALIGNN and M3GNet, which incorporate angle information, have been developed to improve prediction accuracy. (13,18)
In contrast to inorganic materials and polymers, MI research on molecular crystals has made little progress. (19−21) This may be due to the lack of databases linking the structure and properties of the molecular crystals. For example, the Cambridge Structural Database (CSD) and Crystallography Open Database (COD) are useful platforms of molecular crystal structures, (22,23) but they do not include information on physicochemical properties. Therefore, constructing structure–property data sets requires human labor to collect physical properties from publications, experiments, and quantum chemical calculations. Olsthoorn et al. constructed a large database of electronic band gaps (N = 12,500) that relates molecular crystal structures to band gaps obtained from quantum chemical calculations. (24) The band gap is the energy difference between the conduction and valence bands and is related to the optoelectronic properties. They also performed a regression analysis using crystal graph and achieved a mean absolute error (MAE) of 0.388 eV by the ensemble model of the smooth overlap of atomic positions (SOAP) (25) and a kind of GNN, SchNet. (6) It is expected that such database construction and MI-aided material screening will be helpful for the development of novel molecular crystals, as evidenced by some recent works. (26−28)
As such, Olsthoorn et al. conducted a valid prior study but did not compare the graph representations. Because molecular crystals are composed of organic molecules, molecular and crystal graph representations are possible (Figure 1). The complexity of the graph also affects prediction performance, and it is crucial to investigate which graph representation has the much better prediction accuracy. Furthermore, the difference between molecular and crystal graphs will allow quantitative evaluation of the influence of the molecular structure and intermolecular interactions on the target properties, which can provide insight into the physicochemical aspect.

Figure 1

Figure 1. Structure–property relationship and graph representations of molecular crystals. Molecular crystals are formed by organic molecules and thus can be represented by both molecular and crystal graphs.

High Resolution Image
Download MS PowerPoint Slide
This study compares the prediction accuracy of molecular and crystal graphs in band gap prediction. Molecular and crystal graphs of varying complexities were used as input, and GNN models of multiple architectures were used as prediction functions. A comparison of regression results shows that crystal graphs give better prediction accuracy than molecular graphs and the contribution of molecular structure and intermolecular interactions is quantitatively evaluated. The novelty of this study lies in the comparison of molecular crystal representations and in the quantitative evaluation of the contribution of crystal structures to the band gap. Since this algorithm can be applied to other properties of molecular crystals, it is expected to contribute to the efficient screening and design of molecular crystals.

Results and Discussion

Click to copy section linkSection link copied!

Data Set and Structure Representations

A data set where the crystal structure and calculated electronic band gap corresponded was obtained from the Organic Materials Database (OMDB). (24) In the data set, there is no distinction between direct and indirect band gaps; that is, the value was defined as a distance between the minimum energy of the lowest conduction band and the maximum energy of the highest valence band independently. As the original data set did not contain simplified molecular input line entry system (SMILES) corresponding to each molecular structure, SMILES data were extracted from the COD database (29) for this study. Data that caused errors in processing SMILES conversion using rdkit package were excluded, yielding a data set of N = 10472, where SMILES, the crystal structure (as COD ID), and the band gap corresponded. The data set was split into train, validation, and test subset at the ratio of 0.8, 0.05, and 0.15, respectively. All subsets have a similar distribution of band gap (Figure 2).

Figure 2

Figure 2. Data distribution of the band gap in the train, validation, and test subsets.

High Resolution Image
Download MS PowerPoint Slide
Molecular and crystal graphs are formed with different complexities. A SMILES-based molecular graph comprises nodes and edges that reflect atoms and chemical bonds (Table S1). Each node has a 121-dimensional representation that includes atomic information, such as the atom number and charge. Each edge has a 2-dimensional vector based on chemical bonds. As SMILES does not explicitly include hydrogen atoms, it is possible to determine whether they are added to molecular graphs. In the molecular graphs without hydrogen atoms (MolGraph), the number of nodes and edges averages 21.8 and 68.3, respectively (Figure 3a). These averages of molecular graphs with hydrogen atoms (MolGraphH) increased nearly twice (Figure 3a). Figure 3b exemplifies the MolGraph and MolGraphH of a molecule.

Figure 3

Figure 3. Molecular and crystal graphs with different complexities. (a) Comparison of the averages of the number of nodes and edges in different molecular and crystal graphs. (b) Example of molecular graphs without and with hydrogen atoms (MolGraph and MolGraphH, respectively). (c) Dependence of the average of the number of edges in crystal graphs by changing the radius (r) and maximum number of neighbors (max_n). The number of nodes does not change by these parameters. (d) Example of the crystal structure (COD ID = 4030612) and corresponding simple, medium, and complicated CrystGraphs.

High Resolution Image
Download MS PowerPoint Slide
Crystal graphs were created based on atomic coordinates in the crystal structures. The main difference of the crystal graphs from molecular graphs is that edges are formed based on distances rather than chemical bonds. When edges are created in the crystal graphs, there are two parameters: radius and the maximum number of neighbors. The radius defines the maximum distance from one atom to the other atoms for sharing the edges. When the atom–atom distance is within the radius, an edge is allowed between these atoms. The maximum number of neighbors defines the maximum number of edges from a node, and the number of edges is limited to the nearest atoms, even when the atom–atom distance is within the radius. Thus, the number of edges in the crystal graphs depends on the radius (r) in the unit of Angstrom and the maximum number of neighbors (max_n). We considered three different crystal graphs: simple (r = 8 Å, max_n = 6), medium (r = 8 Å, max_n = 12), and complicated (r = 4.2 Å, max_n = 1000) graphs by changing these parameters (Figure 3a,c). Hereafter, they are called simple, medium, and complicated CrystGraphs. The average of the number of edges almost doubled from simple to medium CrystGraph and from medium to complicated CrystGraph (Figure 3a,c). This complexity graph could be achieved with other combinations of r and max_n, but we chose these combinations to vary the number of edges. The number of edges corresponds to the graph complexity, indicating the number of intermolecular interactions considered in graph convolutions. The crystal structure (COD number: 4030612), whose molecule is shown in Figure 3b, exemplifies the difference in graph complexity (Figure 3d). In the crystal structure, C–Br is the longest chemical bond (1.99 Å), and O···H is a short intermolecular interaction (2.69 Å) (Figure 3d). Simple CrystGraph captures the longest chemical bond and the short interaction by forming edges as indicated by solid and dotted arrows, while other longer intermolecular interactions are not considered. Medium CrystGraph captures more intermolecular interactions by a greater number of edges, and complicated CrystGraph contains much more. The difference in the graph complexity on the prediction performance is shown in the following section.

Regression on Band Gap

Molecular and crystal graphs with different complexities were input into the GNN models CGCNN, SchNet, and MEGNet. (6−8) The difference between these models relies on the graph convolution layer, which aggregates and combines node and edge features without changing the graph adjacency (Methods section). The node and edge features after convolution layers were converted to a vector by a readout operation, and the vector was then inputted to the fully connected layers to output a prediction value. The predictive performance was evaluated using the test metrics R2, root-mean-square error (RMSE), and mean absolute error (MAE). The hyperparameters of all of the models were optimized with respect to each input representation (Table S2 and Figures S1–S15).
Table 1 presents the metrics for the test subset using different combinations of graphs and models. The MEGNet models performed better than SchNet and CGCNN in all combinations using molecular and crystal graphs (Table 1). Thus, the difference in graph inputs on test metrics is discussed based on the results of the MEGNet model.
Table 1. Regression Results of Band Gap Analysis Using Molecular and Crystal Graphsa
representationGNNR2MAERMSE
MolGraphSchNet0.732 (0.004)0.401 (0.002)0.527 (0.004)
MEGNet0.776 (0.006)0.360 (0.003)0.482 (0.007)
CGCNN0.754 (0.011)0.375 (0.011)0.505 (0.011)
MolGraphHSchNet0.693 (0.006)0.427 (0.005)0.564 (0.006)
MEGNet0.758 (0.014)0.379 (0.013)0.501 (0.015)
CGCNN0.750 (0.007)0.385 (0.007)0.509 (0.007)
simple CrystGraphSchNet0.862 (0.008)0.279 (0.009)0.378 (0.011)
MEGNet0.892 (0.006)0.249 (0.014)0.335 (0.010)
CGCNN0.858 (0.007)0.283 (0.005)0.383 (0.009)
medium CrystGraphSchNet0.833 (0.055)0.308 (0.055)0.412 (0.066)
MEGNet0.895 (0.004)0.240 (0.005)0.329 (0.006)
CGCNN0.856 (0.010)0.280 (0.014)0.386 (0.013)
complicated CrystGraphSchNet0.810 (0.084)0.331 (0.072)0.437 (0.094)
MEGNet0.879 (0.007)0.259 (0.006)0.354 (0.010)
CGCNN0.870 (0.004)0.277 (0.004)0.367 (0.006)
a

Each value is the average, and the bracket represents standard deviation.

The best representation was medium CrystGraph. Simple CrystGraph afforded nearly high performance with that of medium CrystGraph, and complicated CrystGraph was worse than those CrystGraphs (Table 1). Both molecular graphs, MolGraph and MolGraphH, afforded metrics that were worse than those of CrystGraphs. The better representation was MolGraph rather than MolGraphH. In both cases of molecular and crystal graphs, the most complicated graph did not afford the highest metrics, showing that there is a suitable complexity of representation. In the case of the average model, which represents the reference metrics when assuming there is no relationship between structure and band gap, MAE and RMSE were 0.794 and 1.018 eV, respectively. It is estimated that the combination of MolGraph and the MEGNet model reduced 0.63 eV in MAE by considering molecular structure and that the combination of medium CrystGraph and the MEGNet model further reduced 0.12 eV in MAE by considering intermolecular interactions. Thus, the error using medium CyrstGraph was 0.67 times that of MolGraph. It is reasonable that the intermolecular interactions are weak in many molecular crystals and thus have a moderate effect on the electronic orbitals by crystal packings, and the molecular structure alone can greatly reduce the prediction error. The combination of medium CrystGraph and MEGNet outperformed the ensemble model of SOAP and SchNet developed by Olsthoorn et al. (MAE = 0.388 eV, RMSE = 0.519 eV). (24) This result agreed with SchNet yielding worse metrics in our regression (Table 1). Although this may be due to the reduced number of data, the tuning of crystal graph hyperparameters and the utilization of the MEGNet model afforded better predictive performance. Then, we also performed the regression on the recent architecture, ALIGNN, using the same training data set and found that the MAE for the same test data set was 0.221 eV, which was slightly better than the MEGNet model. Based on this result, it is expected that the latest models such as ALIGNN and M3GNet would result in a smaller prediction error for molecular crystals. However, since it is difficult to input molecular graph used in the current work to these models, and since we want to compare the molecular graph with the crystal graph, we will use the MEGNet model in the following discussion.
Since MolGraph and medium CrystGraph afforded better test metrics as molecular and crystal representations, we performed error analysis of them where error was defined as predicted value minus observed value. First, the observed-predicted plot showed how predictions distributed (Figure 4a,b). The observed-predicted plot of medium CrystGraph presented a smaller variance from the reference line than MolGraph, which is consistent with better test metrics (Table 1). Both training and test MAEs of medium CrystGraph were lower than those of MolGraph, indicating proper representation learning of crystal structures. Even though, predictions of test data ranging from 1 to 5 eV, which was the main fraction of data set (Figure 2), have some variance from the reference line compared to the training data set. Error analysis of MolGraph showed the negative tendency of error (Figure 4c), meaning that MolGraph tends to overestimate for smaller band gaps and to underestimate for larger band gaps. This tendency should indicate there remains bias by missing information on intermolecular interactions. Medium CrystGraph resolved such a negative tendency (Figure 4d). This reason should be the capture of more information on CrystGraph. Crystal graph captures both intermolecular interactions and intramolecular interactions in the edge information, which may have reduced the prediction error. However, we cannot deny that other unidentified factors may influence the prediction.

Figure 4

Figure 4. Predictive performance using molecular and crystal graphs. (a, b) Observed-predicted plots of training and test data set represented by (a) MolGraph and (b) medium CrystGraph. The dashed line is the reference line when predictions perfectly match with DFT data. (c, d) Error analysis of test data set represented by (c) MolGraph and (d) medium CrystGraph. Error was defined as predicted value minus observed value. (e) Error-error plot of MolGraph and medium CrystGraph. The dashed line is the linear regression. The area highlighted in gray is the 95% prediction interval. (f) Dependence of training size on test MAE.

High Resolution Image
Download MS PowerPoint Slide
We then visualized an error–error plot to find the error relationship between MolGraph and medium CrystGraph (Figure 4e). The error–error plot showed roughly linear tendency with a slope of 0.41 and Pearson’s correlation coefficient of 0.62. This result means that medium CrystGraph afforded about 0.4 times smaller error than MolGraph with a mild correlation. This value is smaller than the value seen in the MAE because the MAE averages the absolute errors of all plots, whereas the error–error plot shows the relationship between the medium CrystGraph and MolGraph at the same data point. As a result of direct comparison, we can say that the medium CrystGraph has an error reduction of 0.4 times that of the MolGraph. However, since the correlation coefficient is not high and some predictions do not follow this trend, the MAE reduction effect is considered to have been 0.67 times. We did not identify the relationship between error distribution and structural features even though we checked the crystal structures manually of some outliers. If we find some relationship, we can speculate the strength and weakness of the crystal graph and possibly improve the convolution architecture of GNN, but we did not find so far.
The dependence of the training size on the test MAE was also compared between MolGraph and medium CrystGraph (Figure 4f). MolGraph converged faster than medium CrystGraph probably due to the simpler representation of graph. Medium CrystGraph reached a smaller MAE than MolGarph. This result showed the difference of convergence speed due to the difference of graph complexity and the sufficiency of training when sufficient training size was provided.

Generalization Ability and Large Screening

Subsequently, we confirmed the generalization ability outside the above data set. We collected 21 data from publications where SMILES, crystal structure, and calculated electronic band gap were reported (Table S3). The distribution of the number of constituent elements in newly collected data is similar to the initial data set for model training and testing (Figure S16). In addition, when crystal structures were represented by SOAP and then embedded by t-Distributed Stochastic Neighbor Embedding (t-SNE), the newly collected test data, of which there are only 21, is not widely distributed throughout the space, but it is found to be close to the initial data set (Figure S17). Using the trained MEGNet model, medium CrystGraph yielded better metrics (RMSE = 0.46, MAE = 0.34) than MolGraph (RMSE = 0.59, MAE = 0.43), confirming the better generalization ability of the combination of medium CrystGraph and the MEGNet model (Figure 5a). For further comparison, we calculated predicted values using the ensemble model of SchNet and SOAP developed by Olsthoom et al. on their web application. (23) The metrics of the predictions were comparable with those of MolGraph and worse than those of medium CrystGraph (Figure 5a). The observed-predicted plot showed how predictions are made for each data, and light green plots of medium CrystGraph were tented to be closer to the reference line (Figure 5b), confirming the better generalization. This generalization is consistent with the regression results, as explained in the previous section.

Figure 5

Figure 5. Generalization performance. (a) Comparison of the test metrics between MolGraph and medium CrystGraph developed by this work and the ensemble model developed in the literature. (b) Observed-predicted plot of the generalization data set.

