Representing Global Reactive Potential Energy Surfaces Using Gaussian Processes.

Area of Science:

• Computational chemistry
• Theoretical chemistry
• Chemical physics

Background:

• Accurate potential energy surfaces (PES) are crucial for understanding chemical reactions.
• Generating high-dimensional PES from ab initio calculations is computationally expensive and challenging.

Purpose of the Study:

• To investigate the efficacy of Gaussian process regression (GPR) for constructing global potential energy surfaces.
• To evaluate the number of ab initio data points required for accurate spectral and dynamical calculations.

Main Methods:

• Employed Gaussian process regression, a machine learning technique, to build potential energy surfaces.
• Utilized high-level ab initio calculations to generate data points for training the GPR model.
• Tested the GPR-based PES for the 3A″ state of SH2, involving abstraction and exchange reactions.

Main Results:

• GPR can generate a reasonable PES with approximately 100 ab initio points.
• Accurate convergence of reaction probabilities requires substantially more data points (around 1000).
• The study highlights the trade-off between data quantity and accuracy in PES construction.

Conclusions:

• Gaussian process regression is a promising method for developing potential energy surfaces.
• The number of required ab initio data points depends on the desired accuracy for specific chemical applications.
• Further research is needed to optimize GPR methods for complex chemical systems.