FAIR Chemistry ML-Driven Electrocatalysis ScreeningNew

Background

Electrocatalyst screening by DFT is computationally expensive: each candidate surface-adsorbate configuration needs relaxation plus energy evaluation. For alloy and dopant searches the combinatorial space quickly outruns DFT budgets.

Approach

• Train GNN models (GemNet-OC, eSCN, EquiformerV2) on OC20 relaxations
• Replace the DFT relaxation loop with ML relaxation plus final DFT validation of shortlisted candidates
• Use the same models for dynamics trajectories on reactive surfaces

Reported gains

• Orders-of-magnitude speedups for adsorption-energy and relaxation tasks versus DFT, with accuracy targets set by the OC20 IS2RE/IS2RS metrics

Use cases

• Alloy catalyst composition screening for ORR/OER/HER
• Adsorbate configuration enumeration at scale
• Downstream DFT refinement budgets focused on top candidates

Resources

• FAIR Chemistry repository with pretrained checkpoints
• Open Catalyst papers describing dataset construction and metrics