Generating Converged Quantum-Accurate Free Energy Surfaces for Chemical Reactions with a Force-Matched Semi-Empirical Model

Molecular dynamics using density functional theory (DFT) is a highly accurate approach to predict chemistry, but the extreme computational cost often harshly limits the exploration of long time scale phenomena and many thermodynamic states. We present a general force-matching approach to parameterize quantum-based semi-empirical models that can retain DFT-level accuracy while affording up to a thousandfold reduction in cost. Accelerated sampling is used to simultaneously generate DFT training data and validate the force-matched model for particular reaction paths. We show that a force-matched semi-empirical model for aqueous glycine condensation reactions yields free energy surfaces that are consistent with experiments. Convergence analysis reveals that multiple nanoseconds of combined trajectory are needed to reach a steady-fluctuating free energy estimate, which is accessible with high-throughput models such as those presented here. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.