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

TL;DR

This paper introduces a new machine learning method that improves the accuracy of catalyst predictions using only a few additional data points.

Contribution

A novel few-shot ensemble learning strategy for MLIPs that enables reliable catalyst screening with uncertainty quantification.

Findings

The approach reduces RMSEs by 60% on average with just one DFT-calculated adsorption energy.

The corrected ensemble provides well-calibrated uncertainty estimates with low miscalibration areas.

A promising trimetallic catalyst was identified using only four DFT-calculated adsorption energies.

Abstract

Machine-learned interatomic
potentials (MLIPs) are increasingly
used to accelerate catalyst discovery, but their accuracy and utility
are often unclear, particularly when applying them to different computational
setups or design spaces than those of the training data. This hinders
their effective use in catalyst screening. Here, we improve accuracy
and provide reliable uncertainty quantification through an ensemble-based,
few-shot transfer learning strategy. The framework applies a bias-correction
procedure to an ensemble of catalysis-focused MLIPs using a small
number of density functional theory (DFT) labels from the target setup
and design space. Applied to OH adsorption on bimetallic alloys, the
approach reduces root mean squared errors (RMSEs) by 60% on average
after incorporating just one additional DFT-calculated adsorption
energy. For H adsorption on single-atom alloys, even the zero-shot
ensemble is more accurate than any of the MLIPs, with further improvements
when three to five DFT calculations are used for bias correction.
The corrected ensemble yields well-calibrated uncertainty estimates,
with miscalibration areas of just 0.039 and 0.088 for the two data
sets. In a proof-of-concept screening campaign, the method identified
a promising trimetallic candidate catalyst for the oxygen reduction
reaction using only four total DFT-calculated adsorption energies.
Taken together, these results demonstrate that few-shot bias correction
enables reliable transfer of MLIP predictions across mismatched alloy
search spaces and DFT methodologies, providing a practical route to
accurate, uncertainty-aware catalyst screening with high efficiency.

Figures