Computational Methods for Modeling Reaction Dynamics in Batteries and Catalysts

Graeme Henkelman of the University of Texas at Austin is supported by an award from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry to develop computational methods to understand the function of materials employed for the production and consumption of energy. This research is at the atomic scale and such a fundamental understanding will not only allow us to determine limitations of existing materials, but also consider new materials that have the potential to make energy production and usage more efficient.

The second part of this work will incorporate tools from computer science, including machine learning, to improve the efficiency of our computational methods and accelerate the design of new materials. This project is expected to have a positive impact on the scientific community by providing these new computational tools as well as contribute to the growth and learning of the graduate and undergraduate students who will be developing these tools.

In this project, Graeme Henkelman and his research group are setting out to develop computational methods to model the reaction dynamics in materials related to energy applications, including batteries and catalysts. More specifically, they seek to improve the efficiency of methods based upon transition state theory so that dynamics over experimental time scales can be modeled using forces and energy from density functional theory (DFT).

In order to mitigate the computational expense of DFT a number of strategies will be followed to make the calculations as efficient as possible. First, a public kinetic database will be established with geometric information of reaction mechanisms, which can be used to propose transition states from a query structure to the database. Second, machine learning (ML) methods will be used to accelerate our calculations.

Instead of fitting global potential energy surfaces, ML methods will be used to fit local energies, forces, and curvature to accelerate DFT calculations rather than replace them. Specifically, the ML models will: (1) provide a pre-conditioner for the optimization of minima and saddle points; (2) suggest local minima to facilitate the search for stable materials; (3) efficiently construct a hyperdynamics bias potential for modeling of rare events in energy materials directly with DFT.

As well as using ML methods, this project will investigate how the choice of ML method, including neural networks and Gaussian processes, as well as the numerous hyperparameters, determine the shape and quality of the potential energy surface. Finally, methodology to take trajectories of catalytic reactions will be developed to build reaction networks, from which the overall activity of a catalyst or battery material can be understood.

A broader impact of this research to the scientific community will be in the form of software that is freely distributed. In addition, a team of undergraduate students will be a part of this research and to use these tools.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.