A Gaussian Process Based Δ-Machine Learning Approach to Reactive Potential Energy Surfaces

Area of Science:

• Computational chemistry
• Machine learning
• Physical chemistry

Background:

• Gaussian processes (GPs) are powerful nonparametric machine learning tools.
• GPs offer uncertainty quantification, crucial for error reduction in complex models.
• Potential energy surfaces (PESs) are fundamental for understanding chemical reactions.

Purpose of the Study:

• To apply Gaussian processes within the Δ-machine learning (Δ-ML) framework.
• To construct accurate potential energy surfaces (PESs) using a hybrid approach.
• To evaluate the efficiency and accuracy of GP-based Δ-ML for reactive systems.

Main Methods:

• Utilizing Gaussian processes (GPs) for Δ-machine learning (Δ-ML).
• Constructing a correction PES using high-level ab initio calculations on a low-level PES.
• Testing the method on three benchmark reactive systems: H + H2, OH + H2, and H + CH4.

Main Results:

• The GP-based Δ-ML approach demonstrates higher efficiency than direct GP application for high-level PES construction.
• The method effectively reduces errors in PES calculations.
• GP-based Δ-ML shows comparable performance to neural network-based Δ-ML methods.

Conclusions:

• Gaussian processes provide an efficient and accurate method for constructing potential energy surfaces via Δ-machine learning.
• The GP-based Δ-ML approach is a viable alternative to other machine learning techniques for chemical reaction studies.
• This work highlights the utility of GPs in computational chemistry for improving reaction dynamics calculations.