Diffusion-Advection Transformer for Air Quality Prediction

Luyang Zhang, Chunbo Luo, Geyong Min

Keywords: Spatiotemporal Forecasting, Air Quality, Physics-informed Learning, Transformers, Diffusion and Advection

TL;DR: A physics-informed Transformer that learns temperature-conditioned diffusion and wind-driven advection to improve long-horizon PM2.5 forecasting across regions.

Abstract: Air pollution is a major concern for public health and the environment globally, which highlights the need for effective monitoring and predictive modeling to mitigate its impact. Although data-driven models have shown promising results in air quality prediction, they still struggle to model the underlying physical mechanisms of pollutant dispersion, where diffusion governs small-scale spreading and advection drives large-scale directional transport. To address this limitation, we propose the Diffusion-Advection Transformer (DA-Transformer), a novel physics-informed architecture. Specifically, the model integrates the two key physical mechanisms by embedding diffusion and advection as differential equation-based components. These physics-informed modules are incorporated into a Transformer framework to enable the model to better capture pollutant transport dynamics, such as local diffusion-driven smoothing and wind-induced directional propagation in air quality data. Experiments on three real-world datasets demonstrate that DA-Transformer consistently outperforms baseline models in $\mathrm{PM}_{2.5}$ concentration prediction and achieves substantial gains over its variants that exclude diffusion and advection in their model design.

Primary Area: learning on time series and dynamical systems

Submission Number: 25577