Benchmarking machine-learned interatomic potential methods for reactive molecular dynamics at metal surfaces

Machine-learned interatomic potentials (MLIP) have become widely used tools to accelerate ab initio molecular dynamics simulations in materials science. Many promising MLIPs emerged recently, from simple linear models to deep neural networks (DNN), differing in stability, accuracy, and inference time. The field of reactive dynamics at surfaces has specific requirements on potentials, which make it an interesting area to benchmark different MLIP approaches. Reactive scattering dynamics are highly sensitive to potential corrugation and low reaction probabilities require extensive ensemble averaging. Therefore, MLIPs need to combine smooth and accurate landscapes with extremely efficient inference. In this study, we compare different families of MLIPs, from atomic cluster expansion (ACE), invariant DNN-based SchNet to novel equivariant neural networks such as PaiNN and MACE on the example of reactive molecular hydrogen scattering on copper. We compare these diverse methods by measuring accuracy and inference performance directly on dynamical observables. This provides a detailed picture of MLIP smoothness and corrugation accuracy that goes beyond basic train/test error analysis.