High Resolution Image
Download MS PowerPoint Slide
We also tested the prediction ability for polymorphic crystals using medium CrystGraph, whose compounds are not included in the training data set (Table S4). There have been 8 polymorphs of 5-methyl-2-[(2-nitrophenyl)amino]-3-thiophenecarbonitrile, commonly known as ROY, and their electronic band gaps have been calculated to be 1.12–1.84 eV depending on crystal structures. (30) In another case, band gaps of two polymorphs of 9,10-bis((E)-2-(pyridin-4-yl)vinyl)anthracene (BP4VA) have been reported to be 1.58 and 1.26 eV in the literature. (31) Using the trained model, medium CrystGraph of ROY afforded 0.40 eV MAE, and BP4PV resulted in 0.52 eV MAE (Figure 6). The result, a smaller error of ROY than BP4VA, may be attributed to the difference in a polymorphic manner. In the ROY crystal, molecular conformations differ from each polymorph due to the molecular flexibility, while the stacking manner of relatively rigid molecules differs in the polymorphs of BP4VA. Such polymorphic differences may have resulted in different MAEs. In addition, these values were worse than the previous generalization test, and this result suggests that medium Crystal may have a weakness in the distinction of polymorphs. Error plot also presented a negative tendency, suggesting insufficient learning. This result may originate from the deficiency of training data of polymorphic crystals because the original OMDB database does not contain polymorphic crystals. Such model extensions to the polymorphs should be tackled in the future.

Figure 6

Figure 6. Band gap prediction of polymorphic crystals, ROY and BP4VA, using a medium CrystGraph.

High Resolution Image
Download MS PowerPoint Slide
Since medium CrystGraph afforded better generalization ability, we screened a large data set of crystal structures downloaded from the Cambridge Structural Database (CSD). It is important to find molecular crystals with high and low band gaps by screening for efficient material design. The number of downloaded crystal structures is 111936, identified with the unique CSD reference code. The distribution of the data set for screening was similar to that of the initial data set, after their structures encoded by SOAP descriptor were embedded in 2D plot (Figure S18). The band gap, predicted by medium CrystGraph on the trained MEGNet model, is distributed with a mean of 2.86 eV, a maximum of 7.39 eV, and a minimum of 0.13 eV, respectively (Figure 7a). This distribution is similar to the data distribution curated from the OMDB (Figure 2).

Figure 7

Figure 7. Comprehensive screening of band gap of crystal structures from CSD. (a) Histogram of band gap predicted by the trained model of medium CrystGraph. (b) Top-3 crystal structures with the smallest and largest predicted band gap. CSD reference code, space group, and predicted band gap are also shown.

High Resolution Image
Download MS PowerPoint Slide
Some crystal structures were visualized to find structural features from the screening. The smallest predictions were found in compounds with sulfur atoms, as exemplified by the top-3 smallest predictions (Figure 7b). Two of the three structures consist of thiophene-based compounds, and the smallest prediction is consistent with that thiophene-based compounds such as tetrathiafulvalene (TTF) salt are known as conductive materials. (32) The largest predictions were found in the hydrocarbon and similar compounds (Figure 7b). Because the crystal of propane has the highest band gap (8.54 eV) in the original data set, the inferred structures reflected the trend of a high band gap. The inferred band gap in the large screening was almost consistent with the known structural features, and therefore, we successfully screened potential crystal structures with varying band gaps. All inferred band gaps were available at https://github.com/takuyhaa/OrgCrystGNN/.

Conclusions

Click to copy section linkSection link copied!
In summary, this work compares graph representations of molecular crystals for the exemplified task of band gap prediction. The representation of medium CrystGraph afforded the best regression metrics compared with molecular graphs and other crystal graphs with different complexities. While this result is not surprising, it was beneficial to find that medium CrystGraph reduced error by about 0.4 times with a moderate correlation coefficient than MolGraph. A generalization test using manually collected data validated better performance using medium CrystGraph than MolGraph and the ensemble model developed in the literature. The band gap of large data set downloaded from CSD was screened, and the reasonable inference result was obtained, identifying potential crystals with lowest and highest band gaps. The novelty of this work should lie in clarifying the relative effect of crystal graph over molecule graph through the representation comparison. This workflow can be applied to other properties, potentially contributing to the efficient screening and design of functional molecular crystals.

Methods

Click to copy section linkSection link copied!

Data Collecting

The original data set was obtained from the Organic Materials Database (OMDB). The data set contains 12500 data, where the crystal structure and band gap corresponded. The crystal structures were assigned as the ID number of the Crystallography Open Database (COD). The DFT calculation has been performed using the projector augmented wave method (33,34) implemented in the Vienna Ab initio Simulation Package (VASP). The exchange-correlation function was approximated by the generalized gradient approximation (GGA) according to Perdew, Burke, and Ernzerhof. (35) The band gap was calculated after each crystal structure was geometrically optimized. Because this data set did not contain SMILES, we independently downloaded a list of SMILES with COD ID. The SMILES was added to the data set when the COD ID was matched. Additionally, the data set was modified to exclude data that arise execution errors in the conversion from SMILES to mol object using rdkit package. This modification finally afforded a data set of N = 10472, where SMILES, crystal structure, and band gap correspond. For the generalization test, we collected 21 data from published works, where SMILES and CCDC reference codes and band gaps were reported (Table S3). For large screening, we downloaded the cif files from Cambridge Structural Database (CSD) by the following conditions: only organic, calculated density larger than zero, temperature is not none, R factor less than 0.1, no disorder, and atmospheric pressure. The downloaded data contained various measurement temperatures, and thus, data were limited to data measured in the range of 273–313 K. This curation resulted in data size of 111936 corresponded to unique CSD reference codes.

Molecular and Crystal Representations

For molecular graphs (MolGraph and MolGraphH), SMILES was converted to mol object using the rdkit package, and node and edge features were stored as graph data using PyTorch and PyTorch-Geometric. For crystal graphs (simple, medium, and complicated CrystGraphs), the atomic coordinates in the crystal structures were obtained from cif or json file and then converted to graph data using PyTorch and PyTorch-Geometric.

Graph Neural Networks

The GNN models (CGCNN, SchNet, and MEGNet) were originally developed in independent studies. (6−8) These GNN models commonly consist of convolution layers, readout operation, and a fully connected neural network. The difference in GNN models relies solely on convolutional operations that update node and edge features without changing graph adjacency.
In the SchNet, (6) the convolution operation is
xi′=∑j∈N(i)xj⊙hΘ(exp(−γ(dj,i−μ)))
where xi′ is the aggregated feature added to node feature xi, xj is the adjacency node feature of node i, hΘ denotes a multiple linear perceptron (MLP), dj,i denotes the interatomic distances between atoms, μ is values for Gaussian expansion, and γ is a coefficient. In this work, we set 0 ≤ μ ≤ 1 for equally spaced sampling (n = 50) and γ = −12.5.
In the CGCNN model, (7) the convolutional operation is
xi′=xi+∑j∈N(i)σ(zi,jWf+bf)⊙g(zi,jWs+bs)
where zi,j = [xi,xj,ei,j] denotes the concatenation of central node features, neighboring node features and edge features. Here, ei,j = exp(−γ(dj,i – μ)). In addition, σ and g denote the sigmoid and softplus functions, respectively. W and b represent the weight and bias matrices, respectively.
In the MEGNet model, (8) the convolution operation is
ei,j′=hΘe(xi⊕xj⊕ei,j)
xi′=hΘv((1N(i)∑j∈N(i)ei,j)⊕xi)
where hΘe and hΘv are the edge and node update functions of the MLP, respectively. Two dense layers were added at the beginning of each convolution layer to preprocess the inputs. The update operation is performed in the order of the edges, nodes, and global attributes. These GNN models have the following hyperparameters: the dimension of the dense layer for preprocessing in convolution, the dimension of the update function (MLP), the dimension of the MLP after the readout operation, the number of convolution layers, the number of dense layers after the readout, the pooling operation, and the learning rate. The batch size and number of epochs were fixed to 64 and 200, respectively. Hyperparameter optimization was performed using Optuna in a manually defined search space (Table S2). GNN models were trained to minimize L1 loss on a Windows computer equipped with GPUs (NVIDIA RTX A6000). The codes were written based on PyTorch and PyTorch-Geometric libraries. Codes are available at https://github.com/takuyhaa/OrgCrystGNN/.
The model training was repeated three times, and the test metrics were calculated by averaging three results. For the dependency of training size on test metrics, three different training data sets with a fixed data size were created by different random state and three trainings were performed on a training data set. In all cases, the same test data set was used for evaluation.

Supporting Information

Click to copy section linkSection link copied!

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.3c05224.

  • Hyperparameters, data set for generalization test, data distribution (PDF)

Terms & Conditions

Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.

Author Information

Click to copy section linkSection link copied!
  • Corresponding Author
  • Authors
    • Mayuko Hosokawa - Department of Advanced Science and Engineering, Graduate School of Advanced Science and Engineering, Waseda University, 3-4-1 Okubo, Shinjuku-Ku, Tokyo 169-8555, Japan
    • Toru Asahi - Department of Advanced Science and Engineering, Graduate School of Advanced Science and Engineering, Waseda University, 3-4-1 Okubo, Shinjuku-Ku, Tokyo 169-8555, Japan
  • Author Contributions

    T.T. contributed to conceptualization, investigation, software, formal analysis, visualization, writing – original draft, writing – review and editing, funding acquisition, project administration, supervision. M.H. contributed to data curation. T.A. contributed to supervision for M.H.

  • Notes
    The authors declare no competing financial interest.

Acknowledgments

Click to copy section linkSection link copied!

This work was partly executed under the cooperation of organization between Waseda University and ENEOS Corporation. This work was financially supported by the JSPS Grant-in-Aid (22K14747) and Waseda University Grant for Special Research Projects (2020C-530, 2021C-404, and 2022C-313).

References

Click to copy section linkSection link copied!

This article references 35 other publications.

