High-Accuracy Molecular Simulations with Machine-Learning Potentials and Semiclassical Approximations to Quantum...

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning

Background:

• Accurate molecular simulations necessitate high-level electronic-structure theory and rigorous quantum dynamics approximations.
• Current computational methods are often prohibitively expensive for studying complex chemical reactions.

Purpose of the Study:

• To explore machine-learning approaches for reducing the computational cost of molecular simulations.
• To investigate methods for constructing accurate potential energy surfaces using minimal data.
• To enable advanced semiclassical approximations for quantum dynamics.

Main Methods:

• Utilizing machine learning to construct potential energy surfaces (PES).
• Employing transfer learning for efficient PES construction with minimal training data.
• Applying smooth and differentiable PES to advanced semiclassical quantum dynamics methods.

Main Results:

• Machine-learning significantly reduces computational expense without compromising simulation accuracy.
• Transfer learning allows for PES construction using a minimal number of expensive training points.
• Smooth, differentiable PES facilitate the use of advanced semiclassical methods like perturbatively corrected instanton theory.

Conclusions:

• Machine learning offers a cost-effective solution for high-level molecular simulations.
• The developed methods enable the study of chemical reactions and quantum dynamics with improved accuracy and efficiency.
• Advanced semiclassical approximations can effectively capture quantum mechanical effects such as tunneling and anharmonicity.