Atomistic Probing of Defect-Engineered 2H-MoTe2 Monolayers

Point defects dictate various physical, chemical, and optoelectronic properties of two-dimensional (2D) materials, and therefore, a rudimentary understanding of the formation and spatial distribution of point defects is a key to advancement in 2D material-based nanotechnology. In this work, we performed the demonstration to directly probe the point defects in 2H-MoTe2 monolayers that are tactically exposed to (i) 200 °C-vacuum-annealing and (ii) 532 nm-laser-illumination; and accordingly, we utilize a deep learning algorithm to classify and quantify the generated point defects. We discovered that tellurium-related defects are mainly generated in both 2H-MoTe2 samples; but interestingly, 200 °C-vacuum-annealing and 532 nm-laser-illumination modulate a strong n-type and strong p-type 2H-MoTe2, respectively. While 200 °C-vacuum-annealing generates tellurium vacancies or tellurium adatoms, 532 nm-laser-illumination prompts oxygen atoms to be adsorbed/chemisorbed at tellurium vacancies, giving rise to the p-type characteristic. This work significantly advances the current understanding of point defect engineering in 2H-MoTe2 monolayers and other 2D materials, which is critical for developing nanoscale devices with desired functionality.


Supporting Text 1. Experimental and simulation HAADF-STEM analyses of defects in 200˚C-vacuum-annealed 2H-MoTe2 ML.
To verify the defect species visualized in the 200˚C-vacuum-annealed 2H-MoTe2 (VA 2H-MoTe2) ML, we performed high-resolution HAADF-STEM simulation analysis and compared the intensity profiles with the experimental results.The HAADF-STEM images were simulated from atomic models constructed taking into consideration the possible prevalent defects.and simulated (solid orange/light-gray lines) HAADF-STEM intensity profiles extracted from a-c along the diagonal rectangles.We confirm the defect observed in the experimental HAADF-STEM image as Tead2.The color-codes are the same as in Figure 2, where Mo-related defects are marked in light-gray for comparison.

Supporting Text 2. Experimental and simulation HAADF-(ABF-)STEM analysis of defects in 532-nm-laser-illuminated 2H-MoTe2 ML.
To verify the defect species observed in the 532-nm-laser-illuminated 2H-MoTe2 (LI 2H-MoTe2) ML, we performed high-resolution HAADF-(ABF-)STEM simulation analysis and compared the intensity profiles with the experimental results.The HAADF-(ABF-)STEM images were simulated from atomic models constructed taking into consideration the possible prevalent defects.With the focused the laser beam, the bulk sample prepared by Focus Ion Beam (FIB) was damaged as illustrated in Figure S4a.Here, the distance between laser and sample was 40 cm without ND filter.To reduce the sample damage from focused laser beam, we placed an ND filter between the laser and the sample with a distance of 30 cm (d1), and adjusted the distances between the ND filter and the TEM sample of 30 cm (d2).
To inspect oxygen-related defect in LI 2H-MoTe2 ML, we constructed and performed several atomic simulations for 2H-MoTe2 MLs taking into consideration potential atomic defects (Figure S5).
Concurrently, we incorporated ABF-STEM simulation approach that is highly sensitive for low Z-element (Z = 8) such as oxygen defects present at the Te vacancy site.The HAADF-STEM image is expected to show minimal contrast at the Te vacancy defect sites.For a VTe1+1O, we could hardly detect oxygen contrast in the HAADF-STEM simulation image, due to the present of one Te atom at the same atomic column as oxygen atomic defect.However, a close-up look at the ABF-STEM reveal slightly enhanced contrast as indicated by red arrow in Figure S5c.For VTe2+1O and VTe2+2O, the ABF-STEM simulation clearly revealed enhanced oxygen contrast, since the two Te atoms are missing at the VTe2 defect site (Figure S5e-f).

