Revealing spatiotemporal NOx emission dynamics in the Yangtze River Delta using physics-informed neural networks and TROPOMI observations

Abstract

Accurate estimation of nitrogen oxide (NO_x) emissions is essential for characterizing their impacts on atmospheric environment and supporting air pollution management. This study introduces an innovative inversion framework based on a Physics-Informed Neural Network (PINN), designed to derive NO_x emissions from TROPOspheric Monitoring Instrument (TROPOMI) satellite observations. By embedding the physical laws governing atmospheric processes, this method effectively bridges traditional data assimilation techniques with the computational efficiency of deep learning. Unlike purely data-driven models, it directly integrates physical constraints from atmospheric mass continuity equation into the model training process, thereby avoiding the computational burden associated with chemical transport models (CTMs) and removing the need for CTM simulations for supervision. Application to the Yangtze River Delta region in China (2018–2023) revealed clear spatiotemporal variations in NO_x emissions, capturing the reduction during the COVID-19 pandemic and the subsequent recovery. Emission uncertainties were quantified using Monte Carlo dropout, providing robust error estimates alongside the inversion results. The proposed framework provides a computationally efficient and physically grounded basis for NO_x emission inversion, with broad applicability to air quality assessment and environmental management.