The neural network based Δ-machine learning approach efficiently brings the DFT potential energy surface to the CCSD(T) quality: a case for the OH + CH3OH reaction

Abstract

The recently proposed permutationally invariant polynomial-neural network (PIP-NN) based Δ-machine learning (Δ-ML) approach (PIP-NN Δ-ML) is a flexible, general, and highly cost-efficient method to develop a full dimensional accurate potential energy surface (PES). Only a small portion of points, which can be actively selected from the low-level (often DFT) dataset, with high-level energies are needed to bring a low-level PES to a high-level of quality. The hydrogen abstraction reaction between the methanol and hydroxyl radical, OH + CH₃OH, has been studied using theories and experiments for a long time due to its great importance in combustion, atmospheric and interstellar chemistry. However, it is not trivial to develop the full dimensional accurate PES for it. In this work, the PIP-NN Δ-ML method is successfully applied to the title reaction. The DFT PES was fitted by using 140 192 points. Only 5% of the DFT dataset was needed to be calculated at the level of UCCSD(T)-F12a/AVTZ, aiming to improve the DFT PES to the target high-level, UCCSD(T)-F12a/AVTZ. More than 92% of the original unaffordable calculation costs were saved. The kinetics, including rate coefficients and branching ratios, were then studied by performing quasi-classical trajectory calculations on this newly fitted PES for the title reaction.