Learning Free Energies from Molecular Simulation using Artificial Neural Networks

The use of adaptive biasing potential methods to estimate free energies controlling phase changes and chemical processes from molecular simulation have become increasingly popular in recent years. The most widely applied method, metadynamics, has gained popularity due to its algorithmic simplicity and the ease with which it is incorporated into molecular dynamics simulation, but it is not without limitations due to its use of Gaussian kernels in the biasing process. Recent improvements have utilized basis function expansions to better match highly nuanced and nonlinear free energy landscapes, but present complications of their own in arbitrary bounded domains or on surfaces with sharp curvature. Here, we report on a recently-developed method which uses artificial neural networks to learn the free energy surface from incomplete data. The method is robust to user-defined parameters and shows dramatic improvement over currently available methods in reproducing the free energy of poorly sampled states.