Accelerating Variational Transition State Theory via Artificial Neural Networks

Area of Science:

• Computational Chemistry
• Chemical Physics
• Materials Science

Background:

• Variational transition state theory (VTST) is a powerful method for calculating reaction rate constants.
• Accurate potential energy surfaces are crucial for VTST, but their computation can be resource-intensive.
• Radical-radical reactions are fundamental in many chemical processes, including combustion and atmospheric chemistry.

Purpose of the Study:

• To apply atomistic machine learning to accelerate VTST calculations for radical-radical reactions.
• To develop surrogate potential energy surfaces using artificial neural networks (ANNs).
• To reduce the computational burden of electronic structure calculations in rate constant determination.

Main Methods:

• Artificial neural networks were trained on potential energy data from electronic structure calculations.
• A surrogate potential energy surface was generated using the trained ANNs.
• Classical phase space representations were employed to model radical-radical interactions.
• Variable reaction coordinate transition state theory was used to compute rate constants.

Main Results:

• Training ANNs on potential energy data alone reduced required electronic structure calculations by at least a factor of 4.
• Including forces in the ANN training data led to a more dramatic reduction, at least an order of magnitude.
• The surrogate potential energy surface accurately represented the interactions between radical species.
• Computed rate constants for radical-radical reactions showed good agreement with theoretical predictions.

Conclusions:

• Atomistic machine learning offers a significant acceleration for VTST calculations.
• ANN-based surrogate potential energy surfaces drastically reduce computational costs.
• This methodology enables more efficient and extensive studies of radical-radical reaction dynamics.