Multi-fidelity Gaussian process modeling for chemical energy surfaces

Abstract

There has been a recent surge of interest in using Gaussian process (GP) regression to model chemical energy surfaces. Herein, we discuss an extension of GP modeling called autoregressive Gaussian process (ARGP) modeling, which uses an approximation to the target function to improve learning efficiency and has never been applied to high-accuracy chemical energy surfaces. Our calculations demonstrate that ARGP regression improves the prediction error of a five-point GP regression by two orders of magnitude for an N2 dissociation curve and reproduces the energy surface of H2O with sub-chemical accuracy, using only 25 training points.