C&EN West Coast News Bureau

Computational chemists simulate the behavior of the microscopic. Chemical engineers wrest performance from the macroscopic. One would think the twain should never meet.

But computational chemistry is, in fact, proving to be an extremely useful tool for engineers. The insights gleaned from simulations of the atomistic and molecular wrangling of polymerizations or catalytic reactions can, in turn, shed a surprising amount of light on why an industrial process clogs up or how one auto lubricant is better than another.

In industry, where engineers simply can't spend the money or time to perform every experiment that might improve their process or verify if a plant design is going to work, modeling fills an extremely valuable niche.

That wasn't always the case. Only in recent years have computers become routinely powerful enough, and theory sophisticated enough, to deal with the monstrous complexities of real-world applications. But now, for a relatively reasonable amount of money and frequently within hours or days, chemists can run calculations that ultimately yield predictions of macroscopic parameters such as reaction rates or thermodynamic properties.

Consequently, engineers, recognizing a powerful new tool, and computational scientists, recognizing a rich area of application, are beginning to make their way toward each other. Workshops on computational chemistry and chemical engineering are springing up. Engineers flocked to a symposium on the topic at the American Institute of Chemical Engineer's 1998 annual meeting in Florida. A conference on the Foundations of Molecular Modeling and Simulation--Applications for Industry is scheduled for the summer of 2000 in Colorado.

Most recently, at the American Chemical Society's national meeting in Anaheim, Calif., last month, computational chemists and engineers gathered for a symposium titled "Bridging the Gap between Chemical Modeling & Engineering." Sponsored by the Division of Computers in Chemistry, the symposium was organized by Joseph Golab, a chemist at BP Amoco, and Paul Mathias, a chemical engineer at Aspen Technology, Cambridge, Mass. Nearly 50 speakers representing a liberal mix of the two disciplines discussed their techniques, new methods, successes, and failures. They also looked to the future of the fields, some calling for extra training in chemical engineering for fledgling computational chemists, and vice versa.

"Obviously, chemistry and engineering have already collided," Golab says. "But the computational part is new."

Peter T. Cummings, chemical engineering professor at the University of Tennessee, Knoxville, and group leader in chemical technology at Oak Ridge National Laboratory, Oak Ridge, Tenn., says, "Where we have a significant impact on chemical engineering is in the early stages of process design." During the initial screening for a particular chemical process, most candidates are eliminated for cost reasons. That's also the time when least is known about the materials to be used, he says. "Many of the kinds of calculations we're capable of doing today can be useful in providing properties during the earliest level of design," he notes--properties such as density-phase equilibria, heats of formation, transport properties, or viscosities.

Computational chemists have at their disposal numerous techniques for simulating a wide variety of behaviors. For example, density functional theory (DFT) saves time and money by using electron densities to describe a system, rather than the motion of individual electrons. For extremely accurate but more expensive and time-consuming calculations, chemists may use various methods--such as Hartree-Fock and its variations--for explicitly solving the Schrödinger equation. [The developer of DFT, Walter Kohn, and creator of the widely used Gaussian computational chemistry program, John A. Pople, together won the Nobel Prize in Chemistry last year (C&EN, Oct. 19, 1998, page 12.)

Although numerous existing commercial software programs perform many different tasks, developing new aspects of already existing theory is vital to the progress of computational chemistry, says Andrew M. Rappe, assistant chemistry professor at the University of Pennsylvania. "It means designing our own programs and theory in addition to harvesting results," he says.

For theory and simulation to have a full impact on the development of new materials and processes, structures and properties need to be predicted before synthesis and experiment, says William A. Goddard III, chemistry professor at California Institute of Technology. Starting with electronic descriptions from quantum mechanics, "we want to describe fluid flow, crack initiation, rates, and side products from catalytic processes," Goddard says.

But to span these enormous length and time scales requires bridging the mesoscale gap between those two extremes, Goddard noted at the meeting.

Atomistic techniques deal with time and space scales up to perhaps 10 nanoseconds and nanometers. But between that scale and the macroscopic scale of meters and hours, there's a large space: the mesoscale region of microseconds and micrometers. Developing mesoscale models requires developing ways to average over groups of atoms.

For example, Goddard said at the symposium, Chevron Chemical in Richmond, Calif., is searching for new, improved wear inhibitors that work with lubricants to protect automobile engines. But determining the efficacy of any new candidate requires a laborious, full-scale engine test. Goddard describes the process: "Take the engine apart, measure camshaft tolerances, put the engine together, add oil and inhibitor, run it for 60 days, take it apart, remeasure the tolerances." To do this only once costs $30,000, and several tests to obtain good statistics brings the cost to $150,000, he says.

