Revisiting the Gaussian process regression for fitting high-dimensional potential energy surface and its application to the oh + ho2 → o2 + h2o reaction

Abstract

In this work, Gaussian process regression (GPR) for fitting a high-dimensional potential energy surface (PES) is revisited and implemented to construct the PES of OH + HO2 → O2 + H2O. Using mixed kernel function and optimized distribution of the training database, only ∼3 × 103 energy points are needed to approach convergence, which implies the power of GPR in saving lots of computational cost. Moreover, the convergence of the GPR PES is inspected, leading to discussions on the advantages of the GPR fitting approach. By the segmented strategy [Meng et al., J. Chem. Phys. 144, 154312 (2016)], a GPR PES with a fitting error of ∼21 meV is constructed using ∼4600 energy points at the CCSD(T)-F12a/aug-cc-pVTZ level. The rate coefficients are then computed through the ring-polymer molecular dynamics (RPMD) method. An agreement between the present RPMD calculations and the previous observations is found, implying the accuracy of the present calculations. Moreover, the unusual feature of the Arrhenius curve is interpreted by a coupled harmonic oscillator model [Q. Meng, J. Phys. Chem. A 122, 8320 (2018)] together with a simple kinetics model.