Deep learning-based gas-phase chemical kinetics kernel emulator: Application in a global air quality simulation case

Abstract

The global atmospheric chemical transport model has become a key technology for air quality forecast and management. However, precise and rapid air quality simulations and forecast are frequently limited by the model’s computational performance. The gas-phase chemistry module is the most time-consuming module in air quality models because its traditional solution method is dynamically stiff. To reduce the solving time of the gas phase chemical module, we built an emulator based on a deep residual neural network emulator (NN) for Carbon Bond Mechanism Z (CBM-Z) mechanism implemented in Global Nested Air Quality Prediction Modeling System. A global high resolution cross-life multi-species dataset was built and trained to evaluate multi-species concentration changes at a single time step of CBM-Z. The results showed that the emulator could accelerate to approximately 300–750 times while maintaining an accuracy similar to that of CBM-Z module (the average correlation coefficient squared was 0.97) at the global scale. This deep learning-based emulator could adequately represent the stiff kinetics of CBM-Z, which involves 47 species and 132 reactions. The emulated ozone (O3), nitrogen oxides (NOx), and hydroxyl radical (OH) were consistent with those of the original CBM-Z module in different global regions, heights, and time. Our results suggest that data-driven emulations have great potential in the construction of hybrid models with process-based air quality models, particularly at larger scales.