Enter computational chemistry. Goddard's group developed a model to simulate an oxidized iron camshaft covered with a monolayer of dithiophosphate wear inhibitors, added a lubricant, and calculated the dynamics of the metal surface, wear inhibitor, and lubricant system for the conditions that might occur in extreme conditions in a camshaft in contact with a lifter.

Mechanistic insights from the simulation included the discovery that for the best wear inhibitor, the lubricant molecules in the first layer stayed stuck to the wear inhibitor for long periods--50 to 100 picoseconds. However, with poorer wear inhibitors, the lubricant would stick and slip, stick and slip repeatedly. The study allowed the researchers to calculate friction coefficients and other properties used in continuum modeling of the macroscopic system.

Not only did their data from the simulation agree with experiment, but their calculations led to a model for predicting wear inhibitor performance that was used to study seven new compounds. This, in turn, led to the discovery of another potentially cheaper candidate with comparable performance.

Industrial time scales for producing results are "over coffee, over lunch, or overnight," Anne M. Chaka, a theoretical chemist at Lubrizol, Wickliffe, Ohio, quipped at the meeting. "Trial and error can be reduced by screening ideas before time and money are spent in the lab and in testing." Computational chemistry has proven extremely useful in this regard at her company, which manufactures fuel and lubricant additives.

For example, operations engineers at Lubrizol decided to switch from a batch process of alkylation to a continuous feed stream in order to reduce costs. Two catalytic processes could be used in this reaction: a Friedel-Crafts process, which is a lower temperature Lewis-acid-catalyzed reaction, or a Brönsted-acid-catalyzed reaction that produces less waste. The engineers wanted to know if it was possible to somehow couple the two catalytic processes together in a continuous stream to maximize the low-temperature, low-waste attributes.

Chaka pulled together a bunch of experts at Lubrizol, and discussed various possible mechanisms for the alkylation process. "It was rather exciting. We had people actually pushing at each other to get at the blackboard," she said. They modeled three strong candidates for the mechanisms--two for the Brönsted-acid-catalyzed process, and one for the Friedel-Crafts process--with ab initio methods to determine the reasonableness of each proposed mechanism and identify the rate-limiting steps along the pathways.

One of the three proposed mechanisms was quickly discarded. Understanding the details of the other two pathways indicated that there was little mechanistic similarity between the behavior of the Brönsted and Lewis acid catalysts, thus showing that using the two together would not have a synergistic effect. "There isn't any way you can combine them in a single process," she said.

Chemical engineers are also tackling issues surrounding supercritical carbon dioxide, a promising, environmentally friendly replacement for organic solvents. It dissolves organic substances that are typically intractable in water. But solvents that dissolve apolar organic contaminants also need to dissolve the polar substances such as water that come along with the organic waste. The answer, it would seem, would be a reversed version of the micelle. (Micelles are groups of surfactant molecules with polar heads and organic tails that arrange themselves tail-in around an oil drop.) So-called reverse micelles would have "CO₂-philic" tails and polar heads with which to cluster tail-out around a polar droplet.

But despite hundreds of screenings, only a few surfactants have been found that form reverse micelles. At the symposium, Cummings described modeling efforts that he hopes will help chemists understand how the surfactants work and aid the selection of possible new candidates. "We're trying to develop intuition as to why one class will work and another won't," Cummings says.

Sidebar: Fire-resistant polymer may be useful in airplanes

His group's molecular dynamics model simulates the behavior of a reverse micelle previously studied experimentally. The surfactant is a dual-chain surfactant [(C₇F₁₅)(C₇H₁₅)CHSO₄^-Na⁺], with two seven-carbon tails attached to a sulfate head group with a sodium counter ion. The researchers constructed a model of the surfactant by piecing together established models for each component of the molecule--one for each tail, one for the sulfate head group, and one for the sodium ion. To 30 surfactant molecules, they added 132 water molecules and 2,452 CO₂ molecules, and then simulated their behavior at temperatures and pressures representative of supercritical conditions.

When they ran the simulation, in just 1 nanosecond the disordered micelles had self-assembled into stable, spherical aggregates surrounding a water drop. "We were extraordinarily surprised at how well they worked," Cummings says.

In a simulation without surfactants, the water forms one large insoluble drop in equilibrium with a small number of molecules distributed throughout the CO₂. Once the surfactant is added to the simulation, the surfactants scavenge all of the water by forming micelles, as predicted. One surprising feature of the simulation, Cummings says, was that although water droplets came and went from the reverse micellelike aggregate and occasionally exchanged between micelles, once the water molecules found themselves out in the CO₂, they would beat a hasty retreat back into a self-assembled aggregate.

Cummings and coworkers are now doing a larger simulation with 45,000 atoms that exactly mimics experimental conditions. Stable aggregates are taking much longer to form, Cummings says. And that means much more computer time. "We burn up a lot of CPU hours on massively parallel computers," he says. "We're really at the leading edge of what we can do computationally at this point." They plan to simulate other reverse micelles and are looking at different strategies that would allow them to do the simulation for longer periods of time.

