Employing autoencoders for configuration space sampling: Application to small molecules.

The behavior of molecular systems in different equilibrium physical processes or chemical reactions is governed by the free energy (FE). Calculations of FE require a thorough sampling of configuration space for given external conditions. State-of-the-art sampling techniques, based on molecular dynamics, are formally applicable to systems of arbitrary size. However, in practice, they suffer from the curse of dimensionality and the limitation of the time step by the fastest process present in the system. These make FE calculations for complex molecules, where the entropy effects are of utmost importance, extremely challenging and computationally demanding. Here we propose to use a machine-learning approach based on autoencoders to generate new sampling configurations. By transition to the feature space, we effectively decrease the dimensionality of the problem and resolve the time step limitation. Training autoencoders cost only a small fraction of statistically converged molecular dynamics simulations, paving the way to efficient calculations of thermodynamic properties for complex molecular systems.