Our goal is to build predictive models for gas/solid heterogeneous catalytic systems. We want to go beyond many of the approximations routinely used in simulating such systems. Specifically, we want~to:

  • Improve the accuracy of both enthalpic and entropic contributions in thermodynamic functions of adsorbates at minima and first-order saddle points.

  • Advance the state of the art in modeling the kinetics of catalytic reaction systems by creating automated frameworks with more realistic physical representations.

  • Develop methods applicable to metals and their alloys.

  • Create and extend exascale-ready computational tools that can achieve our scientific goals and help others to adopt our approaches.


Heterogeneous catalysis plays a key role in energy technology and in the chemical industry in general. It is used widely to upgrade heavy fossil fuels, enable the partial reduction of bio-derived feedstocks, and convert small molecules (CO, CO2, CH4, etc.) into larger and more valuable compounds. Typical catalysts used in these processes are metals, their alloys, or metal oxides, often in a nanoparticle form. Catalytic reactions of a gas feedstock over these surfaces are fundamentally complex: they involve many reactions among gas-phase and adsorbed species, and the yields of desired products depend on temperature, pressure, composition and, of course, the nature of the catalyst. The key to understanding, predicting, and eventually optimizing these processes is to develop multiscale modeling approaches, combine cutting-edge theory with modeling, and use large-scale computational resources to handle the requisite problem complexity. Predictive models will help develop green technologies through better biomass utilization, foster sustainable technologies, enable better emission control, and contribute to energy security.


We are developing new beyond-DFT theories for periodic systems to enable accurate, dual-level potential energy surface evaluations in an exascale-enabled electronic structure package, forming the foundations of our thermodynamic and kinetics work. Our machine-learning enhanced methods to include the combination of anharmonic and coverage effects in partition function calculations, along with our advanced automated and highly parallelized configuration and reaction pathway search strategies, will allow us to calculate accurately thermodynamic and kinetic properties for heterogeneous catalytic processes at very large scales. We are advancing our automated microkinetic model generator to incorporate these and further effects, such as site-specificity, which ultimately will permit us to go beyond the mean-field approximation in our modeling paradigm by using a smart adaptive kinetic Monte Carlo model informed by a combination of our microkinetic framework and direct ab initio calculations. Intelligent feedback loops will be created by a suite of sensitivity analysis and uncertainty quantification tools. Our work will yield codes that are open-source and can run effectively on current peta- and upcoming exascale resources, taking advantage of Graphics Processing Unit/Many Integrated Cores (GPU/MIC) architectures.

quantum chemistry

Traditionally in heterogeneous catalysis DFT methods are used to characterize the PES, which put a limit on the accuracy of the models derived from them. Unlike in gas-phase chemistry, there are no well-established higher-level methods. A major goal of this program is to provide new exascale method developments and associated tools for catalytic reactions at surfaces with the latest DFT methods (e.g. new meta-GGA hybrid DFT, and dispersion corrected functionals) and beyond DFT methods (e.g. RPA methods for metal surfaces).


We have previously developed the methodology to achieve accurate adsorbate thermochemistry beyond the harmonic oscillator and free 2-D gas limits under low coverage conditions. Now we go beyond these methods in by expanding our current methodology for anharmonic partition functions to include the effects of co-adsorbates, or lateral interactions, and for saddle points and to enable more accurate predictions for rate constants. We are using a sparse representation of the developed thermodynamic data-sets from these techniques to train a transferable estimate of these dependencies based on molecular descriptors.


Our pynta code is able to search for reaction pathways in a systematic way for simple metal facets in the low-coverage limit. We are expanding this methodology to be able to investigate the effect of co-adsorbates on the reaction pathway energies and ultimately on reaction rate coefficients. Our algorithms will allow the characterization of adsorbates and their reaction pathways on crystal facets in the presence of co-adsorbates. In addition, we are implementing and using free energy techniques to complement and benchmark the systematic approach.

microkinetic model

We us the Reaction Mechanism Generator (RMG) code to build microkinetic mechanisms for H/C/N/O-containing adsorbates on arbitrary metallic catalysts. At present, RMG only considers surfaces in the low-coverage limit, and it only considers a single type of site within the catalyst. RMG will be significantly enhanced to address both of these limitations, so that it can build accurate, predictive microkinetic mechanisms with considerably more chemical complexity. We also improve upon the current hand-crafted decision tree estimators for adsorbates' thermochemical and kinetic parameters and develop APIs for portions of RMG for use in adaptive KMC simulations.

mesoscopic model

There is fundamental physical chemistry insight to be gained from studying coverage-dependent properties. However, to put the learned relations fully to work we need to go beyond the mean field approximation of surface kinetics. We target mesoscopic modeling of surface kinetics using kinetic Monte Carlo (KMC) methods on a lattice. One of the challenges we tackle is the scalability of computational performance on advanced architectures, where we rely on Kokkos-based GPU-enabled open-source KMC code with fine-grain parallelism. We also develop automated exploration of system response using ML techniques.

model analysis

Model analysis tools are required to close the loop in our simulation paradigm shown in the figure higher up on this page. These tools enable the adaptive creation of reaction rate coefficient dictionaries for KMC using RMG and ab initio capabilities, enable uncertainty quantification (UQ) in RMG and in KMC relying on multifidelity methods.