Machine learning dynamic correlation in chemical kinetics

Area of Science:

• Computational chemistry
• Surface science
• Chemical kinetics

Background:

• Lattice models are crucial for simulating surface reaction kinetics.
• Propagating entire lattice configurations is computationally expensive.
• Moment closure approximations simplify these simulations but struggle with long-range correlations.

Purpose of the Study:

• To investigate the use of machine learning (ML) for developing accurate moment closures in chemical kinetics.
• To address the limitations of traditional closures like mean-field and pair approximation (PA).
• To utilize the lattice Lotka-Volterra model as a test system.

Main Methods:

• Training feedforward neural networks on kinetic Monte Carlo (KMC) simulation data.
• Using KMC results for specific rate constants and initial conditions.
• Developing a machine learning moment closure (MLMC) approach.

Main Results:

• MLMC accurately predicts three-site occupation probabilities with the same input as PA.
• MLMC significantly improves the simulation of dynamics and dynamical regimes.
• MLMC demonstrates superior performance compared to traditional closure methods.

Conclusions:

• MLMC is a powerful tool for enhancing the accuracy of lattice-based kinetic simulations.
• MLMC offers a computationally efficient method for interpolating KMC data.
• MLMC enables researchers to gain insights at a reduced computational cost.