  1. 1
    Ramprasad, R.; Batra, R.; Pilania, G.; Mannodi-Kanakkithodi, A.; Kim, C. Machine learning in materials informatics: recent applications and prospects. npj Comput. Mater. 2017, 3, 54,  DOI: 10.1038/s41524-017-0056-5
    Google Scholar
    There is no corresponding record for this reference.
  2. 2
    Butler, K. T.; Davies, D. W.; Cartwright, H.; Isayev, O.; Walsh, A. Machine learning for molecular and materials science. Nature 2018, 559, 547– 555,  DOI: 10.1038/s41586-018-0337-2
    Google Scholar
    2
    Machine learning for molecular and materials science
    Butler, Keith T.; Davies, Daniel W.; Cartwright, Hugh; Isayev, Olexandr; Walsh, Aron
    Nature (London, United Kingdom) (2018), 559 (7715), 547-555CODEN: NATUAS; ISSN:0028-0836. (Nature Research)
    Here we summarize recent progress in machine learning for the chem. sciences. We outline machine-learning techniques that are suitable for addressing research questions in this domain, as well as future directions for the field. We envisage a future in which the design, synthesis, characterization and application of mols. and materials is accelerated by artificial intelligence.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtl2jt7vL&md5=13d36f27db8d59f558fe28e946b4b009
  3. 3
    Schmidt, J.; Marques, M. R.; Botti, S.; Marques, M. A. Recent advances and applications of machine learning in solid-state materials science. npj Comput. Mater. 2019, 5, 83,  DOI: 10.1038/s41524-019-0221-0
    Google Scholar
    There is no corresponding record for this reference.
  4. 4
    Rogers, D.; Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010, 50, 742– 754,  DOI: 10.1021/ci100050t
    Google Scholar
    4
    Extended-Connectivity Fingerprints
    Rogers, David; Hahn, Mathew
    Journal of Chemical Information and Modeling (2010), 50 (5), 742-754CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Extended-connectivity fingerprints (ECFPs) are a novel class of topol. fingerprints for mol. characterization. Historically, topol. fingerprints were developed for substructure and similarity searching. ECFPs were developed specifically for structure-activity modeling. ECFPs are circular fingerprints with a no. of useful qualities: they can be very rapidly calcd.; they are not predefined and can represent an essentially infinite no. of different mol. features (including stereochem. information); their features represent the presence of particular substructures, allowing easier interpretation of anal. results; and the ECFP algorithm can be tailored to generate different types of circular fingerprints, optimized for different uses. While the use of ECFPs has been widely adopted and validated, a description of their implementation has not previously been presented in the literature.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXlt1Onsbg%253D&md5=cd6c736cd7a3d280b67f5316acce8006
  5. 5
    Musil, F.; Grisafi, A.; Bartók, A. P.; Ortner, C.; Csányi, G.; Ceriotti, M. Physics-inspired structural representations for molecules and materials. Chem. Rev. 2021, 121, 9759– 9815,  DOI: 10.1021/acs.chemrev.1c00021
    Google Scholar
    5
    Physics-inspired structural representations for molecules and materials
    Musil, Felix; Grisafi, Andrea; Bartok, Albert P.; Ortner, Christoph; Csanyi, Gabor; Ceriotti, Michele
    Chemical Reviews (Washington, DC, United States) (2021), 121 (16), 9759-9815CODEN: CHREAY; ISSN:0009-2665. (American Chemical Society)
    A review. The first step in the construction of a regression model or a data-driven anal., aiming to predict or elucidate the relationship between the at. scale structure of matter and its properties, involves transforming the Cartesian coordinates of the atoms into a suitable representation. The development of at.-scale representations has played, and continues to play, a central role in the success of machine-learning methods for chem. and materials science. This review summarizes the current understanding of the nature and characteristics of the most commonly used structural and chem. descriptions of atomistic structures, highlighting the deep underlying connections between different frameworks, and the ideas that lead to computationally efficient and universally applicable models. It emphasizes the link between properties, structures, their phys. chem. and their math. description, provides examples of recent applications to a diverse set of chem. and materials science problems, and outlines the open questions and the most promising research directions in the field.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhs1aisL3J&md5=b75f32ea20fa0d83fb028415f540e1f3
  6. 6
    Schütt, K. T.; Sauceda, H. E.; Kindermans, P. J.; Tkatchenko, A.; Müller, K. R. Schnet–a deep learning architecture for molecules and materials. J. Chem. Phys. 2018, 148, 241722  DOI: 10.1063/1.5019779
    Google Scholar
    6
    SchNet - A deep learning architecture for molecules and materials
    Schuett, K. T.; Sauceda, H. E.; Kindermans, P.-J.; Tkatchenko, A.; Mueller, K.-R.
    Journal of Chemical Physics (2018), 148 (24), 241722/1-241722/11CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
    Deep learning has led to a paradigm shift in artificial intelligence, including web, text, and image search, speech recognition, as well as bioinformatics, with growing impact in chem. physics. Machine learning, in general, and deep learning, in particular, are ideally suitable for representing quantum-mech. interactions, enabling us to model nonlinear potential-energy surfaces or enhancing the exploration of chem. compd. space. Here, we present the deep learning architecture SchNet that is specifically designed to model atomistic systems by making use of continuous-filter convolutional layers. We demonstrate the capabilities of SchNet by accurately predicting a range of properties across chem. space for mols. and materials, where our model learns chem. plausible embeddings of atom types across the periodic table. Finally, we employ SchNet to predict potential-energy surfaces and energy-conserving force fields for mol. dynamics simulations of small mols. and perform an exemplary study on the quantum-mech. properties of C20-fullerene that would have been infeasible with regular ab initio mol. dynamics. (c) 2018 American Institute of Physics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXms1Ggurs%253D&md5=988638d520a423f529a16b35031243aa
  7. 7
    Xie, T.; Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys. Rev. Lett. 2018, 120, 145301  DOI: 10.1103/PhysRevLett.120.145301
    Google Scholar
    7
    Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties
    Xie, Tian; Grossman, Jeffrey C.
    Physical Review Letters (2018), 120 (14), 145301CODEN: PRLTAO; ISSN:1079-7114. (American Physical Society)
    The use of machine learning methods for accelerating the design of cryst. materials usually requires manually constructed feature vectors or complex transformation of atom coordinates to input the crystal structure, which either constrains the model to certain crystal types or makes it difficult to provide chem. insights. Here, we develop a crystal graph convolutional neural networks framework to directly learn material properties from the connection of atoms in the crystal, providing a universal and interpretable representation of cryst. materials. Our method provides a highly accurate prediction of d. functional theory calcd. properties for eight different properties of crystals with various structure types and compns. after being trained with 104 data points. Further, our framework is interpretable because one can ext. the contributions from local chem. environments to global properties. Using an example of perovskites, we show how this information can be utilized to discover empirical rules for materials design.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXltFSnu7c%253D&md5=93beb5675af86cf95e07c82c136f3511
  8. 8
    Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; Ong, S. P. Graph networks as a universal machine learning framework for molecules and crystals. Chem. Mater. 2019, 31, 3564– 3572,  DOI: 10.1021/acs.chemmater.9b01294
    Google Scholar
    8
    Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals
    Chen, Chi; Ye, Weike; Zuo, Yunxing; Zheng, Chen; Ong, Shyue Ping
    Chemistry of Materials (2019), 31 (9), 3564-3572CODEN: CMATEX; ISSN:0897-4756. (American Chemical Society)
    Graph networks are a new machine learning (ML) paradigm that supports both relational reasoning and combinatorial generalization. Here, we develop universal MatErials Graph Network (MEGNet) models for accurate property prediction in both mols. and crystals. We demonstrate that the MEGNet models outperform prior ML models such as the SchNet in 11 out of 13 properties of the QM9 mol. data set. Similarly, we show that MEGNet models trained on ∼60 000 crystals in the Materials Project substantially outperform prior ML models in the prediction of the formation energies, band gaps, and elastic moduli of crystals, achieving better than d. functional theory accuracy over a much larger data set. We present two new strategies to address data limitations common in materials science and chem. First, we demonstrate a phys. intuitive approach to unify four sep. mol. MEGNet models for the internal energy at 0 K and room temp., enthalpy, and Gibbs free energy into a single free energy MEGNet model by incorporating the temp., pressure, and entropy as global state inputs. Second, we show that the learned element embeddings in MEGNet models encode periodic chem. trends and can be transfer-learned from a property model trained on a larger data set (formation energies) to improve property models with smaller amts. of data (band gaps and elastic moduli).
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXntFaqt7g%253D&md5=c924547f17083428802d5698682d5971
  9. 9
    Jiang, Y.; Yang, Z.; Guo, J.; Li, H.; Liu, Y.; Guo, Y.; Li, M.; Pu, X. Coupling complementary strategy to flexible graph neural network for quick discovery of coformer in diverse co-crystal materials. Nat. Commun. 2021, 12, 5950  DOI: 10.1038/s41467-021-26226-7
    Google Scholar
    9
    Coupling complementary strategy to flexible graph neural network for quick discovery of coformer in diverse co-crystal materials
    Jiang, Yuanyuan; Yang, Zongwei; Guo, Jiali; Li, Hongzhen; Liu, Yijing; Guo, Yanzhi; Li, Menglong; Pu, Xuemei
    Nature Communications (2021), 12 (1), 5950CODEN: NCAOBW; ISSN:2041-1723. (Nature Research)
    Cocrystal engineering have been widely applied in pharmaceutical, chem. and material fields. However, how to effectively choose coformer has been a challenging task on expts. Here we develop a graph neural network (GNN) based deep learning framework to quickly predict formation of the cocrystal. In order to capture main driving force to crystn. from 6819 pos. and 1052 neg. samples reported by expts., a feasible GNN framework is explored to integrate important prior knowledge into end-to-end learning on the mol. graph. The model is strongly validated against seven competitive models and three challenging independent test sets involving pharmaceutical cocrystals, π-π cocrystals and energetic cocrystals, exhibiting superior performance with accuracy higher than 96%, confirming its robustness and generalization. Furthermore, one new energetic cocrystal predicted is successfully synthesized, showcasing high potential of the model in practice. All the data and source codes are available at https://github.com/Saoge123/ccgnet for aiding cocrystal community.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXit1aqsbjE&md5=e877fd347df52869466130c176db61ec
  10. 10
    Park, C. W.; Wolverton, C. Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery. Phys. Rev. Mater. 2020, 4, 063801  DOI: 10.1103/PhysRevMaterials.4.063801
    Google Scholar
    10
    Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery
    Park, Cheol Woo; Wolverton, Chris
    Physical Review Materials (2020), 4 (6), 063801CODEN: PRMHBS; ISSN:2475-9953. (American Physical Society)
    The recently proposed crystal graph convolutional neural network (CGCNN) offers a highly versatile and accurate machine learning (ML) framework by learning material properties directly from graphlike representations of crystal structures ("crystal graphs"). Here, we develop an improved variant of the CGCNN model (iCGCNN) that outperforms the original by incorporating information of the Voronoi tessellated crystal structure, explicit three-body correlations of neighboring constituent atoms, and an optimized chem. representation of interat. bonds in the crystal graphs. We demonstrate the accuracy of the improved framework in two distinct illustrations: First, when trained/validated on 180 000/20 000 d. functional theory (DFT) calcd. thermodn. stability entries taken from the Open Quantum Materials Database (OQMD) and evaluated on a sep. test set of 230 000 entries, iCGCNN achieves a predictive accuracy that is significantly improved, i.e., 20% higher than that of the original CGCNN. Second, when used to assist a high-throughput search for materials in the ThCr2Si2 structure-type, iCGCNN exhibited a success rate of 31% which is 155 times higher than an undirected high-throughput search and 2.4 times higher than that of the original CGCNN. Using both CGCNN and iCGCNN, we screened 132 600 compds. with elemental decorations of the ThCr2Si2 prototype crystal structure and identified a total of 97 unique stable compds. by performing 757 DFT calcns., accelerating the computational time of the high-throughput search by a factor of 65. Our results suggest that the iCGCNN can be used to accelerate high-throughput discoveries of new materials by quickly and accurately identifying cryst. compds. with properties of interest.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsFKis7fP&md5=0b5903b05d47535d7c2967b7d9d6fc6b
  11. 11
    Louis, S. Y.; Zhao, Y.; Nasiri, A.; Wang, X.; Song, Y.; Liu, F.; Hu, J. Graph convolutional neural networks with global attention for improved materials property prediction. Phys. Chem. Chem. Phys. 2020, 22, 18141– 18148,  DOI: 10.1039/D0CP01474E
    Google Scholar
    11
    Graph convolutional neural networks with global attention for improved materials property prediction
    Louis, Steph-Yves; Zhao, Yong; Nasiri, Alireza; Wang, Xiran; Song, Yuqi; Liu, Fei; Hu, Jianjun
    Physical Chemistry Chemical Physics (2020), 22 (32), 18141-18148CODEN: PPCPFQ; ISSN:1463-9076. (Royal Society of Chemistry)
    The development of an efficient and powerful machine learning (ML) model for materials property prediction (MPP) remains an important challenge in materials science. While various techniques have been proposed to ext. physicochem. features in MPP, graph neural networks (GNN) have also shown very strong capability in capturing effective features for high-performance MPP. Nevertheless, current GNN models do not effectively differentiate the contributions from different atoms. In this paper we develop a novel graph neural network model called GATGNN for predicting properties of inorg. materials. GATGNN is characterized by its compn. of augmented graph-attention layers (AGAT) and a global attention layer. The application of AGAT layers and global attention layers resp. learn the local relationship among neighboring atoms and overall contribution of the atoms to the material's property; together making our framework achieve considerably better prediction performance on various tested properties. Through extensive expts., we show that our method is able to outperform existing state-of-the-art GNN models while it can also provide a measurable insight into the correlation between the atoms and their material property.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsFSmu7bL&md5=9190661d65053bc86d36c270be20cf0d
  12. 12
    Cheng, J.; Zhang, C.; Dong, L. A geometric-information-enhanced crystal graph network for predicting properties of materials. Commun. Mater. 2021, 2, 92,  DOI: 10.1038/s43246-021-00194-3
    Google Scholar
    12
    A geometric-information-enhanced crystal graph network for predicting properties of materials
    Cheng, Jiucheng; Zhang, Chunkai; Dong, Lifeng
    Communications Materials (2021), 2 (1), 92CODEN: CMOAGE; ISSN:2662-4443. (Nature Portfolio)
    Graph neural networks (GNNs) have been used previously for identifying new cryst. materials. However, geometric structure is not usually taken into consideration, or only partially. Here, we develop a geometric-information-enhanced crystal graph neural network (GeoCGNN) to predict the properties of cryst. materials. By considering the distance vector between each node and its neighbors, our model can learn full topol. and spatial geometric structure information. Furthermore, we incorporate an effective method based on the mixed basis functions to encode the geometric information into our model, which outperforms other GNN methods in a variety of databases. For example, for predicting formation energy our model is 25.6%, 14.3% and 35.7% more accurate than CGCNN, MEGNet and iCGCNN models, resp. For band gap, our model outperforms CGCNN by 27.6% and MEGNet by 12.4%.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XotVCku78%253D&md5=44ae02fdbbde3d1eef3d2e2fe77fb860
  13. 13
    Choudhary, K.; DeCost, B. Atomistic line graph neural network for improved materials property predictions. npj Comput. Mater. 2021, 7, 185,  DOI: 10.1038/s41524-021-00650-1
    Google Scholar
    There is no corresponding record for this reference.
  14. 14
    Pandey, S.; Qu, J.; Stevanović, V.; John, P. S.; Gorai, P. Predicting energy and stability of known and hypothetical crystals using graph neural network. Patterns 2021, 2, 100361  DOI: 10.1016/j.patter.2021.100361
    Google Scholar
    14
    Predicting energy and stability of known and hypothetical crystals using graph neural network
    Pandey Shubham; Stevanovic Vladan; Gorai Prashun; Qu Jiaxing; St John Peter
    Patterns (New York, N.Y.) (2021), 2 (11), 100361 ISSN:.
