(431d) Progress Towards Alchemical Simulations Driven By Many-Body Gradient-Domain Machine Learning

Alchemical predictions of molecular properties, such as solvation free energies and pK_as, provide an immense value for understanding chemical and physical processes. However, simulation convergence requires substantial sampling along the alchemical pathway, which is typically only possible with classical potentials. Nowadays, machine learning (ML) potentials can drive molecular simulations with quantum chemical accuracy at a significantly reduce cost—albeit a couple times slower than classical potentials. Training ML potentials often requires a sizable amount (e.g., tens of thousands) of quantum chemistry calculations that diminishes the overall computational cost savings. Many-body gradient-domain machine learning (mbGDML) is a recently developed framework that incorporates the size transferability of many-body expansions with the accuracy and efficiency of GDML potentials. After training, we have n-body potentials that accurately reproduce quantum chemical forces and energies with only 1000 training data points. Here, we describe the mbGDML approach and its use in predicting solvation free energies with alchemical simulations.