Knowing when to trust machine-learned interatomic potentials

TL;DR

This paper introduces PROBE, a post-hoc method for reliable uncertainty quantification in machine-learned interatomic potentials, outperforming ensemble-based methods and providing interpretable diagnostics.

Contribution

PROBE recasts uncertainty quantification as selective classification using frozen embeddings, offering a scalable, architecture-agnostic, and interpretable alternative to ensemble methods.

Findings

PROBE's reliability probability correlates strongly with actual prediction error.

PROBE outperforms ensemble disagreement as a binary reliability signal.

Multi-head self-attention provides chemically interpretable importance maps.

Abstract

Prevailing machine-learned interatomic potential (MLIP) uncertainty-quantification methods rely on ensembles of independently trained backbones. These methods scale unfavorably with foundation-scale MLIPs, and their member-disagreement signals correlate weakly with per-molecule prediction error. Here we probe the frozen per-atom representations of a pretrained MLIP with a compact discriminative classifier, recasting MLIP uncertainty quantification as selective classification rather than error regression. The resulting method, PROBE (Post-hoc Reliability frOm Backbone Embeddings), produces a per-prediction reliability probability that monotonically tracks actual error without modification to the underlying model. Across large held-out evaluation sets and two structurally distinct MLIP architectures, PROBE outperforms ensemble disagreement as a binary reliability signal, which strengthens with the expressiveness of the backbone representation, implying a favorable scaling trajectory toward foundation-scale MLIPs. Multi-head self-attention additionally yields per-atom importance maps, providing chemically interpretable diagnostics at no additional computational cost. PROBE is post-hoc and architecture-agnostic, and is directly deployable on any MLIP that exposes per-atom representations.