    The discovery of new inorganic materials in unexplored chemical spaces necessitates calculating total energy quickly and with sufficient accuracy. Machine learning models that provide such a capability for both ground-state (GS) and higher-energy structures would be instrumental in accelerated screening. Here, we demonstrate the importance of a balanced training dataset of GS and higher-energy structures to accurately predict total energies using a generic graph neural network architecture. Using [Formula: see text] 16,500 density functional theory calculations from the National Renewable Energy Laboratory (NREL) Materials Database and [Formula: see text] 11,000 calculations for hypothetical structures as our training database, we demonstrate that our model satisfactorily ranks the structures in the correct order of total energies for a given composition. Furthermore, we present a thorough error analysis to explain failure modes of the model, including both prediction outliers and occasional inconsistencies in the training data. By examining intermediate layers of the model, we analyze how the model represents learned structures and properties.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB2cfntlejtA%253D%253D&md5=fc71858fe7691e1e930bf632ef025975
  15. 15
    Frey, N. C.; Akinwande, D.; Jariwala, D.; Shenoy, V. B. Machine learning-enabled design of point defects in 2d materials for quantum and neuromorphic information processing. ACS Nano 2020, 14, 13406– 13417,  DOI: 10.1021/acsnano.0c05267
    Google Scholar
    15
    Machine Learning-Enabled Design of Point Defects in 2D Materials for Quantum and Neuromorphic Information Processing
    Frey, Nathan C.; Akinwande, Deji; Jariwala, Deep; Shenoy, Vivek B.
    ACS Nano (2020), 14 (10), 13406-13417CODEN: ANCAC3; ISSN:1936-0851. (American Chemical Society)
    Engineered point defects in two-dimensional (2D) materials offer an attractive platform for solid-state devices that exploit tailored optoelectronic, quantum emission, and resistive properties. Naturally occurring defects are also unavoidably important contributors to material properties and performance. The immense variety and complexity of possible defects make it challenging to exptl. control, probe, or understand at.-scale defect-property relationships. Here, we develop an approach based on deep transfer learning, machine learning, and first-principles calcns. to rapidly predict key properties of point defects in 2D materials. We use physics-informed featurization to generate a minimal description of defect structures and present a general picture of defects across materials systems. We identify over one hundred promising, unexplored dopant defect structures in layered metal chalcogenides, hexagonal nitrides, and metal halides. These defects are prime candidates for quantum emission, resistive switching, and neuromorphic computing.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslKqtLfF&md5=1ade77df47fb8faf2dbe58746ccb4788
  16. 16
    Fung, V.; Zhang, J.; Juarez, E.; Sumpter, B. G. Benchmarking graph neural networks for materials chemistry. npj Comput. Mater. 2021, 7, 84,  DOI: 10.1038/s41524-021-00554-0
    Google Scholar
    16
    Benchmarking graph neural networks for materials chemistry
    Fung, Victor; Zhang, Jiaxin; Juarez, Eric; Sumpter, Bobby G.
    npj Computational Materials (2021), 7 (1), 84CODEN: NCMPCS; ISSN:2057-3960. (Nature Research)
    Graph neural networks (GNNs) have received intense interest as a rapidly expanding class of machine learning models remarkably well-suited for materials applications. To date, a no. of successful GNNs have been proposed and demonstrated for systems ranging from crystal stability to electronic property prediction and to surface chem. and heterogeneous catalysis. However, a consistent benchmark of these models remains lacking, hindering the development and consistent evaluation of new models in the materials field. Here, we present a workflow and testing platform, MatDeepLearn, for quickly and reproducibly assessing and comparing GNNs and other machine learning models. We use this platform to optimize and evaluate a selection of top performing GNNs on several representative datasets in computational materials chem. From our investigations we note the importance of hyperparameter selection and find roughly similar performances for the top models once optimized. We identify several strengths in GNNs over conventional models in cases with compositionally diverse datasets and in its overall flexibility with respect to inputs, due to learned rather than defined representations. Meanwhile several weaknesses of GNNs are also obsd. including high data requirements, and suggestions for further improvement for applications in materials chem. are discussed.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFehs7zM&md5=e49c3e0b3e69dda5e5b41a33e3eba307
  17. 17
    Jalali, M.; Tsotsalas, M.; Wöll, C. MOFSocialNet: Exploiting metal-organic framework relationships via social network analysis. Nanomaterials 2022, 12, 704,  DOI: 10.3390/nano12040704
    Google Scholar
    17
    MOFSocialNet: Exploiting Metal-Organic Framework Relationships via Social Network Analysis
    Jalali, Mehrdad; Tsotsalas, Manuel; Woell, Christof
    Nanomaterials (2022), 12 (4), 704CODEN: NANOKO; ISSN:2079-4991. (MDPI AG)
    The no. of metal-org. frameworks (MOF) as well as the no. of applications of this material are growing rapidly. With the no. of characterized compds. exceeding 100,000, manual sorting becomes impossible. At the same time, the increasing computer power and established use of automated machine learning approaches makes data science tools available, that provide an overview of the MOF chem. space and support the selection of suitable MOFs for a desired application. Among the different data science tools, graph theory approaches, where data generated from numerous real-world applications is represented as a graph (network) of interconnected objects, has been widely used in a variety of scientific fields such as social sciences, health informatics, biol. sciences, agricultural sciences and economics. We describe the application of a particular graph theory approach known as social network anal. to MOF materials and highlight the importance of community (group) detection and graph node centrality. In this first application of the social network anal. approach to MOF chem. space, we created MOFSocialNet. This social network is based on the geometrical descriptors of MOFs available in the CoRE-MOFs database. MOFSocialNet can discover communities with similar MOFs structures and identify the most representative MOFs within a given community. In addn., anal. of MOFSocialNet using social network anal. methods can predict MOF properties more accurately than conventional ML tools. The latter advantage is demonstrated for the prediction of gas storage properties, the most important property of these porous reticular networks.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XlslWmtLo%253D&md5=1bb836c9d57b2127944c9e1b8a45050e
  18. 18
    Chen, C.; Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nat. Comput. Sci. 2022, 2, 718– 728,  DOI: 10.1038/s43588-022-00349-3
    Google Scholar
    There is no corresponding record for this reference.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=&md5=9874b665cc7a056b8e2f928dd3112440
  19. 19
    Egorova, O.; Hafizi, R.; Woods, D. C.; Day, G. M. Multifidelity statistical machine learning for molecular crystal structure prediction. J. Phys. Chem. A 2020, 124, 8065– 8078,  DOI: 10.1021/acs.jpca.0c05006
    Google Scholar
    19
    Multifidelity Statistical Machine Learning for Molecular Crystal Structure Prediction
    Egorova, Olga; Hafizi, Roohollah; Woods, David C.; Day, Graeme M.
    Journal of Physical Chemistry A (2020), 124 (39), 8065-8078CODEN: JPCAFH; ISSN:1089-5639. (American Chemical Society)
    Statistical machine learning was applied to predict expensive hybrid functional DFT (PBE0) calcns. of crystal structures using a multifidelity approach to reevaluate the energies of crystal structures predicted with an inexpensive force field. The method uses an autoregressive Gaussian process, making use of less expensive GGA DFT (PBE) calcns. to bridge the gap between the force field and PBE0 energies. The method is benchmarked on the crystal structure landscapes of 3 small, H-bonded org. mols. and produces accurate predictions of energies and crystal structure ranking using small nos. of the most expensive calcns.; the PBE0 energies can be predicted with errors of <1 kJ mol-1 with between 4.2 and 6.8% of the cost of the full calcns. As the developed model is probabilistic, how the uncertainties in predicted energies impact the assessment of the energetic ranking of crystal structures are discussed.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslequ7jL&md5=15ef21ac6c7aafa85f5e4fc3bb3e0fce
  20. 20
    Han, Y.; Ali, I.; Wang, Z.; Cai, J.; Wu, S.; Tang, J.; Zhang, L.; Ren, J.; Xiao, R.; Lu, Q.; Hang, L. Machine learning accelerates quantum mechanics predictions of molecular crystals. Phys. Rep. 2021, 934, 1– 71,  DOI: 10.1016/j.physrep.2021.08.002
    Google Scholar
    20
    Machine learning accelerates quantum mechanics predictions of molecular crystals
    Han, Yanqiang; Ali, Imran; Wang, Zhilong; Cai, Junfei; Wu, Sicheng; Tang, Jiequn; Zhang, Lin; Ren, Jiahao; Xiao, Rui; Lu, Qianqian; Hang, Lei; Luo, Hongyuan; Li, Jinjin
    Physics Reports (2021), 934 (), 1-71CODEN: PRPLCM; ISSN:0370-1573. (Elsevier B.V.)
    A review. Quantum mechanics (QM) approaches (DFT, MP2, CCSD(T), etc.) play an important role in calcg. mols. and crystals with a high accuracy and acceptable efficiency. In recent years, with the development of artificial intelligence technol., machine learning (ML) has played an increasingly essential role in accelerating the QM calcns. and predictions of mol. crystals, as well as the discovery of novel materials. This review provides state-of-the-art information and prospects for QM theories, fragment-based methods and ML methods, as well as their up-to-date applications in predicting small inorg. mols., large drug mols. and relevant mol. crystals. The discussed applications include ML potential energy surface (PES) construction, crystal structure prediction (CSP), chem. reaction prediction and predictions of a series of properties, such as structure, energy, at. force, bond length, chem. shift, supercond., super-hardness, vibrational spectra, phase transition and diagram. This work also reviews software and packages built recently based on ML methods for property predictions and PES constructions in the field of physics and chem. For the three discussed methods, the most time-consuming one is the high-level all-atom QM method, which is capable of describing electronic structures with high accuracy and thus predicts properties that are consistent with the exptl. results. The second one, fragment-based QM method, requires less computational time than all-atom QM, which can accelerate all-atom QM calcns. for large systems by dividing the entire system into subsystems, presenting a considerable efficiency increase. The computational complexities for fragment-based QM and all-atom QM are N - N2 and N5-N7 (N is the size of the system), resp. A well-trained ML model can make the above predictions within seconds while ensuring a high prediction accuracy, where its prediction cost and accuracy are detd. by the training data and the training process. Therefore, it is challenging for ML applications in physics and chem. to generate highly accurate and powerful ML models while ensuring sufficient datasets. This work not only provides an overview of the recent progress in QM theories, fragment-based methods, ML methods and several ML-based software programs and applications on small inorg. mols., large drug mols. and relevant crystals, but also shed light on ML methods in accelerating QM prediction, optimization and novel crystal material design.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFWiurzI&md5=829b4c028e46feb7be228b2062790582
  21. 21
    Ishizaki, K.; Sugimoto, R.; Hagiwara, Y.; Koshima, H.; Taniguchi, T.; Asahi, T. Actuation performance of a photo-bending crystal modeled by machine learning-based regression. CrystEngComm 2021, 23, 5839– 5847,  DOI: 10.1039/D1CE00208B
    Google Scholar
    21
    Actuation performance of a photo-bending crystal modeled by machine learning-based regression
    Ishizaki, Kazuki; Sugimoto, Ryota; Hagiwara, Yuki; Koshima, Hideko; Taniguchi, Takuya; Asahi, Toru
    CrystEngComm (2021), 23 (34), 5839-5847CODEN: CRECF4; ISSN:1466-8033. (Royal Society of Chemistry)
    Photomech. mol. crystals are a family of mech. responsive materials that have been recognized as a novel type of actuator. The actuation properties of photomech. crystals should be characterized. However, deflection and force, which are crucial for actuators, are dependent on exptl. conditions such as light intensity and crystal size, and thus the no. of combinations under different conditions is infinite. This causes difficulty in obtaining the relationship between the exptl. conditions and actuation outputs. To solve this problem, this paper presents a machine learning-based regression approach for modeling the deflection and force of a photo-bending crystal. The deflection and force were exptl. measured for various light intensities and crystal sizes. These time-series data of deflection and force were represented by feature values via exponential fitting for bending and unbending processes. The features of the max. value and speed of deflection and force were analyzed by polynomial regression and variable selection. Through this process, the most fitted models were constructed for deflection and force and most of them were interpreted by material mechanics. This statistical strategy will potentially be applied to control or optimize other functions of stimuli-responsive crystals.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmtVSqu7k%253D&md5=3ac98d93899abba3690d8a0f9114e0b6
  22. 22
    Groom, C. R.; Bruno, I. J.; Lightfoot, M. P.; Ward, S. C. The Cambridge structural database. Acta Crystallogr., Sect. B 2016, 72, 171– 179,  DOI: 10.1107/S2052520616003954
    Google Scholar
    22
    The Cambridge Structural Database
    Groom, Colin R.; Bruno, Ian J.; Lightfoot, Matthew P.; Ward, Suzanna C.
    Acta Crystallographica, Section B: Structural Science, Crystal Engineering and Materials (2016), 72 (2), 171-179CODEN: ACSBDA; ISSN:2052-5206. (International Union of Crystallography)
    The Cambridge Structural Database (CSD) contains a complete record of all published org. and metal-org. small-mol. crystal structures. The database has been in operation for over 50 years and continues to be the primary means of sharing structural chem. data and knowledge across disciplines. As well as structures that are made public to support scientific articles, it includes many structures published directly as CSD Communications. All structures are processed both computationally and by expert structural chem. editors prior to entering the database. A key component of this processing is the reliable assocn. of the chem. identity of the structure studied with the exptl. data. This important step helps ensure that data is widely discoverable and readily reusable. Content is further enriched through selective inclusion of addnl. exptl. data. Entries are available to anyone through free CSD community web services. Linking services developed and maintained by the CCDC, combined with the use of std. identifiers, facilitate discovery from other resources. Data can also be accessed through CCDC and third party software applications and through an application programming interface.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28Xls1Kntro%253D&md5=f9c65ab86fc9db429588c95b0da3f9b2
  23. 23
    Gražulis, S.; Daškevič, A.; Merkys, A.; Chateigner, D.; Lutterotti, L.; Quirós, M.; Serebryanaya, N. R.; Moeck, P.; Downs, R. T.; Bail, A. L. Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration. Nucleic Acids Res. 2012, 40, D420– D427,  DOI: 10.1093/nar/gkr900
    Google Scholar
    23
    Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration
    Grazulis, Saulius; Daskevic, Adriana; Merkys, Andrius; Chateigner, Daniel; Lutterotti, Luca; Quiros, Miguel; Serebryanaya, Nadezhda R.; Moeck, Peter; Downs, Robert T.; Le Bail, Armel
    Nucleic Acids Research (2012), 40 (D1), D420-D427CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    Using an open-access distribution model, the Crystallog. Open Database (COD, http://www.crystallog.net) collects all known small mol. / small to medium sized unit cell' crystal structures and makes them available freely on the Internet. As of today, the COD has aggregated ∼150 000 structures, offering basic search capabilities and the possibility to download the whole database, or parts thereof using a variety of std. open communication protocols. A newly developed website provides capabilities for all registered users to deposit published and so far unpublished structures as personal communications or pre-publication depositions. Such a setup enables extension of the COD database by many users simultaneously. This increases the possibilities for growth of the COD database, and is the first step towards establishing a world wide Internet-based collaborative platform dedicated to the collection and curation of structural knowledge.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs12htbnP&md5=670b6cc1f2758728c303c4ad89e1cc6f
  24. 24
    Olsthoorn, B.; Geilhufe, R. M.; Borysov, S. S.; Balatsky, A. V. Band gap prediction for large organic crystal structures with machine learning. Adv. Quantum Technol. 2019, 2, 1900023  DOI: 10.1002/qute.201900023
    Google Scholar
    There is no corresponding record for this reference.
  25. 25
    Bartók, A. P.; Kondor, R.; Csányi, G. On representing chemical environments. Phys. Rev. B 2013, 87, 184115  DOI: 10.1103/PhysRevB.87.184115
    Google Scholar
    25
    On representing chemical environments
    Bartok, Albert P.; Kondor, Risi; Csanyi, Gabor
