From Stiff Equations to Deep Learning: Overcoming Challenges in Simulating Complex Atmospheric Aerosol Chemistry

Fine particles in the atmosphere known as Secondary Organic Aerosols (SOA) have a considerable impact on the Earth's energy budget, as they interact with clouds and radiation. The formation of SOA is a complex process that involves various chemical reactions in the gas phase, aqueous aerosols, and clouds. This process is computationally expensive for three-dimensional chemical transport models, as it requires solving a stiff set of differential equations. Deep neural networks (DNNs) can be used to represent the nonlinear changes in the physical and chemical processes of aerosols. However, their use is limited due to several challenges such as generalizability, preservation of mass balance, simulating sparse model outputs, and maintaining physical constraints.  To address these challenges, we have developed a physics-informed DNN approach that can simulate the complex physical and chemical formation processes of isoprene epoxydiol SOA (IEPOX-SOA) over the Amazon rainforest. The DNN is trained over a short period of 7 hours of simulated IEPOX-SOA over the entire atmospheric column using the Weather Research and Forecasting Model coupled with Chemistry (WRF-Chem). The trained DNN is then embedded within WRF-Chem to replace the default solver of IEPOX-SOA formation, which is computationally expensive. The approach shows promise, as the trained DNN generalizes well and agrees with the default model simulation of the IEPOX-SOA mass concentrations and its size distribution over several days of simulations in both dry and wet seasons. Additionally, the computational expense of WRF-Chem is reduced by a factor of 2. The approach has the potential to be applied to other computationally expensive chemistry solvers in climate models, which could greatly speed up the models while maintaining complexity.