Towards Causal Understanding of Urban Air Pollution: Mechanistic Models under Sparse Sensing

Understanding the causal drivers of urban air pollution remains a central challenge for environmental science and policy. While high-resolution source apportionment typically relies on dense monitoring networks or receptor-based chemical analysis, many cities must operate with sparse sensors and incomplete emission inventories. We frame air pollution source attribution as a causal inference problem, linking emissions to observed concentrations through mechanistic dispersion models. Using Gaussian plume formulations, we combine multiple emission categories—vehicular traffic, domestic emissions, brick kilns, industries, and power plants—with real-world sensor data from New Delhi. Our methodology estimates source-specific contributions under sparse observations via parameterized dispersion modeling, while also capturing the influence of missing or unobserved sources. By situating source apportionment within a causal modeling perspective, we emphasize both the opportunities and limitations of mechanistic approaches under real-world constraints, and propose a causal learning framework at the intersection of environmental science and machine learning.