    Physical Review B: Condensed Matter and Materials Physics (2013), 87 (18), 184115/1-184115/16CODEN: PRBMDO; ISSN:1098-0121. (American Physical Society)
    We review some recently published methods to represent at. neighborhood environments, and analyze their relative merits in terms of their faithfulness and suitability for fitting potential energy surfaces. The crucial properties that such representations (sometimes called descriptors) must have are differentiability with respect to moving the atoms and invariance to the basic symmetries of physics: rotation, reflection, translation, and permutation of atoms of the same species. We demonstrate that certain widely used descriptors that initially look quite different are specific cases of a general approach, in which a finite set of basis functions with increasing angular wave nos. are used to expand the at. neighborhood d. function. Using the example system of small clusters, we quant. show that this expansion needs to be carried to higher and higher wave nos. as the no. of neighbors increases in order to obtain a faithful representation, and that variants of the descriptors converge at very different rates. We also propose an altogether different approach, called Smooth Overlap of Atomic Positions, that sidesteps these difficulties by directly defining the similarity between any two neighborhood environments, and show that it is still closely connected to the invariant descriptors. We test the performance of the various representations by fitting models to the potential energy surface of small silicon clusters and the bulk crystal.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXpvFClu7Y%253D&md5=f7739275562b8e77d4532f00da8814fb
  26. 26
    Wengert, S.; Csányi, G.; Reuter, K.; Margraf, J. T. Data-efficient machine learning for molecular crystal structure prediction. Chem. Sci. 2021, 12, 4536– 4546,  DOI: 10.1039/D0SC05765G
    Google Scholar
    26
    Data-efficient machine learning for molecular crystal structure prediction
    Wengert, Simon; Csanyi, Gabor; Reuter, Karsten; Margraf, Johannes T.
    Chemical Science (2021), 12 (12), 4536-4546CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    The combination of modern machine learning (ML) approaches with high-quality data from quantum mech. (QM) calcns. can yield models with an unrivalled accuracy/cost ratio. However, such methods are ultimately limited by the computational effort required to produce the ref. data. In particular, ref. calcns. for periodic systems with many atoms can become prohibitively expensive for higher levels of theory. This trade-off is crit. in the context of org. crystal structure prediction (CSP). Here, a data-efficient ML approach would be highly desirable, since screening a huge space of possible polymorphs in a narrow energy range requires the assessment of a large no. of trial structures with high accuracy. In this contribution, we present tailored Δ-ML models that allow screening a wide range of crystal candidates while adequately describing the subtle interplay between intermol. interactions such as H-bonding and many-body dispersion effects. This is achieved by enhancing a physics-based description of long-range interactions at the d. functional tight binding (DFTB) level-for which an efficient implementation is available-with a short-range ML model trained on high-quality first-principles ref. data. The presented workflow is broadly applicable to different mol. materials, without the need for a single periodic calcn. at the ref. level of theory. We show that this even allows the use of wavefunction methods in CSP.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXjslWrtr4%253D&md5=b168772ffa3e022a39d1e57a954804ff
  27. 27
    Takagi, D.; Ishizaki, K.; Asahi, T.; Taniguchi, T. Molecular screening for solid–solid phase transition by machine learning. Digital Discovery 2023, 2, 1126,  DOI: 10.1039/D3DD00034F
    Google Scholar
    27
    Molecular screening for solid-solid phase transitions by machine learning
    Takagi, Daisuke; Ishizaki, Kazuki; Asahi, Toru; Taniguchi, Takuya
    Digital Discovery (2023), 2 (4), 1126-1133CODEN: DDIIAI; ISSN:2635-098X. (Royal Society of Chemistry)
    The solid-solid phase transition in mol. crystals is generally found by chance empirically. In this study, we constructed a machine learning framework to screen mols. that will exhibit solid-solid phase transitions in their cryst. states, based on pos.-unlabeled learning. We trained classification models using the pos. dataset we constructed manually and the unlabeled data extd. from the Cambridge Structural Database. The best classifier works as a suggester, and 9 substances among the suggested 113 mols. were found to exhibit solid-solid phase transitions according to the literature and expts. The finding probability of 8.0% is much higher than the probability of phase transition in the database, suggesting the effectiveness of mol. selection by this workflow. We also found that the mol. structure is weakly related to the transition temp. by regression anal. The findings of this study are useful for designing functional mol. crystals with solid-solid phase transitions.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3sXhvVygurvF&md5=027166e2d42022ecf59c4e8d6846490d
  28. 28
    Musil, F.; De, S.; Yang, J.; Campbell, J. E.; Day, G. M.; Ceriotti, M. Machine learning for the structure–energy–property landscapes of molecular crystals. Chem. Sci. 2018, 9, 1289– 1300,  DOI: 10.1039/C7SC04665K
    Google Scholar
    28
    Machine learning for the structure-energy-property landscapes of molecular crystals
    Musil, Felix; De, Sandip; Yang, Jack; Campbell, Joshua E.; Day, Graeme M.; Ceriotti, Michele
    Chemical Science (2018), 9 (5), 1289-1300CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Mol. crystals play an important role in several fields of science and technol. They frequently crystallize in different polymorphs with substantially different phys. properties. To help guide the synthesis of candidate materials, at.-scale modeling can be used to enumerate the stable polymorphs and to predict their properties, as well as to propose heuristic rules to rationalize the correlations between crystal structure and materials properties. Here we show how a recently-developed machine-learning (ML) framework can be used to achieve inexpensive and accurate predictions of the stability and properties of polymorphs, and a data-driven classification that is less biased and more flexible than typical heuristic rules. We discuss, as examples, the lattice energy and property landscapes of pentacene and two azapentacene isomers that are of interest as org. semiconductor materials. We show that we can est. force field or DFT lattice energies with sub-kJ mol-1 accuracy, using only a few hundred ref. configurations, and reduce by a factor of ten the computational effort needed to predict charge mobility in the crystal structures. The automatic structural classification of the polymorphs reveals a more detailed picture of mol. packing than that provided by conventional heuristics, and helps disentangle the role of hydrogen bonded and π-stacking interactions in detg. mol. self-assembly. This observation demonstrates that ML is not just a black-box scheme to interpolate between ref. calcns., but can also be used as a tool to gain intuitive insights into structure-property relations in mol. crystal engineering.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhvFGnsbfE&md5=48b184722478d1a988a2090f765135e2
  29. 29
    Quirós, M.; Gražulis, S.; Girdzijauskaitė, S.; Merkys, A.; Vaitkus, A. Using SMILES strings for the description of chemical connectivity in the Crystallography Open Database. J. Cheminf. 2018, 10, 23,  DOI: 10.1186/s13321-018-0279-6
    Google Scholar
    29
    Using SMILES strings for the description of chemical connectivity in the Crystallography Open Database
    Quiros, Miguel; Grazulis, Saulius; Girdzijauskaite, Saule; Merkys, Andrius; Vaitkus, Antanas
    Journal of Cheminformatics (2018), 10 (), 23/1-23/17CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    Computer descriptions of chem. mol. connectivity are necessary for searching chem. databases and for predicting chem. properties from mol. structure. In this article, the ongoing work to describe the chem. connectivity of entries contained in the Crystallog. Open Database (COD) in SMILES format is reported. This collection of SMILES is publicly available for chem. (substructure) search or for any other purpose on an open-access basis, as is the COD itself. The conventions that have been followed for the representation of compds. that do not fit into the valence bond theory are outlined for the most frequently found cases. The procedure for getting the SMILES out of the CIF files starts with checking whether the atoms in the asym. unit are a chem. acceptable image of the compd. When they are not (mol. in a symmetry element, disorder, polymeric species, etc.), the previously published cif_mol. program is used to get such image in many cases. The program package Open Babel is then applied to get SMILES strings from the CIF files (either those directly taken from the COD or those produced by cif_mol. when applicable). The results are then checked and/or fixed by a human editor, in a computer-aided task that at present still consumes a great deal of human time. Even if the procedure still needs to be improved to make it more automatic (and hence faster), it has already yielded more than 160,000 curated chem. structures and the purpose of this article is to announce the existence of this work to the chem. community as well as to spread the use of its results.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWiurk%253D&md5=99cd2cb8206426aae7de48ea932fe169
  30. 30
    Feng, X.; Becke, A. D.; Johnson, E. R. Theoretical investigation of polymorph-and coformer-dependent photoluminescence in molecular crystals. CrystEngComm 2021, 23, 4264– 4271,  DOI: 10.1039/D1CE00383F
    Google Scholar
    30
    Theoretical investigation of polymorph- and coformer-dependent photoluminescence in molecular crystals
    Feng, Xibo; Becke, Axel D.; Johnson, Erin R.
    CrystEngComm (2021), 23 (24), 4264-4271CODEN: CRECF4; ISSN:1466-8033. (Royal Society of Chemistry)
    Polymorph- and coformer-dependent photoluminescence (PL) are among the variety of novel solid-state PL phenomena recently obsd. in many mol. crystals. They are of particular research interest due to their direct connections to two heavily investigated topics in crystal engineering: polymorphism and cocrystn. Herein, we apply a novel computational methodol., initially proposed and successfully applied in our previous investigation of piezochromism, to theor. modeling of the polymorph- and coformer-dependent PL in the well-known ROY polymorphs and the recently synthesized 9-acetylanthracene (9-ACA) cocrystals, resp. Our methodol. offers satisfactory prediction of the exptl. obsd. color zoning for the ROY polymorphs and provides good qual. and quant. accuracy for the emission (fluorescence) energies of the 9-ACA cocrystals, although the results in both cases may be adversely affected by delocalization error in the d.-functional methods employed. While the polymorph-dependent PL in ROY is found to be controlled by the intramol. geometry, modeling of the periodic crystal environment is necessary for accurate prediction of the coformer-dependent PL in the 9-ACA cocrystals, which is driven by charge transfer.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhtFSkurrK&md5=4789ac696c9e43357bc15af8988abfed
  31. 31
    Aziz, A.; Sidat, A.; Talati, P.; Crespo-Otero, R. Understanding the solid state luminescence and piezochromic properties in polymorphs of an anthracene derivative. Phys. Chem. Chem. Phys. 2022, 24, 2832– 2842,  DOI: 10.1039/D1CP05192J
    Google Scholar
    31
    Understanding the solid state luminescence and piezochromic properties in polymorphs of an anthracene derivative
    Aziz, Alex; Sidat, Amir; Talati, Priyesh; Crespo-Otero, Rachel
    Physical Chemistry Chemical Physics (2022), 24 (5), 2832-2842CODEN: PPCPFQ; ISSN:1463-9076. (Royal Society of Chemistry)
    Luminescent mol. crystals have gained significant research interest for optoelectronic applications. However, fully understanding their structural and electronic relationships in the condensed phase and under external stimuli remains a significant challenge. Here, piezochromism in the mol. crystal 9,10-bis((E)-2-(pyridin-4-yl)vinyl)anthracene (BP4VA) is studied using a combination of d. functional theory (DFT) and time-dependent TD-DFT. We investigate the effects that mol. packing and geometry have on the electronic and phonon structure and the excited state properties in this archetypal system. We find that the luminescence properties are red-shifted with the transition from a herringbone to a sheet packing arrangement. An almost continuous red-shift in the band gap is found with the application of an external pressure through the enhancement of π-π and CH-π interactions, and is a mechanism in fine tuning an emissive response. The anal. of the phonon structure of the mol. crystal suggests restriction of motion in the herringbone packing arrangement, with motion restricted at higher pressure. This is supported by the Huang-Rhys factors which show a decrease in the reorganisation energy with the application of pressure. Ultimately, a balance between the decrease in reorganisation energies and the increase in exciton coupling will det. whether nonradiative decay is enhanced or decreased with the increase in pressure in these systems.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhsF2ls7g%253D&md5=7bb4c15ceff7e88447a96420c707f5a2
  32. 32
    Odom, S. A.; Caruso, M. M.; Finke, A. D.; Prokup, A. M.; Ritchey, J. A.; Leonard, J. H.; White, S. R.; Sottos, N. R.; Moore, J. S. Restoration of conductivity with TTF-TCNQ charge-transfer salts. Adv. Funct. Mater. 2010, 20, 1721– 1727,  DOI: 10.1002/adfm.201000159
    Google Scholar
    32
    Restoration of Conductivity with TTF-TCNQ Charge-Transfer Salts
    Odom, Susan A.; Caruso, Mary M.; Finke, Aaron D.; Prokup, Alex M.; Ritchey, Joshua A.; Leonard, John H.; White, Scott R.; Sottos, Nancy R.; Moore, Jeffrey S.
    Advanced Functional Materials (2010), 20 (11), 1721-1727CODEN: AFMDC6; ISSN:1616-301X. (Wiley-VCH Verlag GmbH & Co. KGaA)
    The formation of the conductive TTF-TCNQ (tetrathiafulvalene-tetracyanoquinodimethane) charge-transfer salt via rupture of microencapsulated solns. of its individual components is reported. Solns. of TTF and TCNQ in various solvents are sep. incorporated into poly(urea-formaldehyde) core-shell microcapsules. Rupture of a mixt. of TTF-contg. microcapsules and TCNQ-contg. microcapsules gave the cryst. salt, as verified by FTIR spectroscopy and powder x-ray diffraction. Preliminary measurements demonstrate the partial restoration of cond. of severed gold electrodes in the presence of TTF-TCNQ derived in situ. This is the 1st microcapsule system for the restoration of cond. in mech. damaged electronic devices in which the repairing agent is not conductive until its release.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXntVyju7Y%253D&md5=a590d7aeaab369f76599fcd14cf8ab70
  33. 33
    Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 1994, 50, 17953,  DOI: 10.1103/PhysRevB.50.17953
    Google Scholar
    33
    Projector augmented-wave method
    Blochl
    Physical review. B, Condensed matter (1994), 50 (24), 17953-17979 ISSN:0163-1829.
    There is no expanded citation for this reference.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC2sfjslSntA%253D%253D&md5=1853d67af808af2edab58beaab5d3051
  34. 34
    Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 1999, 59, 1758,  DOI: 10.1103/PhysRevB.59.1758
    Google Scholar
    34
    From ultrasoft pseudopotentials to the projector augmented-wave method
    Kresse, G.; Joubert, D.
    Physical Review B: Condensed Matter and Materials Physics (1999), 59 (3), 1758-1775CODEN: PRBMDO; ISSN:0163-1829. (American Physical Society)
    The formal relationship between ultrasoft (US) Vanderbilt-type pseudopotentials and Blochl's projector augmented wave (PAW) method is derived. The total energy functional for US pseudopotentials can be obtained by linearization of two terms in a slightly modified PAW total energy functional. The Hamilton operator, the forces, and the stress tensor are derived for this modified PAW functional. A simple way to implement the PAW method in existing plane-wave codes supporting US pseudopotentials is pointed out. In addn., crit. tests are presented to compare the accuracy and efficiency of the PAW and the US pseudopotential method with relaxed-core all-electron methods. These tests include small mols. (H2, H2O, Li2, N2, F2, BF3, SiF4) and several bulk systems (diamond, Si, V, Li, Ca, CaF2, Fe, Co, Ni). Particular attention is paid to the bulk properties and magnetic energies of Fe, Co, and Ni.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK1MXkt12nug%253D%253D&md5=78a73e92a93f995982fc481715729b14
  35. 35
    Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 1996, 77, 3865,  DOI: 10.1103/PhysRevLett.77.3865
    Google Scholar
    35
    Generalized gradient approximation made simple
    Perdew, John P.; Burke, Kieron; Ernzerhof, Matthias
    Physical Review Letters (1996), 77 (18), 3865-3868CODEN: PRLTAO; ISSN:0031-9007. (American Physical Society)
    Generalized gradient approxns. (GGA's) for the exchange-correlation energy improve upon the local spin d. (LSD) description of atoms, mols., and solids. We present a simple derivation of a simple GGA, in which all parameters (other than those in LSD) are fundamental consts. Only general features of the detailed construction underlying the Perdew-Wang 1991 (PW91) GGA are invoked. Improvements over PW91 include an accurate description of the linear response of the uniform electron gas, correct behavior under uniform scaling, and a smoother potential.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK28XmsVCgsbs%253D&md5=55943538406ee74f93aabdf882cd4630

