Enhancing Air Quality Simulations With Neural Downscaling Architectures

High-resolution air pollution datasets are crucial for exposure assessment and policy support but are computationally demanding to produce with traditional models. We present a machine learning-based framework for spatial downscaling of chemistry-transport model outputs from 6 km to 1 km horizontal resolution over a large European domain. The model is trained on base case simulations and validated across multiple emission scenarios and temporal scales. Results show good performance in reproducing both absolute concentrations and scenario-induced differences. Applying the model to monthly-averaged fields offers a computationally efficient solution, supporting rapid scenario analysis.