Constructing Accurate Machine Learning Force Fields for Flexible Molecules

Machine learning (ML) models can reproduce potential energy surfaces (PES) for molecules containing up to a few tens of atoms with an accuracy comparable to the most exact ab initio methods. This provides a tool for computing thermodynamic properties that would require millions of CPU years otherwise. For instance, a recently developed sGDML¹ model predicts forces and energy with CCSD(T) accuracy using just a few hundreds of configurations for training. However, up to now ML has been mainly applied to rather rigid molecules. In this regard, our objective is to test ML models for flexible molecules and out-of-equilibrium configurations along transition paths. For this, we select molecules (e.g. azobenzene, stilbene) with relatively complex transition paths, which result from an interplay between long- and short-range interactions. Then, different paths connecting PES minima are tested using sGDML. This allows us to define optimal descriptors and the appropriate strategies for choosing the training sets, which is crucial for ML models relying on a limited number of training points. Our results open an avenue for calculating transport paths, transition rates and other "out-of-equilibrium” properties with previously unattained accuracy.

1. Chmiela et al., Nat. Commun. 9, 3887 (2018)