Atmospheric Chemistry Surrogate Modeling With Sparse Identification of Mass Action Dynamics

Abstract Modeling gas‐phase atmospheric chemistry is computationally challenging because the underlying dynamics are high‐dimensional (requiring many state variables) and stiff (characterized by Jacobian eigenvalues that span many orders of magnitude). Previous work has demonstrated the promise of machine learning (ML) to accelerate air quality model simulations, but such models commonly suffer from dynamical instability or runaway error propagation. Here, to address this limitation, we introduce a physics‐informed parsimonious machine‐learning framework for atmospheric chemistry. This Sparse Identification of Mass Action Dynamics (SIMADy) framework operates by (a) using a linear autoencoder to create an interpretably‐compressed low‐dimensional state representation, and (b) representing the underlying dynamics of that state constrained by the law of mass action and therefore ensuring boundedness. Using the full Master Chemical Mechanism—one of the most detailed atmospheric chemical mechanisms—as the reference model, the root mean square (RMS) error of the ML model prediction for ozone () concentration over 10 days is 25% of the RMS reference model concentration across all simulations in testing data set when using six state variables. As expected, the ML model predictions remain bounded and exhibit no numerical divergence in all tested simulations. The surrogate model is 5,300 times faster on a CPU (for one simulation) and approximately 1.1 million times faster on a GPU (for many simulations in parallel) compared to the reference model. In future work, the fast, stable, and relatively accurate surrogate models learned by SIMADy can be used to accelerate 3D models of air quality and climate.