Cited By

Click to copy section linkSection link copied!

This article is cited by 1 publications.

  1. Xinxin Niu, Qian Zhang, Yanfeng Dang, Wenping Hu, Yajing Sun. MolPackL: Quantification and Interpretation of Intermolecular Interactions Driven by Molecular Packing. Journal of the American Chemical Society 2024, 146 (34) , 24075-24084. https://doi.org/10.1021/jacs.4c08132
Download PDF
Open PDF
Go to ACS Omega
Get e-Alerts
Get e-Alerts

ACS Omega

Cite this: ACS Omega 2023, 8, 42, 39481–39489
Click to copy citationCitation copied!
https://doi.org/10.1021/acsomega.3c05224
Published October 16, 2023

Copyright © 2023 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Article Views

1289

Altmetric

-

Citations

1
Learn about these metrics

Article Views are the COUNTER-compliant sum of full text article downloads since November 2008 (both PDF and HTML) across all institutions and individuals. These metrics are regularly updated to reflect usage leading up to the last few days.

Citations are the number of other articles citing this article, calculated by Crossref and updated daily. Find more information about Crossref citation counts.

The Altmetric Attention Score is a quantitative measure of the attention that a research article has received online. Clicking on the donut icon will load a page at altmetric.com with additional details about the score and the social media presence for the given article. Find more information on the Altmetric Attention Score and how the score is calculated.

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. Structure–property relationship and graph representations of molecular crystals. Molecular crystals are formed by organic molecules and thus can be represented by both molecular and crystal graphs.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Data distribution of the band gap in the train, validation, and test subsets.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Molecular and crystal graphs with different complexities. (a) Comparison of the averages of the number of nodes and edges in different molecular and crystal graphs. (b) Example of molecular graphs without and with hydrogen atoms (MolGraph and MolGraphH, respectively). (c) Dependence of the average of the number of edges in crystal graphs by changing the radius (r) and maximum number of neighbors (max_n). The number of nodes does not change by these parameters. (d) Example of the crystal structure (COD ID = 4030612) and corresponding simple, medium, and complicated CrystGraphs.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Predictive performance using molecular and crystal graphs. (a, b) Observed-predicted plots of training and test data set represented by (a) MolGraph and (b) medium CrystGraph. The dashed line is the reference line when predictions perfectly match with DFT data. (c, d) Error analysis of test data set represented by (c) MolGraph and (d) medium CrystGraph. Error was defined as predicted value minus observed value. (e) Error-error plot of MolGraph and medium CrystGraph. The dashed line is the linear regression. The area highlighted in gray is the 95% prediction interval. (f) Dependence of training size on test MAE.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 5

    Figure 5. Generalization performance. (a) Comparison of the test metrics between MolGraph and medium CrystGraph developed by this work and the ensemble model developed in the literature. (b) Observed-predicted plot of the generalization data set.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 6

    Figure 6. Band gap prediction of polymorphic crystals, ROY and BP4VA, using a medium CrystGraph.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 7

    Figure 7. Comprehensive screening of band gap of crystal structures from CSD. (a) Histogram of band gap predicted by the trained model of medium CrystGraph. (b) Top-3 crystal structures with the smallest and largest predicted band gap. CSD reference code, space group, and predicted band gap are also shown.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 35 other publications.

