Reviewed by Frances BriggsJun 24 2025
In a new study published in Chemical Science, researchers at the University of Rochester have developed an algorithm that shows enormous potential for incorporating machine learning and artificial intelligence into battery research.
Image Credit: Connect Images - Curated/Shutterstock.com
Scientists and engineers often use density functional theory (DFT) to study the atomic interactions of material surfaces to develop more energy-efficient batteries, capacitors, and other electrical devices.
However, DFT simulations have a massive computational cost and a large configuration space, which limit their application in understanding multi-reactant catalysis of geometrically diverse surfaces.
Currently it’s prohibitive and there’s no supercomputer in the world that can do an analysis like that. We need clever ways to manage that large data set, use intuition to understand the most important interactions on the surface, and apply data-driven methods to reduce the sample space.
Siddharth Deshpande, Assistant Professor, Department of Chemical Engineering, University of Rochester
By evaluating the structural similarities among different atomic configurations, Deshpande and his students discovered they could accurately capture the underlying chemical processes by analyzing just two percent or less of all possible surface interactions.
Building on this insight, they developed an algorithm that significantly streamlines the analysis, detailed in the recent Chemical Science publication.
In the study, the team applied the algorithm to investigate, for the first time, how defects on a metal surface influence the carbon monoxide oxidation reaction—a process relevant to understanding energy losses in alcohol fuel cells.
Deshpande explains that their approach enhances density functional theory (DFT), a widely used computational method he describes as the “workhorse” for materials modeling over the past several decades.
This new method may become the building ground to incorporate machine learning and artificial intelligence. We want to take this to more difficult and challenging applications, like understanding the electrode-electrolyte interference in batteries, the solvent-surface interactions for catalysis, and multi-component materials such as alloys.
Siddharth Deshpande, Assistant Professor, Department of Chemical Engineering, University of Rochester
Journal Reference:
Zeng, J. et al. (2025) A structural similarity based data-mining algorithm for modeling multi-reactant heterogeneous catalysts. Chemical Science. doi.org/10.1039/D5SC02117K.