Chemical reaction rate coefficient optimization using Fourier neural operator

Abstract

Modeling reaction rates for atmospheric chemical reactions is essential for predicting chemical concentrations and accelerating the simulation of atmospheric chemical processes. Traditional approaches rely on numerical integration of Ordinary Differential Equations (ODEs), which becomes computationally expensive for stiff systems. This study aims to explore the use of the Fourier Neural Operator (FNO) to efficiently learn and predict chemical concentrations. A two-step approach is used: first, the model learns the relationship between initial chemical concentrations, reaction coefficients, and actual concentrations at different time points in the frequency domain, enabling faster convergence and better generalization compared to conventional neural networks. The process is followed by inverse optimization to estimate reaction coefficients. The model was trained on synthetic reaction datasets, namely Robertson and POLLU, with a combination of different parameters to approximate reaction rates under varying conditions. Extensive experiments were performed, and the findings were promising for both datasets, with the model accurately estimating all three reaction coefficients in Robertson and 12 of 25 in POLLU. Although the model performed poorly in the latter due to increased variability and interaction complexity, it still demonstrated the ability to identify dominant kinetic parameters and approximate the overall system dynamics with reasonable accuracy.