Accelerating Ab Intio-Based Free Energy Sampling with Machine Learning

Ab initio Molecular Dynamics (AIMD) simulations are powerful for studying quantum mechanical phenomena, including chemical reactions, diffusion, and defect generations in materials, which require breaking and formation of bonds. Understanding not only the kinetics but also the thermodynamic properties of materials often require combining molecular dynamics with enhanced sampling techniques to capture the free energy landscape of the underlying phenomena. Unfortunately, free energy sampling with AIMD has been infeasible for large complex systems due to the computational cost of electronic structure calculations that are needed for each point in the configurational space. In this talk, we propose two approaches to overcome such bottleneck by using machine learning, namely the artificial neural network: (1) active learning of free energy surface during the adaptive biasing simulations and (2) on-the-fly learning of both ab initio force field and free energy surfaces during biased AIMD simulations. We demonstrate these two methods by applying them to catalytic reactions and defect migrations in covalently bonded solids.