CatalysisMat
Learning adsorption energies for catalyst discovery
Catalyst design hinges on adsorption energies, but computing them with DFT for every candidate is expensive. This project trains graph neural networks (ALIGNN) to predict adsorbate-on-slab binding energies across a dozen diverse datasets — surface energies, oxygen-reduction and CO₂-reduction intermediates, chemisorption energies, and large Open Catalyst (OC20) initial-to-relaxed-energy tasks.
The work characterizes where ML surrogates reach near-DFT accuracy and where they plateau, providing guidance on data requirements for reliable catalyst screening. It connects to my earlier work on high-entropy fuel-cell catalysts, published in Nano Futures (2025).