How accurate are foundational machine learning interatomic potentials for heterogeneous catalysis?

TL;DR

This paper systematically evaluates the zero-shot performance of 80 foundational MLIPs in heterogeneous catalysis, revealing their strengths, limitations, and the need for application-specific assessment.

Contribution

It provides the first comprehensive benchmark of MLIPs across diverse catalytic tasks and highlights their variable performance and current limitations.

Findings

MLIPs perform well for vacancy and zero-point energy predictions.

Many MLIPs fail catastrophically on magnetic materials.

No single MLIP universally outperforms others.

Abstract

Foundational machine learning interatomic potentials (MLIPs) are being developed at a rapid pace, promising closer and closer approximation to ab initio accuracy. This unlocks the possibility to simulate much larger length and time scales. However, benchmarks for these MLIPs are usually limited to ordered, crystalline and bulk materials. Hence, reported performance does not necessarily accurately reflect MLIP performance in real applications such as heterogeneous catalysis. Here, we systematically analyze zero-shot performance of 80 different MLIPs, evaluating tasks typical for heterogeneous catalysis across a range of different data sets, including adsorption and reaction on surfaces of alloyed metals, oxides, and metal-oxide interfacial systems. We demonstrate that current-generation foundational MLIPs can already perform at high accuracy for applications such as predicting vacancy formation energies of perovskite oxides or zero-point energies of supported nanoclusters. However, limitations also exist. We find that many MLIPs catastrophically fail when applied to magnetic materials, and structure relaxation in the MLIP generally increases the energy prediction error compared to single-point evaluation of a previously optimized structure. Comparing low-cost task-specific models to foundational MLIPs, we highlight some core differences between these model approaches and show that -- if considering only accuracy -- these models can compete with the current generation of best-performing MLIPs. Furthermore, we show that no single MLIP universally performs best, requiring users to investigate MLIP suitability for their desired application.