A Neural Network-Assisted Euler Integrator for Stiff Kinetics in Atmospheric Chemistry
Yuanlong Huang, John H. Seinfeld
Abstract
Atmospheric chemistry, characterized by highly coupled sets of ordinary differential equations (ODEs), is dynamically stiff owing to the fact that both fast and slow processes exist simultaneously. We develop here a neural network-assisted Euler integrator for the kinetics of atmospheric chemical reactions. We show that the integral kernel of the chemical reaction system can be represented by a neural network. The stiff kinetics of the atmospheric H2O2/OH/HO2 system, involving 3 species and 4 reactions, and a simplified air pollution mechanism, involving 20 species and 25 reactions, are developed here in detail as illustrations of the neural network Euler integrator. The algorithm developed accelerates the numerical integration of large sets of coupled stiff ODEs by at least one order of magnitude by avoiding the intensive linear algebra that is required in traditional stiff ODE solvers; moreover, the mechanism-specific neural network-assisted algorithm can be readily coupled to other modules in a three-dimensional atmospheric chemical transport model.