Supporting Text 4. Experimental HAADF/ABF-STEM analysis of oxygen plasma-treated 2H-MoTe2 ML.
To directly verify the presence of oxygen atoms at the Te vacancy-site, we also analyzed HAADFand ABF-STEM images of 2H-MoTe2 MLs exposed to oxygen plasma.Previous studies have demonstrated that exposing 2H-MoTe2 flakes to oxygen-rich conditions, e.g., oxygen plasma, results in the creation of Te vacancies with subsequent adsorption/chemisorption of oxygen at these sites resulting in p-type characteristics [1][2][3] .As expected, we could identify a weak atomic contrast at the VTe2 site, consistent with our finding in LI 2H-MoTe2 ML.   Figure S10a presents the architecture of FCN to classify Te defect as an example.The configuration of FCN is divided into an "encoding" module and a "decoding" module, respectively.The encoding module consists of maxpooling layer, convolution layer, batch normalization layer, and PReLU activation function.The decoding module is changed to a transposed convolution layer instead of a maxpooling layer in the encoding module.Feature concatenation is used to combine feature maps in the encoding module in the same stage with the decoding module.The point defect types are classified through encoding and decoding process.Either Te defect or Mo defect can be identified.Figures S10b-c S11a illustrates estimated point defects and corresponding ground truth (labelled by human), respectively.In Figure S11b, we present the overall summary of species and concentration of point defects; VTe1 (red) with concentration of 0.58×10 14 /cm 2 (total analyzed area of 1.38×10 15 /cm 2 (13.8 nm 2 ). Figure S11c is a confusion matrix of point defect classification performances; the defect classification accuracies for each defect species were 100.0%(Perfect), 100.0%(VTe1) of total 202 classifications.Figure The top and bottom panels in Figures S12a-b illustrate the estimated point defects and corresponding ground truth (labelled by human), respectively.The orange arrows in top panel are misclassified to Perfect; ground truth of Tead2.All the color-codes for defect species are the same as Figure 2. Figure S12c is a confusion matrix of point defect classification performances; the defect classification accuracies for each defect species were 100.0%(Perfect), 100.0%(VTe1), 100.0%(VTe2), 80.0% (Tead1), and 63.6% (Tead2) of total 297 classifications.Figure S12d corresponds to device characterization (back gated-FET) of VA 2H-    respectively.The (i) orange, (ii) light-green, (iii) red arrows in top panels are all misclassified to Perfect; ground truth of (i) Tead2, (ii) Tead1, and (iii) VTe1+1O, respectively.All the color-codes for defect species are the same as Figure 3. Figure S14d is a confusion matrix of point defect classification performances; the defect classification accuracies for each defect species were 100.0%(Perfect), 97.5% (VTe1+1O), 100.0%(VTe2+2O), 85.7% (Tead1), and 88.0% (Tead2) of total 477 classifications.In Figure S15e, we present the overall summary of species and distribution of point defects (left), and corresponding device characterization (back gated-FET) of PT 2H-MoTe2, exhibiting strong p-type character (green).It is evident that oxygen plasma treatment generates VTe1+O (blue), Tead2 (light-blue), Tead1 (sky-blue), VTe2+2O (lightgray).The most dominant defect species is (i) VTe1+O, followed by (ii) Tead2, (iii) Tead1, and (iv) VTe2+2O, with the concentration being (i) 1.06×10 14 /cm 2 , (ii) 0.60×10 14 /cm 2 , (iii) 0.160×10 14 /cm 2 and (iv) 0.14×10 14 /cm 2 respectively (total analyzed area of 3.68×10 15 /cm 2 (36.8 nm 2 ).For VA 2H-MoTe2 and LI 2H-MoTe2, the classification accuracies for Tead2 and VMo were the lowest; 63.6% and 83.3%, respectively.For a close-up view of VA 2H-MoTe2 (left and middle panel in Figure S15a), the orange arrow denotes the mis-classification to Perfect; the ground truth of Tead2.These trends are often observed as illustrated in confusion matrix (Figure S12).To address this confusion of FCN for Tead2 vs Perfect type in VA 2H-MoTe2, we profiled the intensities in HAADF-STEM of (i) simulated Tead2 (orange open squares), (ii) simulated Perfect (gray open squares), and (iii) experimental Tead2 (orange solid line), which extracted from white dotted diagonal rectangles (Figure S15c).By comparing the intensities, the contrast of "experimental Tead2 site" locates between "simulated Tead2" and "simulated Perfect type", though it reaches to the simulated Tead2.These contrasts variation at Tead2 may contribute the confusion for FCN between Tead2 vs Perfect types.Also, the absolute number of Tead2 observed in our experimental images i.e. 11 counts can contribute to the superficially-lower accuracies for Tead2.
For the LI 2H-MoTe2, the lowest accuracies were found with VMo.In left and middle panels of Figure S15b, the blue arrow denotes the mis-classified to Perfect; the ground truth of VMo.Obviously, as indicated by white circles, there exist a Moint in a blue unit cell (ground truth) i.e., there exist couple of VMo and Moint, in a unit cell.The FCN model essentially confuses the defect types since we targeted to define the one-type-of defect in a unit cell.The pair of VMo-Moint defects in a unit cell was also confirmed in Figure 3b, which was revealed by 532-nm laser-illumination.As mentioned in the conclusion section, the interstitial defects are challenging due to the random-spatial distribution in a crystal matrix.Also, the absolute number of VMo was very small, since VMo is less likely to be transpired than VTe1 as mentioned in results section.These two types of low accuracies for Tead2 (VMo) for 2H-VA 2H-MoTe2 (LI 2H-MoTe2) can be improved by compiling the imaging data and database.More dependable analytic results would be achieved and expanded to more types of point defect types.
Figure S1 shows a experimental HAADF-STEM image of Tead1, b (c) and simulated HAADF-STEM images of Tead1 (Moad1).The intensity profiles shown in d are extracted from a-b light-green solid, and c light-gray solid diagonal rectangles, respectively.The corresponding atomic models are presented at the bottom for clarity.Based on intensity profiles, the experimental result (open light-green squares) is comparable with that of Tead1 (light-green solid line).

