Physically and chemically inspired kernel-based neural network for constructing accurate and efficient machine learning force fields for hundreds of atoms.

Accurate and efficient machine learning (ML) models have become a game-changing element in modern (bio)chemistry and materials science during the last decade. Among other applications, the ability to run ab initio level atomistic molecular dynamics for nanoseconds enabled an understanding of multiple challenging processes observed by experiments. The kernel-based methods and artificial neural networks are the two primary techniques for constructing ML models, each having its strengths and weaknesses. In our work, we combined the advantages of both approaches in a single ML architecture capable of accurately reconstructing force fields (FF) for systems of varying complexity from small (MD17 dataset [1]) and intermediate-sized (MD22 dataset [2]) molecules and molecular complexes to large-scale periodic systems with hundreds of atoms per unit cell (MAPbI3). The proposed MLFFs efficiently capture the multiscale nature of interatomic interactions and the indistinguishability of atoms with identical chemical environments. Moreover, the architecture can be equally used as a global ML model providing the ultimate accuracy and in a localized version enabling transferability.

1) S. Chmiela et al., Sci. Adv. 3, e1603015 (2017).

2) S. Chmiela et al., arXiv.2209.14865 (2022).