Challenges and Opportunities of Pretrained Machine Learning Interatomic Potentials in Heterogeneous Catalysis

TL;DR

This paper explores how machine learning can improve catalyst design by offering faster and accurate atomistic modeling compared to traditional methods.

Contribution

The paper introduces and evaluates pretrained machine learning interatomic potentials for heterogeneous catalysis, emphasizing their potential and challenges.

Findings

Pretrained MLIPs can match DFT accuracy at lower computational cost.

Standardized protocols are needed to benchmark MLIP performance across architectures.

Transferability and integration challenges limit the widespread use of MLIPs in catalysis.

Abstract

The design of catalysts gets its fundamental rationale
from accurate
and efficient modeling of reactivity on surfaces and materials. To
reach this detailed atomistic understanding, density functional theory
(DFT) has been the key computational technique. However, the emergence
of machine learning interatomic potentials (MLIPs) marks a significant
paradigm shift, offering the potential to match DFT accuracy at a
drastically reduced computational cost. This perspective provides
an overview of state-of-the-art MLIPs for heterogeneous catalysis
as “out-of-the-box” tools. We summarize the different
families of MLIPs and their training processes and then apply these
pretrained models to heterogeneous catalysis problems. Furthermore,
we critically address the challenges of model transferability and
integration in unified frameworks, underscoring the necessity for
standardized protocols to benchmark performance across different architectures.
Finally, we assess the capacity of pretrained models to democratize
computational catalysis, highlighting the specific hurdles that remain
in achieving reliable, predictive power for widespread use.

Figures