Development of Spatiotemporal Physically-informed Top-Down NH3 and NOX Emissions over the U.S. using an AI Machine-Learning Inverse Modeling System with Observations

Timely accurate estimation of NH₃ and NO_X emissions plays a critical role in forming PM_2.5 concentrations in the atmosphere. While the bottom-up method can provide an averaged value, the satellite-based top-down methods can generate near-real-time constraints on emissions; however, the existing numerical models (e.g., chemical transport model, CTM) can be computationally expensive, and its efficiency can be largely limited by efforts in dealing with the complex emission-concentration response. However, the computational burden can be significantly improved with the use of a deep neural network trained with CTM simulations, note as DeepCTM. We apply this novel machine-learning-based method (DeepCTM) using a physically informed variational autoencoder (VAE) emission predictor to infer NH₃ emissions from satellite-retrieved and ground-based concentrations of NO₂ and NH₃. The VAE emission predictor has successfully implemented in NO₂ concentrations with the satellite-retrieved surface NO₂ concentrations. The proven interpretability of the VAE emission predictor will be applied using sensitivity analysis by modulating each feature, indicating that NH₃ and NO₂ concentrations and local meteorology are highly correlated for estimating NH₃ emissions. The advantages of the VAE emission predictor in efficiency, flexibility, and accuracy demonstrate its great potential in estimating the latest spatiotemporal emissions and its future applications.