AeroGraph-PINN: Physics-Informed Neural Networks for Air Quality and Pollution Prediction

Abstract

Predicting air pollution accurately is vital because pollution directly impacts human health. However, generating reliable pollution maps is difficult when physical monitoring stations are limited. A standard deep learning model learns strictly from observed data, often producing unrealistic predictions in unmonitored spatial gaps. In this work, I propose AeroGraph-PINN, a Physics-Informed Neural Network that combines deep learning with the 2D Advection-Diffusion Partial Differential Equation (PDE). Evaluated on a synthetic 50,000-point spatiotemporal domain using 200 sparse sensor observations and 2,000 collocation points, AeroGraph-PINN achieves an MSE of 0.0315 and an R^2 score of 0.9274, outperforming purely data-driven baselines.