Creating Gaussian Process Regression Models for Molecular Simulations Using Adaptive Sampling

FFLUX is a new force field that combines the accuracy of quantum mechanics with the speed of force fields, without any link to the architecture of classical force fields. This force field is atom-focused and adopts the parameter-free topological atom from Quantum Chemical Topology (QCT). FFLUX uses Gaussian Process Regression (GPR) (aka kriging) models to make predictions of atomic properties, which in this work are atomic energies according to QCT’s Interacting Quantum Atom (IQA) approach. Here we report the adaptive sampling technique Maximum Expected Prediction Error (MEPE) to create data-compact, efficient and accurate kriging models (sub kJ mol-1 for water, ammonia, methane and methanol, and sub kcal mol-1 for N-methylacetamide (NMA)). The models cope with large molecular distortions and are ready for use in molecular simulation. A brand new press-one-button Python pipeline, called ICHOR, carries out the training.