Potential energy surfaces: Δ-machine learning from analytical functional forms.

Area of Science:

• Computational Chemistry
• Chemical Physics
• Machine Learning in Chemistry

Background:

• Developing accurate potential energy surfaces (PESs) is crucial for understanding chemical reactions.
• High-level electronic structure calculations are accurate but computationally expensive.
• Machine learning (ML) offers a promising avenue for developing cost-effective PESs.

Purpose of the Study:

• To introduce and validate a Delta-machine learning (Δ-ML) approach for constructing accurate PESs.
• To assess the efficiency of Δ-ML for polyatomic systems using the H + CH4 reaction as a benchmark.
• To compare Δ-ML PES with high-level theoretical methods for kinetics and dynamics.

Main Methods:

• Utilized a flexible analytical potential energy surface to efficiently sample low-level data.
• Integrated information from a high-accuracy permutation invariant polynomial neural network (PIP-NN) surface.
• Performed kinetic studies using variational transition state theory with multidimensional tunneling corrections.
• Conducted dynamic studies using quasiclassical trajectory calculations on the H + CD4 reaction.

Main Results:

• The Δ-ML approach successfully reproduced the kinetics and dynamics of the H + CH4 reaction.
• The constructed Δ-ML PES demonstrated high accuracy comparable to high-level surfaces.
• The method proved efficient in describing the multidimensional polyatomic system.

Conclusions:

• Delta-machine learning (Δ-ML) provides a highly cost-effective strategy for generating accurate potential energy surfaces.
• The developed Δ-ML method is effective for modeling the complex kinetics and dynamics of polyatomic chemical reactions.
• This approach shows significant promise for computational chemistry applications requiring accurate PESs.