Global Particulate Matter Forecasting Using Lightweight, Region-Specific Deep Learning Models

Ansh Kushwaha, Kaushik Gopalan

Abstract: Global particulate matter (PM) forecasting is critical for air quality management, yet regional variability in emission sources and atmospheric processes poses challenges for unified modeling approaches. We present a lightweight nested-domain deep learning framework using U-Net architecture for short-range forecasting of $PM_{1}$, $PM_{2.5}$, and $PM_{10}$. Our approach trains separate models for 10 different spatial regions and 3 PM species, using overlapping $256 \times 256$ input grids to predict $192 \times 192$ forecast regions with explicit spatial context. Using CAMS reanalysis data spanning 2021--2024, we train independent U-Net models for each region/PM species combination using a model size of $\approx 4$ million parameters per model, for a total of $\approx 120$ million parameters for all the models combined. Evaluated against the $\approx 1$ billion-parameter Aurora foundation model, our framework achieves competitive root mean square error at 6--24 hour forecast horizons while consistently resulting in slightly higher structural similarity indices. These results demonstrate that lightweight, regionally-specialized models offer a viable alternative to large-scale foundation models for PM forecasting, providing computational efficiency without sacrificing forecast accuracy.

Submission Number: 60