Figure S1 .
Figure S1.Inspection of Tead1 defect.a Experimental HAADF-STEM image of Tead1 observed in VA 2H-MoTe2 ML. b-c (Top panels) Simulated HAADF-STEM image of Tead1 and Moad1 respectively.The defect sites are indicated by the colored arrows.(Bottom panels) Corresponding atomic models as top panels.Scale bars; 0.2 nm.d Experimental (open light-green squares) and simulated (light-green/light-gray line) HAADF-STEM intensity profiles extracted from a-c along the diagonal rectangles.We confirm the defect observed in the experimental HAADF-STEM image as Tead1.The color-codes are the same as in Figure 2, where Mo-related defects are marked in light-gray for comparison.

Figure
Figure S2 shows a experimental HAADF-STEM image of Tead2, and b (c) simulated HAADF-

Figure S2 .
Figure S2.Inspection of Tead2 defect.a Experimental HAADF-STEM image of Tead2 observed in VA 2H-MoTe2 ML. b-c (Top panels) Simulated HAADF-STEM image of Tead2 and Mo adatom on Mo; Moad2, respectively.The defect sites are indicated by the colored arrows.(Bottom panels) Corresponding atomic models of the simulated HAADF-STEM images in the top panels.Scale bars; 0.2 nm.d Experimental (open orange squares)

Figure
Figure S3 shows a experimental HAADF-STEM image of VMo & Moint, and b (c) simulated

Figure S3 .
Figure S3.Inspection of VMo & Moint.a Experimental HAADF-STEM image of VMo & Moint pair observed in LI 2H-MoTe2 ML. b-c (Top panels) Simulated HAADF-STEM image of VMo & Moint.The defect sites are indicated by the colored arrows.(Bottom panels) Corresponding atomic models as top panels.Scale bars; 0.2 nm.d Experimental (open purple squares) and simulated (solid dark yellow/purple) HAADF-STEM intensity profiles extracted from a-c along the diagonal rectangles.We confirm the defect observed in the experimental HAADF-STEM image as VMo & Moint pair.The color-codes are the same as in Figure 2, where Teint are marked in dark yellow for comparison.

Figure
Figure S4illustrates the experimental set-up to optimize the 532-nm-laser illumination set-up.

Figure
Figure S4. a (Top) Experimental set-up for laser-illumination before optimization to illuminate the TEM sample.(Bottom-left) ((Bottom-right)) Cross-sectional FIB sample of bulk 2H-MoTe2 before (after) illuminated with focused laser.Even for the bulk 2H-MoTe2, sample was hugely damaged.b Optimized laser-illumination set up for TEM sample by inserting ND filter with adjusting distance between laser source-ND filter (d1) and ND filter-TEM sample (d2).The distance changes result in defocus to the TEM sample, which suggests the mild laserillumination condition to monolayer 2H-MoTe2 sample.On the right is a schematic illustration of the experimental set-up.

