Molecular Force Fields with Gradient-Domain Machine Learning: Dynamics of Small Molecules with Coupled Cluster Forces

Molecular dynamics (MD) simulations using conventional force fields constitute the cornerstone of contemporary atomistic modeling in biology, chemistry, and materials science. However, the predictive power of these simulations is only as good as the underlying interatomic potential. Here we present the reconstruction of molecular force fields for small molecules using the recently developed symmetric gradient-domain machine learning (sGDML) approach. The sGDML approach faithfully reproduces complex high-dimensional potential-energy surfaces from just a few 100s of molecular conformations generated by ab-initio MD simulations. The data efficiency of the model allows employing high-level wavefunction-based atomic forces and energies for training, such as the “gold standard” CCSD(T) method. We demonstrate that the flexible nature of this fully data-driven model recovers any local and non-local quantum interaction coming from -F=〈φ*│∂H/∂x│φ〉 (e.g. H-bonding, proton transfer, lone pairs, changes in hybridization states, steric repulsion and n->pi* transitions) without relying on prior knowledge of the phenomena. The analysis of MD@sGDML trajectories yields new qualitative insights into dynamics, chemistry, and spectroscopy of small molecules close to spectroscopic accuracy.