Few-Shot Ensemble Learning for Catalysis and Application to Trimetallics for Oxygen Reduction

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine-learned interatomic potentials (MLIPs) accelerate catalyst discovery but often lack accuracy and reliable uncertainty quantification when applied to new systems.
• This limitation hinders their effective use in high-throughput catalyst screening.

Purpose of the Study:

• To improve the accuracy and provide reliable uncertainty quantification for MLIPs in catalyst screening.
• To develop a few-shot transfer learning strategy for adapting MLIPs to different computational setups and design spaces.

Main Methods:

• An ensemble of catalysis-focused MLIPs was bias-corrected using a small number of density functional theory (DFT) labels from the target setup.
• A few-shot transfer learning approach was employed for bias correction and uncertainty quantification.

Main Results:

• The approach reduced root mean squared errors (RMSEs) by 60% for OH adsorption on bimetallic alloys with one additional DFT calculation.
• For H adsorption on single-atom alloys, the zero-shot ensemble outperformed individual MLIPs, with further improvements using 3-5 DFT calculations.
• Well-calibrated uncertainty estimates were achieved, with low miscalibration areas (0.039 and 0.088).
• A promising trimetallic catalyst for oxygen reduction was identified using only four DFT calculations.

Conclusions:

• Few-shot bias correction enables reliable transfer of MLIP predictions across different alloy search spaces and DFT methodologies.
• This provides an efficient and practical route to accurate, uncertainty-aware catalyst screening.