Accurate and Efficient ML Force Fields for Hundreds of Atoms

In order to faithfully represent non-local interatomic interactions, a molecular force field has to allow interactions between all degrees of freedom, without resorting to localization or other approximations. This has been a challenge for machine learning (ML) based approaches, because even a simple pairwise correlation implies a poor quadratic scaling behavior in the number of atoms, which quickly becomes computationally prohibitive for training. To date, no fully-correlated global ML models exist that are applicable to systems with more than a few dozen atoms.

To overcome this limitation, we develop an efficient iterative, parameter-free solver to train symmetric gradient domain machine learning (sGDML) [Chmiela et al., 2018] potentials for systems with several hundred atoms. Our approach keeps all correlations of this global model intact, allowing the accurate description of complex molecules, materials and molecular assemblies.