Nanoporous materials are becoming increasingly important in industrial processes, but with their random distribution of wormlike pores of various shapes and sizes, they've been difficult to model. Lev D. Gelb, a postdoctoral researcher in the North Carolina State University, Raleigh, lab of chemical engineering professor Keith E. Gubbins, described new inroads into the problem.

Controlled-pore glass can be designed with a large range of pore sizes and porosities. In most previous work, people have modeled such materials very simply--for example, if the material had cylindrically shaped pores, they were modeled by a single cylinder. But as Gelb points out, that excludes the effects of disorder and connectivity. To develop a more realistic picture of what's going on, he and his colleagues have developed a model that actually simulates the production of the glass.

In the real world, such glass is synthesized by heating a mixture of silica and boric acid to high temperatures. As this mixture is cooled, it separates into two interconnected phases, one of which is nearly pure porous silica. The pore size can be controlled by the amount of time spent cooling. The borate phase inside the pores is removed with acid. The key to Gelb's simulation is to ignore the complicated chemistry of the silicate and borosilicate and treat the model as a mixture of Lennard-Jones fluids, theoretical descriptions of rare gases. The researchers can justify this treatment because the network depends mostly on the viscosities and densities of the two liquids, and not on parameters such as the intermolecular potentials of the materials.

Their simulation of the mixture in a cell containing half a million atoms showed the development of pores that grew larger over time as the atoms rearranged themselves.

Their models may also shed light on substances inside the pores. "Fluids behave strangely in small spaces," Gelb says. "They show enormous shifts in critical temperatures and densities, and they also behave as if they're under enormous pressure, which people don't really understand."

For example, in tiny pores, vapors will condense onto the highly curved surfaces at temperatures and pressures where they wouldn't ordinarily be liquids. This phenomenon is known as capillary condensation and is a result of surface tension and the attraction between the pore walls and the vapor molecules. The group's simulations also show this.

Gelb and Gubbins want to use these techniques to generate models of other materials, such as sol-gel materials and activated carbons, as well use them as a test to screen potential new materials and methods.

More often than not, chemical processes involve catalysts. But how catalysts work is often mysterious. Numerous computational efforts are centering around modeling the behavior of catalysts to help increase their efficiencies, to design new ones, or to improve old ones.

For example, Matthew Neurock, chemical engineering professor at the University of Virginia, Charlottesville, has developed a hybrid computational method using DFT and kinetic Monte Carlo algorithms to simulate the behavior of catalysis, such as the selective hydrogenation of maleic anhydride to tetrahydrofuran on different metal surfaces. They've been able to predict a number of key quantities for various steps in the reaction, including chemisorption energies and activation barriers.

It's been known experimentally that interfacing one metal with another can markedly change their catalytic surface properties. But there's been no way to predict a priori how they'll be affected. The researchers' kinetic studies on ethylene and maleic anhydride hydrogenation on different bimetallic palladium surfaces have led to a general description of how the electronic structure of the surface controls adsorption and activation energies for different bimetallic surfaces. "You can begin to think about interesting ways to alter the surface to modify its catalytic behavior," Neurock says.

They have found that their models generally agree with experiment to within 5 kcal per mole, which, he says, can provide relative trends. In many cases, knowing these trends is useful to engineers.

"It's important to keep pushing to get extremely accurate results, but you can also gain ideas in terms of direction and what is controlling the chemistry. First-principle quantum mechanical methods provide trends that are fairly reliable. That's what we're after right now," Neurock says.

William F. Schneider, a theoretical chemist at Ford Research Laboratory, Dearborn, Mich., also models catalytic behavior. He's already had success in modeling the behavior of copper zeolites for reducing emissions of nitrogen oxides (NO_x). Now he and coworkers are tackling another difficult problem, that of trying to understand the chemistry of materials used to temporarily trap NO_x for later reduction.

These NO_x traps are composed of two principal components--a noble-metal catalyst that oxidizes NO to NO₂, and a barium oxide adsorbate, which reacts with the NO₂ and additional O₂ to form a barium nitrate. The trap is periodically purged by running the chemistry in reverse, liberating the NO that is later reduced. But this system breaks down in the presence of sulfur. During combustion and catalytic oxidation, any sulfur in the diesel or gasoline fuel is converted to SO₂ and SO₃. The barium oxide has a high affinity for these sulfur species, and their buildup eventually clogs the trap.

