Emulating tropospheric chemistry mechanisms with deep neural networks

Chemical transport models (CTMs) are essential tools for investigating the chemical processes at stake in the atmosphere and supporting needs on air quality assessment and planning. Yet, they typically require massive computational resources to solve the system of stiff ordinary differential equations governing atmospheric chemical kinetics (many reactions and species with highly variable abundances and kinetic time scales), which limits the resolution of simulations and/or the level of complexity of the chemistry representation. In the frame of the EACH (Emulating Atmospheric Chemistry) project involving the Barcelona Supercomputing Center, the Max Planck Institute for Chemistry,and Freie Universität Berlin, we are investigating the potential of deep learning to emulate the chemistry, focusing first on gas phase chemistry, with the ultimate goal of being able to accelerate CTM simulations. More specifically, we assess the performance and generalization capability of dense deep feedforward neural networks based on the multilayer perceptron (MLP) architecture using two test mechanisms: POLLU[1], a simplified tropospheric ozone formation mechanism (20 species, 25 reactions), and CB05[2], a condensed mechanism of atmospheric oxidant chemistry (59 species, 156 reactions) used in many CTMs. Training datasets were generated from millions of 0D chemical box-model simulations, with initial conditions sampled from uniform multidimensional distributions. For each experimental setup, a systematic hyper-parameter search was conducted to identify the optimal configuration. We trained several MLP variants incorporating physical consistency through both hard (architectural) and soft (loss-function-based) physical constraints designed to preserve stoichiometric relationships and enforce non-negativity of concentrations, and we assessed mass conservation using tailored evaluation metrics. The sensitivity of the MLPs performances to the number of training time series and their length was explored to examine the impact of data design on model performance.[1] Verwer, J. G. (1994). Gauss–Seidel iteration for stiff ODEs from chemical kinetics. SIAM Journal on Scientific Computing, 15(5), 1243-1250.[2] Yarwood, Greg, et al. (2005). Final report to the US EPA, RT-0400675 8: 13