Machine Learning for Observables: Reactant to Product State Distributions for Atom-Diatom Collisions.

Area of Science:

• Chemical Physics
• Computational Chemistry
• Machine Learning

Background:

• Atom-diatom collisions are fundamental to chemical reactions and energy transfer.
• Predicting product state distributions from reactant conditions is computationally challenging.
• Existing methods often lack accuracy or generality for complex systems.

Purpose of the Study:

• To develop and test machine learning models for predicting product state distributions in atom-diatom collisions.
• To compare the performance of function-, kernel-, and grid-based machine learning approaches.
• To assess the applicability of these models to nonequilibrium conditions, such as in hypersonic flows.

Main Methods:

• Developed machine learning models using function-, kernel-, and grid-based representations.
• Trained models on data from quasi-classical trajectory simulations.
• Quantitatively tested prediction accuracy using R-squared values.
• Investigated the impact of function choice in the function-based approach.
• Applied the grid-based approach to multitemperature initial distributions.

Main Results:

• All three machine learning approaches achieved high prediction accuracy (R² > 0.998).
• The grid-based approach demonstrated superior accuracy, practicality, and generality compared to function- and kernel-based methods.
• The function-based approach showed better computational performance but was sensitive to function choice.
• The grid-based model successfully predicted distributions for nonequilibrium, multitemperature initial states.

Conclusions:

• Machine learning models, particularly the grid-based approach, are highly effective for predicting product state distributions in atom-diatom collisions.
• The grid-based method is recommended for its accuracy, practicality, and generality.
• These models have significant potential for integration into computational fluid dynamics and direct simulation Monte Carlo methods for simulating complex phenomena like hypersonic flows.