AoLNS Doctoral Talk – Molecular Innovation and Advanced Materials Theme Week1

Title 1

ReactBench: Evaluating Foundational Machine-Learning Interatomic Potentials for Chemical Reactions

Abstract

The growing use of machine learning interatomic potentials (MLIPs) as general-purpose engines for atomistic simulations calls for benchmarks that test not only predictive accuracy, but also reliability in the applications for which these models are ultimately intended. Here, we introduce ReactBench, a benchmark suite designed to evaluate foundational MLIPs for chemical reactions across tasks spanning static reaction-path energetics, reaction-path optimisation, and reactive dynamics. ReactBench contains 5,315 homogeneous and heterogeneous reactions and evaluates 12 foundational MLIPs. ReactBench therefore goes beyond static energy benchmarks to evaluate MLIPs in realistic reactive applications, identify their limitations and guide the development of more reliable models.

Title 2

Machine Learning Based Coarse-Grained Force Fields for Liquids

Abstract

Molecular dynamics of liquids demands forces that are both accurate and affordable. Machine-learned coarse-grained force fields promise both, by learning the potential of mean force directly from simulation data. This talk introduces the physical foundations of machine-learned force fields — symmetry, locality, and equivariance — and explains why equivariance becomes essential at the coarse-grained level, where many-body effects are the potential. We compare two models — So3krate and NEP — asking how many-body interactions can be represented, whether phase separation can be captured, and how overfitting to intrinsic noise can be avoided.