Figure S5 .
Figure S5.Inspection of oxygen-related defects.a-f (Top-left)-(Top-right) Simulated HAADF-(grey scale) and ABF-(false color) STEM analysis of a Perfect (control group), b VTe1, c VTe1+1O, d VTe2, e VTe2+1O, and f VTe2+2O, respectively.(Bottom-left)-(Bottom-right) Corresponding atomic models and intensity profile as top panels.From all intensity profiling, oxygen contrast is clearly detectable in the ABF-STEM images of VTe2+1O and VTe2+2O.The VTe1+1O shows negligible oxygen contrast compared to VTe1 and VTe1+1O due to the presence of Te atom in the same column as oxygen-related defect.Scale bars; 0.2 nm.

Figure
Figure S6.a-b (Top panels) ((Middle panels)) Experimental (simulated) HAADF-and ABF-STEM images of LI 2H-MoTe2 ML, with VTe1+1O indicated by red dotted circles.Experimental (open red squares) and simulated (solid red lines) HAADF-(ABF-)STEM intensity profiles extracted from a-b along the diagonal rectangles are shown below the simulated images.c-d (Top panels) ((Middle panels)) Experimental (simulated) HAADF-and ABF-STEM images of LI 2H-MoTe2 ML with VTe2+2O indicated by yellow dotted circles.Experimental (open yellow squares) and simulated (solid yellow lines) HAADF-(ABF-)STEM intensity profiles extracted from c-d along the diagonal rectangles are shown at the bottom.The oxygen contrast and intensity ratio for VTe1+1O is diminished.However, the oxygen contrast is well pronounced at the VTe2+2O site.For clarity, the atomic models embedded with VTe1+1O and VTe2+2O are presented at the bottom of ABF-STEM intensity profiles.Scale bars; 0.2 nm.

Figure
Figure S7.a-b (Top panels) ABF-STEM intensity profiles for VA 2H-MoTe2 ML with VTe2 and LI 2H-MoTe2 ML with VTe2+2O, respectively.The corresponding atomic models with the same color-codes as Figures 2-3 are presented at the bottom.Scale bars; 0.2 nm.c ABF-intensity profile of VA 2H-MoTe2 (open yellow squares) and LI 2H-MoTe2 (filled yellow squares).The weak contrast attributed to chemisorbed oxygen at the Te vacancy site (filled yellow arrow) is only evident in LI 2H-MoTe2.

Figure
Figure S9 illustrates the point defect classification workflow by deep learning models.In the deep presents the train/validation loss curves during FCN training for Te defect and Mo defect types.Note that the train/validation loss values are very low even at the early stage of training, which means learning process fits well to the point defect analysis.The left panels of Figures S11d-e are input HAADF-STEM images with randomly distributed point defects in crystal matrices.Compared to the ground truth of input images (middle panels), the trained-FCN models identify the defect types in either Te on-site or Mo on-site defects, as illustrated in the right panels in Figure S10 and as mentioned in Figure S9.

Figure S9 .
Figure S9.Schematic representation of deep learning method.Deep learning (DL) processing: Input STEM images are fed to each of the three deep learning models.Post-processing: Point defect classification results of Te defect and Mo defect, each image is fed to Faster R-CNN to detect unit cell and unit cell locations.With the combination of Te defect and Mo defect for each unit cell area, the final point defect types are determined.Finally, the classification results corresponding to each types are overlaid on the input STEM images.

Figure
Figure S11 depicts application of automatic point defect classification algorithm to the

Figure
Figure S10.a Fully Convolution Network (FCN) that segments defect types in the Te on-site.The configuration of the network is divided into an "encoding" module and a "decoding" module, respectively.The encoding module consists of maxpooling layer, convolution layer, batch normalization layer, and PReLU (Parametric Rectified Linear Unit) activation function.The decoding module is changed to a transposed convolution layer instead of a maxpooling layer in the encoding module.Feature concatenation is used to combine feature maps in the encoding module in the same stage with the decoding module.b-c The loss curves for train/validation for Te on-site and Mo on-site defect classifications by FCN model.d-e (Left panel) Input HAADF-STEM images for defect examination with randomly distributed point defects.(Middle panel) Ground truth of left panels.(Right panel) The predicted point defect types by trained FCN models.The color-codes are the same as Figures 2-3, except the Perfect type (blue).Scale bars; 0.5 nm.

