DATE/TIME: Friday, November 20, 2026, 2:30 pm PLACE: ENR building, room 223
Chi Li
Washington University in St. Louis
Physics-Guided AI for Air Quality: From Satellite Radiances to Exposure and Emission-Driven Trends
Air pollution is a major environmental risk to human health, yet large gaps in monitoring coverage and incomplete representations of atmospheric processes in models limit our ability to characterize exposure and explain changes in air quality. Artificial intelligence (AI) can help bridge these gaps, but its scientific value depends on whether it captures relevant physical relationships and generalizes beyond its training data. In this talk, I will present how integrating AI with observations and physics-based atmospheric chemistry models advances air pollution measurement, exposure estimation, and attribution.
First, a neural network trained on more than 360,000 radiative transfer simulations retrieves nitrogen dioxide directly from satellite radiances. In cloud-free, polluted scenes, it achieves performance comparable to the operational product at roughly one-tenth the computational cost. Its strengths and limitations are consistent with retrieval theory, illustrating how AI can capture physical relationships. Second, deep learning informed by GEOS-Chem simulations improves estimates of ground-level fine particulate matter from satellite observations relative to conventional AI approaches. Evaluation using buffered-distance cross-validation tests performance as the distance from training monitors increases, demonstrating the value of physical knowledge in GEOS-Chem for estimating exposure in sparsely monitored regions. Third, machine learning separates meteorological influences from observed ground-level ozone variability to reveal emission-driven trends across Chinese cities. Interpreting these trends with atmospheric chemistry modeling helps elucidate the diverse local contributions of changes in volatile organic compound emissions and particulate matter to short-term ozone trends.
Together, these examples show how physics-guided AI can move beyond predictive accuracy toward more generalizable exposure estimates and a clearer understanding of the drivers of air quality change.
Seminar Host:
Deepangsu Chatterjee
Department of Environmental Sciences, SEBS, Rutgers University