    1. 1
      Ramprasad, R.; Batra, R.; Pilania, G.; Mannodi-Kanakkithodi, A.; Kim, C. Machine learning in materials informatics: recent applications and prospects. npj Comput. Mater. 2017, 3, 54,  DOI: 10.1038/s41524-017-0056-5
      There is no corresponding record for this reference.
    2. 2
      Butler, K. T.; Davies, D. W.; Cartwright, H.; Isayev, O.; Walsh, A. Machine learning for molecular and materials science. Nature 2018, 559, 547– 555,  DOI: 10.1038/s41586-018-0337-2
      2
      Machine learning for molecular and materials science
      Butler, Keith T.; Davies, Daniel W.; Cartwright, Hugh; Isayev, Olexandr; Walsh, Aron
      Nature (London, United Kingdom) (2018), 559 (7715), 547-555CODEN: NATUAS; ISSN:0028-0836. (Nature Research)
      Here we summarize recent progress in machine learning for the chem. sciences. We outline machine-learning techniques that are suitable for addressing research questions in this domain, as well as future directions for the field. We envisage a future in which the design, synthesis, characterization and application of mols. and materials is accelerated by artificial intelligence.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtl2jt7vL&md5=13d36f27db8d59f558fe28e946b4b009
    3. 3
      Schmidt, J.; Marques, M. R.; Botti, S.; Marques, M. A. Recent advances and applications of machine learning in solid-state materials science. npj Comput. Mater. 2019, 5, 83,  DOI: 10.1038/s41524-019-0221-0
      There is no corresponding record for this reference.
    4. 4
      Rogers, D.; Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010, 50, 742– 754,  DOI: 10.1021/ci100050t
      4
      Extended-Connectivity Fingerprints
      Rogers, David; Hahn, Mathew
      Journal of Chemical Information and Modeling (2010), 50 (5), 742-754CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Extended-connectivity fingerprints (ECFPs) are a novel class of topol. fingerprints for mol. characterization. Historically, topol. fingerprints were developed for substructure and similarity searching. ECFPs were developed specifically for structure-activity modeling. ECFPs are circular fingerprints with a no. of useful qualities: they can be very rapidly calcd.; they are not predefined and can represent an essentially infinite no. of different mol. features (including stereochem. information); their features represent the presence of particular substructures, allowing easier interpretation of anal. results; and the ECFP algorithm can be tailored to generate different types of circular fingerprints, optimized for different uses. While the use of ECFPs has been widely adopted and validated, a description of their implementation has not previously been presented in the literature.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXlt1Onsbg%253D&md5=cd6c736cd7a3d280b67f5316acce8006
    5. 5
      Musil, F.; Grisafi, A.; Bartók, A. P.; Ortner, C.; Csányi, G.; Ceriotti, M. Physics-inspired structural representations for molecules and materials. Chem. Rev. 2021, 121, 9759– 9815,  DOI: 10.1021/acs.chemrev.1c00021
      5
      Physics-inspired structural representations for molecules and materials
      Musil, Felix; Grisafi, Andrea; Bartok, Albert P.; Ortner, Christoph; Csanyi, Gabor; Ceriotti, Michele
      Chemical Reviews (Washington, DC, United States) (2021), 121 (16), 9759-9815CODEN: CHREAY; ISSN:0009-2665. (American Chemical Society)
      A review. The first step in the construction of a regression model or a data-driven anal., aiming to predict or elucidate the relationship between the at. scale structure of matter and its properties, involves transforming the Cartesian coordinates of the atoms into a suitable representation. The development of at.-scale representations has played, and continues to play, a central role in the success of machine-learning methods for chem. and materials science. This review summarizes the current understanding of the nature and characteristics of the most commonly used structural and chem. descriptions of atomistic structures, highlighting the deep underlying connections between different frameworks, and the ideas that lead to computationally efficient and universally applicable models. It emphasizes the link between properties, structures, their phys. chem. and their math. description, provides examples of recent applications to a diverse set of chem. and materials science problems, and outlines the open questions and the most promising research directions in the field.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhs1aisL3J&md5=b75f32ea20fa0d83fb028415f540e1f3
    6. 6
      Schütt, K. T.; Sauceda, H. E.; Kindermans, P. J.; Tkatchenko, A.; Müller, K. R. Schnet–a deep learning architecture for molecules and materials. J. Chem. Phys. 2018, 148, 241722  DOI: 10.1063/1.5019779
      6
      SchNet - A deep learning architecture for molecules and materials
      Schuett, K. T.; Sauceda, H. E.; Kindermans, P.-J.; Tkatchenko, A.; Mueller, K.-R.
      Journal of Chemical Physics (2018), 148 (24), 241722/1-241722/11CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
      Deep learning has led to a paradigm shift in artificial intelligence, including web, text, and image search, speech recognition, as well as bioinformatics, with growing impact in chem. physics. Machine learning, in general, and deep learning, in particular, are ideally suitable for representing quantum-mech. interactions, enabling us to model nonlinear potential-energy surfaces or enhancing the exploration of chem. compd. space. Here, we present the deep learning architecture SchNet that is specifically designed to model atomistic systems by making use of continuous-filter convolutional layers. We demonstrate the capabilities of SchNet by accurately predicting a range of properties across chem. space for mols. and materials, where our model learns chem. plausible embeddings of atom types across the periodic table. Finally, we employ SchNet to predict potential-energy surfaces and energy-conserving force fields for mol. dynamics simulations of small mols. and perform an exemplary study on the quantum-mech. properties of C20-fullerene that would have been infeasible with regular ab initio mol. dynamics. (c) 2018 American Institute of Physics.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXms1Ggurs%253D&md5=988638d520a423f529a16b35031243aa
    7. 7
      Xie, T.; Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Phys. Rev. Lett. 2018, 120, 145301  DOI: 10.1103/PhysRevLett.120.145301
      7
      Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties
      Xie, Tian; Grossman, Jeffrey C.
      Physical Review Letters (2018), 120 (14), 145301CODEN: PRLTAO; ISSN:1079-7114. (American Physical Society)
      The use of machine learning methods for accelerating the design of cryst. materials usually requires manually constructed feature vectors or complex transformation of atom coordinates to input the crystal structure, which either constrains the model to certain crystal types or makes it difficult to provide chem. insights. Here, we develop a crystal graph convolutional neural networks framework to directly learn material properties from the connection of atoms in the crystal, providing a universal and interpretable representation of cryst. materials. Our method provides a highly accurate prediction of d. functional theory calcd. properties for eight different properties of crystals with various structure types and compns. after being trained with 104 data points. Further, our framework is interpretable because one can ext. the contributions from local chem. environments to global properties. Using an example of perovskites, we show how this information can be utilized to discover empirical rules for materials design.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXltFSnu7c%253D&md5=93beb5675af86cf95e07c82c136f3511
    8. 8
      Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; Ong, S. P. Graph networks as a universal machine learning framework for molecules and crystals. Chem. Mater. 2019, 31, 3564– 3572,  DOI: 10.1021/acs.chemmater.9b01294
      8
      Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals
      Chen, Chi; Ye, Weike; Zuo, Yunxing; Zheng, Chen; Ong, Shyue Ping
      Chemistry of Materials (2019), 31 (9), 3564-3572CODEN: CMATEX; ISSN:0897-4756. (American Chemical Society)
      Graph networks are a new machine learning (ML) paradigm that supports both relational reasoning and combinatorial generalization. Here, we develop universal MatErials Graph Network (MEGNet) models for accurate property prediction in both mols. and crystals. We demonstrate that the MEGNet models outperform prior ML models such as the SchNet in 11 out of 13 properties of the QM9 mol. data set. Similarly, we show that MEGNet models trained on ∼60 000 crystals in the Materials Project substantially outperform prior ML models in the prediction of the formation energies, band gaps, and elastic moduli of crystals, achieving better than d. functional theory accuracy over a much larger data set. We present two new strategies to address data limitations common in materials science and chem. First, we demonstrate a phys. intuitive approach to unify four sep. mol. MEGNet models for the internal energy at 0 K and room temp., enthalpy, and Gibbs free energy into a single free energy MEGNet model by incorporating the temp., pressure, and entropy as global state inputs. Second, we show that the learned element embeddings in MEGNet models encode periodic chem. trends and can be transfer-learned from a property model trained on a larger data set (formation energies) to improve property models with smaller amts. of data (band gaps and elastic moduli).
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXntFaqt7g%253D&md5=c924547f17083428802d5698682d5971
    9. 9
      Jiang, Y.; Yang, Z.; Guo, J.; Li, H.; Liu, Y.; Guo, Y.; Li, M.; Pu, X. Coupling complementary strategy to flexible graph neural network for quick discovery of coformer in diverse co-crystal materials. Nat. Commun. 2021, 12, 5950  DOI: 10.1038/s41467-021-26226-7
      9
      Coupling complementary strategy to flexible graph neural network for quick discovery of coformer in diverse co-crystal materials
      Jiang, Yuanyuan; Yang, Zongwei; Guo, Jiali; Li, Hongzhen; Liu, Yijing; Guo, Yanzhi; Li, Menglong; Pu, Xuemei
      Nature Communications (2021), 12 (1), 5950CODEN: NCAOBW; ISSN:2041-1723. (Nature Research)
      Cocrystal engineering have been widely applied in pharmaceutical, chem. and material fields. However, how to effectively choose coformer has been a challenging task on expts. Here we develop a graph neural network (GNN) based deep learning framework to quickly predict formation of the cocrystal. In order to capture main driving force to crystn. from 6819 pos. and 1052 neg. samples reported by expts., a feasible GNN framework is explored to integrate important prior knowledge into end-to-end learning on the mol. graph. The model is strongly validated against seven competitive models and three challenging independent test sets involving pharmaceutical cocrystals, π-π cocrystals and energetic cocrystals, exhibiting superior performance with accuracy higher than 96%, confirming its robustness and generalization. Furthermore, one new energetic cocrystal predicted is successfully synthesized, showcasing high potential of the model in practice. All the data and source codes are available at https://github.com/Saoge123/ccgnet for aiding cocrystal community.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXit1aqsbjE&md5=e877fd347df52869466130c176db61ec
    10. 10
      Park, C. W.; Wolverton, C. Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery. Phys. Rev. Mater. 2020, 4, 063801  DOI: 10.1103/PhysRevMaterials.4.063801
      10
      Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery
      Park, Cheol Woo; Wolverton, Chris
      Physical Review Materials (2020), 4 (6), 063801CODEN: PRMHBS; ISSN:2475-9953. (American Physical Society)
      The recently proposed crystal graph convolutional neural network (CGCNN) offers a highly versatile and accurate machine learning (ML) framework by learning material properties directly from graphlike representations of crystal structures ("crystal graphs"). Here, we develop an improved variant of the CGCNN model (iCGCNN) that outperforms the original by incorporating information of the Voronoi tessellated crystal structure, explicit three-body correlations of neighboring constituent atoms, and an optimized chem. representation of interat. bonds in the crystal graphs. We demonstrate the accuracy of the improved framework in two distinct illustrations: First, when trained/validated on 180 000/20 000 d. functional theory (DFT) calcd. thermodn. stability entries taken from the Open Quantum Materials Database (OQMD) and evaluated on a sep. test set of 230 000 entries, iCGCNN achieves a predictive accuracy that is significantly improved, i.e., 20% higher than that of the original CGCNN. Second, when used to assist a high-throughput search for materials in the ThCr2Si2 structure-type, iCGCNN exhibited a success rate of 31% which is 155 times higher than an undirected high-throughput search and 2.4 times higher than that of the original CGCNN. Using both CGCNN and iCGCNN, we screened 132 600 compds. with elemental decorations of the ThCr2Si2 prototype crystal structure and identified a total of 97 unique stable compds. by performing 757 DFT calcns., accelerating the computational time of the high-throughput search by a factor of 65. Our results suggest that the iCGCNN can be used to accelerate high-throughput discoveries of new materials by quickly and accurately identifying cryst. compds. with properties of interest.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsFKis7fP&md5=0b5903b05d47535d7c2967b7d9d6fc6b
    11. 11
      Louis, S. Y.; Zhao, Y.; Nasiri, A.; Wang, X.; Song, Y.; Liu, F.; Hu, J. Graph convolutional neural networks with global attention for improved materials property prediction. Phys. Chem. Chem. Phys. 2020, 22, 18141– 18148,  DOI: 10.1039/D0CP01474E
      11
      Graph convolutional neural networks with global attention for improved materials property prediction
      Louis, Steph-Yves; Zhao, Yong; Nasiri, Alireza; Wang, Xiran; Song, Yuqi; Liu, Fei; Hu, Jianjun
      Physical Chemistry Chemical Physics (2020), 22 (32), 18141-18148CODEN: PPCPFQ; ISSN:1463-9076. (Royal Society of Chemistry)
      The development of an efficient and powerful machine learning (ML) model for materials property prediction (MPP) remains an important challenge in materials science. While various techniques have been proposed to ext. physicochem. features in MPP, graph neural networks (GNN) have also shown very strong capability in capturing effective features for high-performance MPP. Nevertheless, current GNN models do not effectively differentiate the contributions from different atoms. In this paper we develop a novel graph neural network model called GATGNN for predicting properties of inorg. materials. GATGNN is characterized by its compn. of augmented graph-attention layers (AGAT) and a global attention layer. The application of AGAT layers and global attention layers resp. learn the local relationship among neighboring atoms and overall contribution of the atoms to the material's property; together making our framework achieve considerably better prediction performance on various tested properties. Through extensive expts., we show that our method is able to outperform existing state-of-the-art GNN models while it can also provide a measurable insight into the correlation between the atoms and their material property.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsFSmu7bL&md5=9190661d65053bc86d36c270be20cf0d
    12. 12
      Cheng, J.; Zhang, C.; Dong, L. A geometric-information-enhanced crystal graph network for predicting properties of materials. Commun. Mater. 2021, 2, 92,  DOI: 10.1038/s43246-021-00194-3
      12
      A geometric-information-enhanced crystal graph network for predicting properties of materials
      Cheng, Jiucheng; Zhang, Chunkai; Dong, Lifeng
      Communications Materials (2021), 2 (1), 92CODEN: CMOAGE; ISSN:2662-4443. (Nature Portfolio)
      Graph neural networks (GNNs) have been used previously for identifying new cryst. materials. However, geometric structure is not usually taken into consideration, or only partially. Here, we develop a geometric-information-enhanced crystal graph neural network (GeoCGNN) to predict the properties of cryst. materials. By considering the distance vector between each node and its neighbors, our model can learn full topol. and spatial geometric structure information. Furthermore, we incorporate an effective method based on the mixed basis functions to encode the geometric information into our model, which outperforms other GNN methods in a variety of databases. For example, for predicting formation energy our model is 25.6%, 14.3% and 35.7% more accurate than CGCNN, MEGNet and iCGCNN models, resp. For band gap, our model outperforms CGCNN by 27.6% and MEGNet by 12.4%.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XotVCku78%253D&md5=44ae02fdbbde3d1eef3d2e2fe77fb860
    13. 13
      Choudhary, K.; DeCost, B. Atomistic line graph neural network for improved materials property predictions. npj Comput. Mater. 2021, 7, 185,  DOI: 10.1038/s41524-021-00650-1
      There is no corresponding record for this reference.
    14. 14
      Pandey, S.; Qu, J.; Stevanović, V.; John, P. S.; Gorai, P. Predicting energy and stability of known and hypothetical crystals using graph neural network. Patterns 2021, 2, 100361  DOI: 10.1016/j.patter.2021.100361
      14
      Predicting energy and stability of known and hypothetical crystals using graph neural network
      Pandey Shubham; Stevanovic Vladan; Gorai Prashun; Qu Jiaxing; St John Peter
      Patterns (New York, N.Y.) (2021), 2 (11), 100361 ISSN:.
      The discovery of new inorganic materials in unexplored chemical spaces necessitates calculating total energy quickly and with sufficient accuracy. Machine learning models that provide such a capability for both ground-state (GS) and higher-energy structures would be instrumental in accelerated screening. Here, we demonstrate the importance of a balanced training dataset of GS and higher-energy structures to accurately predict total energies using a generic graph neural network architecture. Using [Formula: see text] 16,500 density functional theory calculations from the National Renewable Energy Laboratory (NREL) Materials Database and [Formula: see text] 11,000 calculations for hypothetical structures as our training database, we demonstrate that our model satisfactorily ranks the structures in the correct order of total energies for a given composition. Furthermore, we present a thorough error analysis to explain failure modes of the model, including both prediction outliers and occasional inconsistencies in the training data. By examining intermediate layers of the model, we analyze how the model represents learned structures and properties.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB2cfntlejtA%253D%253D&md5=fc71858fe7691e1e930bf632ef025975
    15. 15
      Frey, N. C.; Akinwande, D.; Jariwala, D.; Shenoy, V. B. Machine learning-enabled design of point defects in 2d materials for quantum and neuromorphic information processing. ACS Nano 2020, 14, 13406– 13417,  DOI: 10.1021/acsnano.0c05267
      15
      Machine Learning-Enabled Design of Point Defects in 2D Materials for Quantum and Neuromorphic Information Processing
      Frey, Nathan C.; Akinwande, Deji; Jariwala, Deep; Shenoy, Vivek B.
      ACS Nano (2020), 14 (10), 13406-13417CODEN: ANCAC3; ISSN:1936-0851. (American Chemical Society)
      Engineered point defects in two-dimensional (2D) materials offer an attractive platform for solid-state devices that exploit tailored optoelectronic, quantum emission, and resistive properties. Naturally occurring defects are also unavoidably important contributors to material properties and performance. The immense variety and complexity of possible defects make it challenging to exptl. control, probe, or understand at.-scale defect-property relationships. Here, we develop an approach based on deep transfer learning, machine learning, and first-principles calcns. to rapidly predict key properties of point defects in 2D materials. We use physics-informed featurization to generate a minimal description of defect structures and present a general picture of defects across materials systems. We identify over one hundred promising, unexplored dopant defect structures in layered metal chalcogenides, hexagonal nitrides, and metal halides. These defects are prime candidates for quantum emission, resistive switching, and neuromorphic computing.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslKqtLfF&md5=1ade77df47fb8faf2dbe58746ccb4788
    16. 16
      Fung, V.; Zhang, J.; Juarez, E.; Sumpter, B. G. Benchmarking graph neural networks for materials chemistry. npj Comput. Mater. 2021, 7, 84,  DOI: 10.1038/s41524-021-00554-0
      16
      Benchmarking graph neural networks for materials chemistry
      Fung, Victor; Zhang, Jiaxin; Juarez, Eric; Sumpter, Bobby G.
      npj Computational Materials (2021), 7 (1), 84CODEN: NCMPCS; ISSN:2057-3960. (Nature Research)
      Graph neural networks (GNNs) have received intense interest as a rapidly expanding class of machine learning models remarkably well-suited for materials applications. To date, a no. of successful GNNs have been proposed and demonstrated for systems ranging from crystal stability to electronic property prediction and to surface chem. and heterogeneous catalysis. However, a consistent benchmark of these models remains lacking, hindering the development and consistent evaluation of new models in the materials field. Here, we present a workflow and testing platform, MatDeepLearn, for quickly and reproducibly assessing and comparing GNNs and other machine learning models. We use this platform to optimize and evaluate a selection of top performing GNNs on several representative datasets in computational materials chem. From our investigations we note the importance of hyperparameter selection and find roughly similar performances for the top models once optimized. We identify several strengths in GNNs over conventional models in cases with compositionally diverse datasets and in its overall flexibility with respect to inputs, due to learned rather than defined representations. Meanwhile several weaknesses of GNNs are also obsd. including high data requirements, and suggestions for further improvement for applications in materials chem. are discussed.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFehs7zM&md5=e49c3e0b3e69dda5e5b41a33e3eba307
    17. 17
      Jalali, M.; Tsotsalas, M.; Wöll, C. MOFSocialNet: Exploiting metal-organic framework relationships via social network analysis. Nanomaterials 2022, 12, 704,  DOI: 10.3390/nano12040704
      17
      MOFSocialNet: Exploiting Metal-Organic Framework Relationships via Social Network Analysis
      Jalali, Mehrdad; Tsotsalas, Manuel; Woell, Christof
      Nanomaterials (2022), 12 (4), 704CODEN: NANOKO; ISSN:2079-4991. (MDPI AG)
      The no. of metal-org. frameworks (MOF) as well as the no. of applications of this material are growing rapidly. With the no. of characterized compds. exceeding 100,000, manual sorting becomes impossible. At the same time, the increasing computer power and established use of automated machine learning approaches makes data science tools available, that provide an overview of the MOF chem. space and support the selection of suitable MOFs for a desired application. Among the different data science tools, graph theory approaches, where data generated from numerous real-world applications is represented as a graph (network) of interconnected objects, has been widely used in a variety of scientific fields such as social sciences, health informatics, biol. sciences, agricultural sciences and economics. We describe the application of a particular graph theory approach known as social network anal. to MOF materials and highlight the importance of community (group) detection and graph node centrality. In this first application of the social network anal. approach to MOF chem. space, we created MOFSocialNet. This social network is based on the geometrical descriptors of MOFs available in the CoRE-MOFs database. MOFSocialNet can discover communities with similar MOFs structures and identify the most representative MOFs within a given community. In addn., anal. of MOFSocialNet using social network anal. methods can predict MOF properties more accurately than conventional ML tools. The latter advantage is demonstrated for the prediction of gas storage properties, the most important property of these porous reticular networks.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XlslWmtLo%253D&md5=1bb836c9d57b2127944c9e1b8a45050e
    18. 18
      Chen, C.; Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nat. Comput. Sci. 2022, 2, 718– 728,  DOI: 10.1038/s43588-022-00349-3
      There is no corresponding record for this reference.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=&md5=9874b665cc7a056b8e2f928dd3112440
    19. 19
      Egorova, O.; Hafizi, R.; Woods, D. C.; Day, G. M. Multifidelity statistical machine learning for molecular crystal structure prediction. J. Phys. Chem. A 2020, 124, 8065– 8078,  DOI: 10.1021/acs.jpca.0c05006
      19
      Multifidelity Statistical Machine Learning for Molecular Crystal Structure Prediction
      Egorova, Olga; Hafizi, Roohollah; Woods, David C.; Day, Graeme M.
      Journal of Physical Chemistry A (2020), 124 (39), 8065-8078CODEN: JPCAFH; ISSN:1089-5639. (American Chemical Society)
      Statistical machine learning was applied to predict expensive hybrid functional DFT (PBE0) calcns. of crystal structures using a multifidelity approach to reevaluate the energies of crystal structures predicted with an inexpensive force field. The method uses an autoregressive Gaussian process, making use of less expensive GGA DFT (PBE) calcns. to bridge the gap between the force field and PBE0 energies. The method is benchmarked on the crystal structure landscapes of 3 small, H-bonded org. mols. and produces accurate predictions of energies and crystal structure ranking using small nos. of the most expensive calcns.; the PBE0 energies can be predicted with errors of <1 kJ mol-1 with between 4.2 and 6.8% of the cost of the full calcns. As the developed model is probabilistic, how the uncertainties in predicted energies impact the assessment of the energetic ranking of crystal structures are discussed.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslequ7jL&md5=15ef21ac6c7aafa85f5e4fc3bb3e0fce
    20. 20
      Han, Y.; Ali, I.; Wang, Z.; Cai, J.; Wu, S.; Tang, J.; Zhang, L.; Ren, J.; Xiao, R.; Lu, Q.; Hang, L. Machine learning accelerates quantum mechanics predictions of molecular crystals. Phys. Rep. 2021, 934, 1– 71,  DOI: 10.1016/j.physrep.2021.08.002
      20
      Machine learning accelerates quantum mechanics predictions of molecular crystals
