Toward optimal descriptors for accurate machine learning of flexible molecules

Robust machine learning (ML) models should be able to reliably predict global molecular potential-energy surfaces (PES) including equilibrium and “far-from-equilibrium” geometries. However, existing molecular ML models are substantially biased towards “close-to-equilibrium” geometries. Indeed, the difficulty of the learning task increases with increasing flexibility of a molecule, due to a vast manifold of configurations with a complex interplay of covalent and non-covalent interactions to be learned.

Our objective is to test how the ability to accurately reproduce PES depends upon the choice of a molecular descriptor. We use azobenzene, aspirin and salicylic acid molecules as our test systems, and the sGDML code¹ for building ML force-fields. We found that the descriptors which demonstrate excellent performance for “close-to-equilibrium” parts of PES are inefficient for building global PES models. To resolve this challenge, we propose new descriptors that allow building accurate and data-efficient ML models for flexible molecules.

1. Chmiela, S. et al., Sci. Adv. 3, e1603015 (2017) ; Chmiela, S. et al., Nat. Commun. 9, 3887 (2018).