Because a metal oxide is an extended structure, and not a discrete molecule, Schneider and coworkers are employing two approaches to simplify the simulation of a big chunk of matter. One, known as the supercell approximation, represents an oxide surface as an infinite slab compartmentalized into cells several atomic layers thick and perhaps five atomic rows on a side. This method more realistically represents the bulk properties of a material, but it's computationally expensive. Another, known as the cluster method, is a more approximate representation of a small section of the metal-oxide surface, containing only a handful of formula units. The advantage of this technique is that these small systems can be more cheaply and accurately modeled. To describe the bigger environment of the bulk system, one could do a series of cluster calculations from small to large and extrapolate, or--in what's known as the embedded cluster method--model a cluster embedded in a field that simulates the bulk material, he says.

Schneider's models confirm experimentalists' difficulties. They show that the SO₂ strongly reacts with the surface and that the primary locus of the reaction is at the oxide site. The simulation provides a detailed view of the reaction in a series of "snapshots": An SO₂ molecule moves toward the surface of a unit cell of metal oxide and a sulfur atom grabs an oxygen from the surface and ultimately ends up locked onto the surface as a sulfite.

The process is even more vigorous at defect sites--steps, corners, or bumps on the surface. Such imperfections are extremely common. "If you really want to realistically look at reactivity, you have to look at those types of defects," Schneider says.

Modeling oxygen as it spontaneously dissociates and forms rows of atomic oxygen on a rhodium surface has provided Rappe with an explanation of a long-standing problem: Oxygen adsorption inexplicably stops at half coverage. Intuitively, the reaction should proceed until the surface is completely covered, but "there are entire rows that it refuses to occupy," Rappe says.

Reasoning that there must be some unknown kinetic barrier, he and postdoctoral researcher Steven P. Lewis and graduate student Eric J. Walter developed a model to study what happens to the system as oxygen molecules adsorb. They identified the transition state associated with oxygen dissociation on the surface. As a result of repulsive interaction with neighboring oxygen atoms, they saw that the energy of this transition state increases as the oxygen coverage increases around the dissociation site. When half the site has oxygen atoms, the energy required for O₂ to dissociate is more than that of a free molecule.

The group also studies vibrational decay of adsorbates on surfaces. How rapidly molecules return to equilibrium and by what mechanism is important in surface processes such as catalysis. "If you could run a catalytic process at lower temperature, that would be advantageous," Rappe says.

Their methods allowed Rappe and Lewis to resolve a theoretical conundrum about the vibrational motion of carbon monoxide on a copper surface. Experimental values for the vibrational decay times of this system were very short, but when theorists tried to model the system, they calculated a much longer time. Even though they were using the correct mechanism, their time was an order of magnitude too long. Rappe's group showed that adsorbate-adsorbate interactions through the substrate play a critical role, so that the ordered overlayer of molecules vibrationally couples to the underlying surface in a completely different way than a single molecule does. "So the answer they got was right for a single molecule, but a single molecule and a layer are so different," Rappe says.

Meanwhile, David A. Dixon, a theoretical chemist at Pacific Northwest National Laboratory, Richland, Wash., is pushing the accuracy envelope. His concern is to try to get the accuracy of his predictions down to less than 1 kcal per mole, which is needed in many engineering applications.

Accuracy is particularly critical in calculating values for quantities that have never been measured experimentally. As a foundation, Dixon is using calculation methods developed by his colleagues, theoretical chemists Thom H. Dunning Jr. and David F. Feller, that can be extrapolated to a complete basis set limit--that is, that include functions that describe the behavior of every electron and cover space as broadly as possible.

But to be extremely precise, a number of mathematical terms that are small but important must be included in the description, Dixon says.

Sidebar: Light makes membranes mighty

For example, corrections need to be made for the interactions of core electrons with valence electrons. In the first-row atoms, the 1s orbital will interact slightly with 2s and 2p orbitals. And relativistic effects, stemming from some electrons traveling around the nucleus at close to the speed of light, also come into play. "If you do all those corrections, you can talk about highly accurate thermochemistry," Dixon says. Of course, the limitations, as always, are cost and computer time.

Dixon's group has already calculated a number of quantities useful to chemical engineers. They've calculated the heats of formation of some fluorocarbons, for example, and of the silicon atom. "You really can get the kind of numbers that go to process simulation," he says.

As these types of applications become more commonplace, budding chemists and engineers should be more roundly trained from the start, says Stuart W. Churchill, emeritus chemical engineering professor at the University of Pennsylvania. He calls for chemistry students to gain greater understanding of the other discipline by taking a course or two in chemical engineering.

Golab agrees, saying in the future, perhaps both chemical engineers and computational chemists alike will acquaint themselves with each field.

The application of theory will always be the primary goal of chemical engineers, Golab says. "I think that's really where the fun is going to be for industry," he says. "They've been realizing chemical modeling is finally going to work out."

Chemical & Engineering News
Copyright © 1999 American Chemical Society