      Han, Yanqiang; Ali, Imran; Wang, Zhilong; Cai, Junfei; Wu, Sicheng; Tang, Jiequn; Zhang, Lin; Ren, Jiahao; Xiao, Rui; Lu, Qianqian; Hang, Lei; Luo, Hongyuan; Li, Jinjin
      Physics Reports (2021), 934 (), 1-71CODEN: PRPLCM; ISSN:0370-1573. (Elsevier B.V.)
      A review. Quantum mechanics (QM) approaches (DFT, MP2, CCSD(T), etc.) play an important role in calcg. mols. and crystals with a high accuracy and acceptable efficiency. In recent years, with the development of artificial intelligence technol., machine learning (ML) has played an increasingly essential role in accelerating the QM calcns. and predictions of mol. crystals, as well as the discovery of novel materials. This review provides state-of-the-art information and prospects for QM theories, fragment-based methods and ML methods, as well as their up-to-date applications in predicting small inorg. mols., large drug mols. and relevant mol. crystals. The discussed applications include ML potential energy surface (PES) construction, crystal structure prediction (CSP), chem. reaction prediction and predictions of a series of properties, such as structure, energy, at. force, bond length, chem. shift, supercond., super-hardness, vibrational spectra, phase transition and diagram. This work also reviews software and packages built recently based on ML methods for property predictions and PES constructions in the field of physics and chem. For the three discussed methods, the most time-consuming one is the high-level all-atom QM method, which is capable of describing electronic structures with high accuracy and thus predicts properties that are consistent with the exptl. results. The second one, fragment-based QM method, requires less computational time than all-atom QM, which can accelerate all-atom QM calcns. for large systems by dividing the entire system into subsystems, presenting a considerable efficiency increase. The computational complexities for fragment-based QM and all-atom QM are N - N2 and N5-N7 (N is the size of the system), resp. A well-trained ML model can make the above predictions within seconds while ensuring a high prediction accuracy, where its prediction cost and accuracy are detd. by the training data and the training process. Therefore, it is challenging for ML applications in physics and chem. to generate highly accurate and powerful ML models while ensuring sufficient datasets. This work not only provides an overview of the recent progress in QM theories, fragment-based methods, ML methods and several ML-based software programs and applications on small inorg. mols., large drug mols. and relevant crystals, but also shed light on ML methods in accelerating QM prediction, optimization and novel crystal material design.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFWiurzI&md5=829b4c028e46feb7be228b2062790582
    21. 21
      Ishizaki, K.; Sugimoto, R.; Hagiwara, Y.; Koshima, H.; Taniguchi, T.; Asahi, T. Actuation performance of a photo-bending crystal modeled by machine learning-based regression. CrystEngComm 2021, 23, 5839– 5847,  DOI: 10.1039/D1CE00208B
      21
      Actuation performance of a photo-bending crystal modeled by machine learning-based regression
      Ishizaki, Kazuki; Sugimoto, Ryota; Hagiwara, Yuki; Koshima, Hideko; Taniguchi, Takuya; Asahi, Toru
      CrystEngComm (2021), 23 (34), 5839-5847CODEN: CRECF4; ISSN:1466-8033. (Royal Society of Chemistry)
      Photomech. mol. crystals are a family of mech. responsive materials that have been recognized as a novel type of actuator. The actuation properties of photomech. crystals should be characterized. However, deflection and force, which are crucial for actuators, are dependent on exptl. conditions such as light intensity and crystal size, and thus the no. of combinations under different conditions is infinite. This causes difficulty in obtaining the relationship between the exptl. conditions and actuation outputs. To solve this problem, this paper presents a machine learning-based regression approach for modeling the deflection and force of a photo-bending crystal. The deflection and force were exptl. measured for various light intensities and crystal sizes. These time-series data of deflection and force were represented by feature values via exponential fitting for bending and unbending processes. The features of the max. value and speed of deflection and force were analyzed by polynomial regression and variable selection. Through this process, the most fitted models were constructed for deflection and force and most of them were interpreted by material mechanics. This statistical strategy will potentially be applied to control or optimize other functions of stimuli-responsive crystals.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmtVSqu7k%253D&md5=3ac98d93899abba3690d8a0f9114e0b6
    22. 22
      Groom, C. R.; Bruno, I. J.; Lightfoot, M. P.; Ward, S. C. The Cambridge structural database. Acta Crystallogr., Sect. B 2016, 72, 171– 179,  DOI: 10.1107/S2052520616003954
      22
      The Cambridge Structural Database
      Groom, Colin R.; Bruno, Ian J.; Lightfoot, Matthew P.; Ward, Suzanna C.
      Acta Crystallographica, Section B: Structural Science, Crystal Engineering and Materials (2016), 72 (2), 171-179CODEN: ACSBDA; ISSN:2052-5206. (International Union of Crystallography)
      The Cambridge Structural Database (CSD) contains a complete record of all published org. and metal-org. small-mol. crystal structures. The database has been in operation for over 50 years and continues to be the primary means of sharing structural chem. data and knowledge across disciplines. As well as structures that are made public to support scientific articles, it includes many structures published directly as CSD Communications. All structures are processed both computationally and by expert structural chem. editors prior to entering the database. A key component of this processing is the reliable assocn. of the chem. identity of the structure studied with the exptl. data. This important step helps ensure that data is widely discoverable and readily reusable. Content is further enriched through selective inclusion of addnl. exptl. data. Entries are available to anyone through free CSD community web services. Linking services developed and maintained by the CCDC, combined with the use of std. identifiers, facilitate discovery from other resources. Data can also be accessed through CCDC and third party software applications and through an application programming interface.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28Xls1Kntro%253D&md5=f9c65ab86fc9db429588c95b0da3f9b2
    23. 23
      Gražulis, S.; Daškevič, A.; Merkys, A.; Chateigner, D.; Lutterotti, L.; Quirós, M.; Serebryanaya, N. R.; Moeck, P.; Downs, R. T.; Bail, A. L. Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration. Nucleic Acids Res. 2012, 40, D420– D427,  DOI: 10.1093/nar/gkr900
      23
      Crystallography Open Database (COD): an open-access collection of crystal structures and platform for world-wide collaboration
      Grazulis, Saulius; Daskevic, Adriana; Merkys, Andrius; Chateigner, Daniel; Lutterotti, Luca; Quiros, Miguel; Serebryanaya, Nadezhda R.; Moeck, Peter; Downs, Robert T.; Le Bail, Armel
      Nucleic Acids Research (2012), 40 (D1), D420-D427CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      Using an open-access distribution model, the Crystallog. Open Database (COD, http://www.crystallog.net) collects all known small mol. / small to medium sized unit cell' crystal structures and makes them available freely on the Internet. As of today, the COD has aggregated ∼150 000 structures, offering basic search capabilities and the possibility to download the whole database, or parts thereof using a variety of std. open communication protocols. A newly developed website provides capabilities for all registered users to deposit published and so far unpublished structures as personal communications or pre-publication depositions. Such a setup enables extension of the COD database by many users simultaneously. This increases the possibilities for growth of the COD database, and is the first step towards establishing a world wide Internet-based collaborative platform dedicated to the collection and curation of structural knowledge.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs12htbnP&md5=670b6cc1f2758728c303c4ad89e1cc6f
    24. 24
      Olsthoorn, B.; Geilhufe, R. M.; Borysov, S. S.; Balatsky, A. V. Band gap prediction for large organic crystal structures with machine learning. Adv. Quantum Technol. 2019, 2, 1900023  DOI: 10.1002/qute.201900023
      There is no corresponding record for this reference.
    25. 25
      Bartók, A. P.; Kondor, R.; Csányi, G. On representing chemical environments. Phys. Rev. B 2013, 87, 184115  DOI: 10.1103/PhysRevB.87.184115
      25
      On representing chemical environments
      Bartok, Albert P.; Kondor, Risi; Csanyi, Gabor
      Physical Review B: Condensed Matter and Materials Physics (2013), 87 (18), 184115/1-184115/16CODEN: PRBMDO; ISSN:1098-0121. (American Physical Society)
      We review some recently published methods to represent at. neighborhood environments, and analyze their relative merits in terms of their faithfulness and suitability for fitting potential energy surfaces. The crucial properties that such representations (sometimes called descriptors) must have are differentiability with respect to moving the atoms and invariance to the basic symmetries of physics: rotation, reflection, translation, and permutation of atoms of the same species. We demonstrate that certain widely used descriptors that initially look quite different are specific cases of a general approach, in which a finite set of basis functions with increasing angular wave nos. are used to expand the at. neighborhood d. function. Using the example system of small clusters, we quant. show that this expansion needs to be carried to higher and higher wave nos. as the no. of neighbors increases in order to obtain a faithful representation, and that variants of the descriptors converge at very different rates. We also propose an altogether different approach, called Smooth Overlap of Atomic Positions, that sidesteps these difficulties by directly defining the similarity between any two neighborhood environments, and show that it is still closely connected to the invariant descriptors. We test the performance of the various representations by fitting models to the potential energy surface of small silicon clusters and the bulk crystal.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXpvFClu7Y%253D&md5=f7739275562b8e77d4532f00da8814fb
    26. 26
      Wengert, S.; Csányi, G.; Reuter, K.; Margraf, J. T. Data-efficient machine learning for molecular crystal structure prediction. Chem. Sci. 2021, 12, 4536– 4546,  DOI: 10.1039/D0SC05765G
      26
      Data-efficient machine learning for molecular crystal structure prediction
      Wengert, Simon; Csanyi, Gabor; Reuter, Karsten; Margraf, Johannes T.
      Chemical Science (2021), 12 (12), 4536-4546CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      The combination of modern machine learning (ML) approaches with high-quality data from quantum mech. (QM) calcns. can yield models with an unrivalled accuracy/cost ratio. However, such methods are ultimately limited by the computational effort required to produce the ref. data. In particular, ref. calcns. for periodic systems with many atoms can become prohibitively expensive for higher levels of theory. This trade-off is crit. in the context of org. crystal structure prediction (CSP). Here, a data-efficient ML approach would be highly desirable, since screening a huge space of possible polymorphs in a narrow energy range requires the assessment of a large no. of trial structures with high accuracy. In this contribution, we present tailored Δ-ML models that allow screening a wide range of crystal candidates while adequately describing the subtle interplay between intermol. interactions such as H-bonding and many-body dispersion effects. This is achieved by enhancing a physics-based description of long-range interactions at the d. functional tight binding (DFTB) level-for which an efficient implementation is available-with a short-range ML model trained on high-quality first-principles ref. data. The presented workflow is broadly applicable to different mol. materials, without the need for a single periodic calcn. at the ref. level of theory. We show that this even allows the use of wavefunction methods in CSP.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXjslWrtr4%253D&md5=b168772ffa3e022a39d1e57a954804ff
    27. 27
      Takagi, D.; Ishizaki, K.; Asahi, T.; Taniguchi, T. Molecular screening for solid–solid phase transition by machine learning. Digital Discovery 2023, 2, 1126,  DOI: 10.1039/D3DD00034F
      27
      Molecular screening for solid-solid phase transitions by machine learning
      Takagi, Daisuke; Ishizaki, Kazuki; Asahi, Toru; Taniguchi, Takuya
      Digital Discovery (2023), 2 (4), 1126-1133CODEN: DDIIAI; ISSN:2635-098X. (Royal Society of Chemistry)
      The solid-solid phase transition in mol. crystals is generally found by chance empirically. In this study, we constructed a machine learning framework to screen mols. that will exhibit solid-solid phase transitions in their cryst. states, based on pos.-unlabeled learning. We trained classification models using the pos. dataset we constructed manually and the unlabeled data extd. from the Cambridge Structural Database. The best classifier works as a suggester, and 9 substances among the suggested 113 mols. were found to exhibit solid-solid phase transitions according to the literature and expts. The finding probability of 8.0% is much higher than the probability of phase transition in the database, suggesting the effectiveness of mol. selection by this workflow. We also found that the mol. structure is weakly related to the transition temp. by regression anal. The findings of this study are useful for designing functional mol. crystals with solid-solid phase transitions.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3sXhvVygurvF&md5=027166e2d42022ecf59c4e8d6846490d
    28. 28
      Musil, F.; De, S.; Yang, J.; Campbell, J. E.; Day, G. M.; Ceriotti, M. Machine learning for the structure–energy–property landscapes of molecular crystals. Chem. Sci. 2018, 9, 1289– 1300,  DOI: 10.1039/C7SC04665K
      28
      Machine learning for the structure-energy-property landscapes of molecular crystals
      Musil, Felix; De, Sandip; Yang, Jack; Campbell, Joshua E.; Day, Graeme M.; Ceriotti, Michele
      Chemical Science (2018), 9 (5), 1289-1300CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Mol. crystals play an important role in several fields of science and technol. They frequently crystallize in different polymorphs with substantially different phys. properties. To help guide the synthesis of candidate materials, at.-scale modeling can be used to enumerate the stable polymorphs and to predict their properties, as well as to propose heuristic rules to rationalize the correlations between crystal structure and materials properties. Here we show how a recently-developed machine-learning (ML) framework can be used to achieve inexpensive and accurate predictions of the stability and properties of polymorphs, and a data-driven classification that is less biased and more flexible than typical heuristic rules. We discuss, as examples, the lattice energy and property landscapes of pentacene and two azapentacene isomers that are of interest as org. semiconductor materials. We show that we can est. force field or DFT lattice energies with sub-kJ mol-1 accuracy, using only a few hundred ref. configurations, and reduce by a factor of ten the computational effort needed to predict charge mobility in the crystal structures. The automatic structural classification of the polymorphs reveals a more detailed picture of mol. packing than that provided by conventional heuristics, and helps disentangle the role of hydrogen bonded and π-stacking interactions in detg. mol. self-assembly. This observation demonstrates that ML is not just a black-box scheme to interpolate between ref. calcns., but can also be used as a tool to gain intuitive insights into structure-property relations in mol. crystal engineering.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhvFGnsbfE&md5=48b184722478d1a988a2090f765135e2
    29. 29
      Quirós, M.; Gražulis, S.; Girdzijauskaitė, S.; Merkys, A.; Vaitkus, A. Using SMILES strings for the description of chemical connectivity in the Crystallography Open Database. J. Cheminf. 2018, 10, 23,  DOI: 10.1186/s13321-018-0279-6
      29
      Using SMILES strings for the description of chemical connectivity in the Crystallography Open Database
      Quiros, Miguel; Grazulis, Saulius; Girdzijauskaite, Saule; Merkys, Andrius; Vaitkus, Antanas
      Journal of Cheminformatics (2018), 10 (), 23/1-23/17CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      Computer descriptions of chem. mol. connectivity are necessary for searching chem. databases and for predicting chem. properties from mol. structure. In this article, the ongoing work to describe the chem. connectivity of entries contained in the Crystallog. Open Database (COD) in SMILES format is reported. This collection of SMILES is publicly available for chem. (substructure) search or for any other purpose on an open-access basis, as is the COD itself. The conventions that have been followed for the representation of compds. that do not fit into the valence bond theory are outlined for the most frequently found cases. The procedure for getting the SMILES out of the CIF files starts with checking whether the atoms in the asym. unit are a chem. acceptable image of the compd. When they are not (mol. in a symmetry element, disorder, polymeric species, etc.), the previously published cif_mol. program is used to get such image in many cases. The program package Open Babel is then applied to get SMILES strings from the CIF files (either those directly taken from the COD or those produced by cif_mol. when applicable). The results are then checked and/or fixed by a human editor, in a computer-aided task that at present still consumes a great deal of human time. Even if the procedure still needs to be improved to make it more automatic (and hence faster), it has already yielded more than 160,000 curated chem. structures and the purpose of this article is to announce the existence of this work to the chem. community as well as to spread the use of its results.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWiurk%253D&md5=99cd2cb8206426aae7de48ea932fe169
    30. 30
      Feng, X.; Becke, A. D.; Johnson, E. R. Theoretical investigation of polymorph-and coformer-dependent photoluminescence in molecular crystals. CrystEngComm 2021, 23, 4264– 4271,  DOI: 10.1039/D1CE00383F
      30
      Theoretical investigation of polymorph- and coformer-dependent photoluminescence in molecular crystals
      Feng, Xibo; Becke, Axel D.; Johnson, Erin R.
      CrystEngComm (2021), 23 (24), 4264-4271CODEN: CRECF4; ISSN:1466-8033. (Royal Society of Chemistry)
      Polymorph- and coformer-dependent photoluminescence (PL) are among the variety of novel solid-state PL phenomena recently obsd. in many mol. crystals. They are of particular research interest due to their direct connections to two heavily investigated topics in crystal engineering: polymorphism and cocrystn. Herein, we apply a novel computational methodol., initially proposed and successfully applied in our previous investigation of piezochromism, to theor. modeling of the polymorph- and coformer-dependent PL in the well-known ROY polymorphs and the recently synthesized 9-acetylanthracene (9-ACA) cocrystals, resp. Our methodol. offers satisfactory prediction of the exptl. obsd. color zoning for the ROY polymorphs and provides good qual. and quant. accuracy for the emission (fluorescence) energies of the 9-ACA cocrystals, although the results in both cases may be adversely affected by delocalization error in the d.-functional methods employed. While the polymorph-dependent PL in ROY is found to be controlled by the intramol. geometry, modeling of the periodic crystal environment is necessary for accurate prediction of the coformer-dependent PL in the 9-ACA cocrystals, which is driven by charge transfer.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhtFSkurrK&md5=4789ac696c9e43357bc15af8988abfed
    31. 31
      Aziz, A.; Sidat, A.; Talati, P.; Crespo-Otero, R. Understanding the solid state luminescence and piezochromic properties in polymorphs of an anthracene derivative. Phys. Chem. Chem. Phys. 2022, 24, 2832– 2842,  DOI: 10.1039/D1CP05192J
      31
      Understanding the solid state luminescence and piezochromic properties in polymorphs of an anthracene derivative
      Aziz, Alex; Sidat, Amir; Talati, Priyesh; Crespo-Otero, Rachel
      Physical Chemistry Chemical Physics (2022), 24 (5), 2832-2842CODEN: PPCPFQ; ISSN:1463-9076. (Royal Society of Chemistry)
      Luminescent mol. crystals have gained significant research interest for optoelectronic applications. However, fully understanding their structural and electronic relationships in the condensed phase and under external stimuli remains a significant challenge. Here, piezochromism in the mol. crystal 9,10-bis((E)-2-(pyridin-4-yl)vinyl)anthracene (BP4VA) is studied using a combination of d. functional theory (DFT) and time-dependent TD-DFT. We investigate the effects that mol. packing and geometry have on the electronic and phonon structure and the excited state properties in this archetypal system. We find that the luminescence properties are red-shifted with the transition from a herringbone to a sheet packing arrangement. An almost continuous red-shift in the band gap is found with the application of an external pressure through the enhancement of π-π and CH-π interactions, and is a mechanism in fine tuning an emissive response. The anal. of the phonon structure of the mol. crystal suggests restriction of motion in the herringbone packing arrangement, with motion restricted at higher pressure. This is supported by the Huang-Rhys factors which show a decrease in the reorganisation energy with the application of pressure. Ultimately, a balance between the decrease in reorganisation energies and the increase in exciton coupling will det. whether nonradiative decay is enhanced or decreased with the increase in pressure in these systems.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhsF2ls7g%253D&md5=7bb4c15ceff7e88447a96420c707f5a2
    32. 32
      Odom, S. A.; Caruso, M. M.; Finke, A. D.; Prokup, A. M.; Ritchey, J. A.; Leonard, J. H.; White, S. R.; Sottos, N. R.; Moore, J. S. Restoration of conductivity with TTF-TCNQ charge-transfer salts. Adv. Funct. Mater. 2010, 20, 1721– 1727,  DOI: 10.1002/adfm.201000159
      32
      Restoration of Conductivity with TTF-TCNQ Charge-Transfer Salts
      Odom, Susan A.; Caruso, Mary M.; Finke, Aaron D.; Prokup, Alex M.; Ritchey, Joshua A.; Leonard, John H.; White, Scott R.; Sottos, Nancy R.; Moore, Jeffrey S.
      Advanced Functional Materials (2010), 20 (11), 1721-1727CODEN: AFMDC6; ISSN:1616-301X. (Wiley-VCH Verlag GmbH & Co. KGaA)
      The formation of the conductive TTF-TCNQ (tetrathiafulvalene-tetracyanoquinodimethane) charge-transfer salt via rupture of microencapsulated solns. of its individual components is reported. Solns. of TTF and TCNQ in various solvents are sep. incorporated into poly(urea-formaldehyde) core-shell microcapsules. Rupture of a mixt. of TTF-contg. microcapsules and TCNQ-contg. microcapsules gave the cryst. salt, as verified by FTIR spectroscopy and powder x-ray diffraction. Preliminary measurements demonstrate the partial restoration of cond. of severed gold electrodes in the presence of TTF-TCNQ derived in situ. This is the 1st microcapsule system for the restoration of cond. in mech. damaged electronic devices in which the repairing agent is not conductive until its release.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXntVyju7Y%253D&md5=a590d7aeaab369f76599fcd14cf8ab70
    33. 33
      Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 1994, 50, 17953,  DOI: 10.1103/PhysRevB.50.17953
      33
      Projector augmented-wave method
      Blochl
      Physical review. B, Condensed matter (1994), 50 (24), 17953-17979 ISSN:0163-1829.
      There is no expanded citation for this reference.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC2sfjslSntA%253D%253D&md5=1853d67af808af2edab58beaab5d3051
    34. 34
      Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 1999, 59, 1758,  DOI: 10.1103/PhysRevB.59.1758
      34
      From ultrasoft pseudopotentials to the projector augmented-wave method
      Kresse, G.; Joubert, D.
      Physical Review B: Condensed Matter and Materials Physics (1999), 59 (3), 1758-1775CODEN: PRBMDO; ISSN:0163-1829. (American Physical Society)
      The formal relationship between ultrasoft (US) Vanderbilt-type pseudopotentials and Blochl's projector augmented wave (PAW) method is derived. The total energy functional for US pseudopotentials can be obtained by linearization of two terms in a slightly modified PAW total energy functional. The Hamilton operator, the forces, and the stress tensor are derived for this modified PAW functional. A simple way to implement the PAW method in existing plane-wave codes supporting US pseudopotentials is pointed out. In addn., crit. tests are presented to compare the accuracy and efficiency of the PAW and the US pseudopotential method with relaxed-core all-electron methods. These tests include small mols. (H2, H2O, Li2, N2, F2, BF3, SiF4) and several bulk systems (diamond, Si, V, Li, Ca, CaF2, Fe, Co, Ni). Particular attention is paid to the bulk properties and magnetic energies of Fe, Co, and Ni.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK1MXkt12nug%253D%253D&md5=78a73e92a93f995982fc481715729b14
    35. 35
      Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 1996, 77, 3865,  DOI: 10.1103/PhysRevLett.77.3865
      35
      Generalized gradient approximation made simple
      Perdew, John P.; Burke, Kieron; Ernzerhof, Matthias
      Physical Review Letters (1996), 77 (18), 3865-3868CODEN: PRLTAO; ISSN:0031-9007. (American Physical Society)
      Generalized gradient approxns. (GGA's) for the exchange-correlation energy improve upon the local spin d. (LSD) description of atoms, mols., and solids. We present a simple derivation of a simple GGA, in which all parameters (other than those in LSD) are fundamental consts. Only general features of the detailed construction underlying the Perdew-Wang 1991 (PW91) GGA are invoked. Improvements over PW91 include an accurate description of the linear response of the uniform electron gas, correct behavior under uniform scaling, and a smoother potential.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK28XmsVCgsbs%253D&md5=55943538406ee74f93aabdf882cd4630

  • Supporting Information

    Supporting Information

    The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.3c05224.

    • Hyperparameters, data set for generalization test, data distribution (PDF)


    Terms & Conditions

    Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.