Figure
Figure S11.a (Top) Estimated point defect classification results of pristine 2H-MoTe2 ML.The color-codes are the same as in Figure 2. (Bottom) Corresponding ground truth of top panel.b (Top) Defect concentration in pristine 2H-MoTe2 revealed by deep learning.c Confusion matrices for point defect estimation performance of total 202 input unit cells of pristine ML 2H-MoTe2.d (Bottom) Transfer (I-Vg) plot for pristine 2H-MoTe2 ML exhibiting weak n-type character by VTe1 as revealed by a.

Figure S12 .
Figure S12.Point defect classification performance evaluation of VA 2H-MoTe2 MLs.a-b (Top panel) Estimated point defect classification results by deep learning for VA 2H-MoTe2 ML. (Bottom panel) Corresponding ground truth of top panel.The orange arrows in a and b are misclassified to Perfect; ground truth of Tead1.Note that Perfect ones are not indicated though they were classified by deep learning.Color-codes are all the same as Figures 2-3 for defect types.Scale bars; 0.5 nm.The interstitial defects are denoted by white circles.c Confusion matrices for point defect estimation performance of total 297 input unit cells of vacuumannealed ML 2H-MoTe2.d Transfer (I-Vg) plot for VA 2H-MoTe2 ML exhibiting strong n-type character (black) compared to pristine 2H-MoTe2 (weak n-type, grey) by defined defect types.

Figures
Figures S13 presents additional statistical point defect classification results of LI 2H-MoTe2.The

Figure S13 .
Figure S13.Point defect classification performance evaluation of LI 2H-MoTe2 MLs.a-d (Top panels) Estimated point defect classification results by deep learning for LI 2H-MoTe2 ML. (Bottom panels) Corresponding ground truth of top panels.The blue arrow in d is misclassified to Perfect; ground truth of VMo.Note that Perfect ones are not indicated though they were classified by deep learning.Color-codes are all the same as Figures 2-3 for defect types.Scale bars; 0.5 nm.The interstitial defects are denoted by white circles.e Confusion matrices for point defect estimation performance of total 616 input unit cells of LI 2H-MoTe2 ML. f Transfer (I-Vg) plot for LI 2H-MoTe2 ML exhibiting strong p-type character (green) compared to pristine 2H-MoTe2 (weak n-type, gray) by defined defect types

Figure S14 .
Figure S14.Point defect classification performance evaluation of PT 2H-MoTe2 MLs.a-c (Top panels) Estimated point defect classification results by deep learning for PT 2H-MoTe2 ML. (Bottom panels) Corresponding ground truth of top panels.The orange arrow in b is misclassified to Perfect; ground truth of Tead2.Similarly, the green, yellow, and red arrows in c are all misclassified to Perfect; ground truth of Tead1, VTe2+2O, and VTe1+1O, respectively.Note that Perfect regions are not indicated though they were classified by deep learning.Color-codes are all the same as Figures 2-3 for defect types.Scale bars; 0.5 nm.The interstitial defects are denoted by white circles.d Confusion matrices for point defect estimation performance of total 477 input unit cells of PT 2H-MoTe2 ML. e (Left) Statistical point defects classification by deep learning in PT 2H-MoTe2.VTe1+1O (blue), Tead2 (light-blue), Tead1 (sky-blue), VTe2+2O (light-gray).(Right) Transfer (I-Vg) plot for PT 2H-MoTe2 ML exhibiting strong p-type character (green) by defined defect types.

Figure
Figure S15.a-b (Left) and (right) Estimated point defects and ground truth of VA 2H-MoTe2 and LI 2H-MoTe2, respectively.Scale bars; 0.1 nm.The color-codes are the same as Figures 2-3.Note that the orange (blue) arrow indicates the mis-classified to Perfect; while the ground truth of Tead2 (VMo) for VA 2H-MoTe2 (LI 2H-MoTe2).The white circle denotes the Moint.c Intensity profile extracted from white dotted diagonal rectangles in a. Orange solid line; experimental intensity profile of Tead2, Open orange (gray) squares; simulated intensity profile of Tead2 (Perfect).