Abstract Accurate forecasting of PM 2.5 concentrations is essential for air quality management and public health protection. Previous methods for correcting biases in chemical transport model simulations are computationally expensive, limiting their operational applications. We introduce the Transport-Informed Graph Neural Network (TransNet), a machine learning model that forecasts PM 2.5 concentrations for +72-h by learning coupled Advection-Diffusion-Reaction (ADR) operators, enabling efficient operational deployment. Evaluated over 170 stations across South Korea, TransNet outperforms the bias-correction model Adaptive Graph Attention Network (AGATNet) at short-term lead times ( + 1 h to +48 h), achieving initial MAE of 2.60 μg/m 3 compared to AGATNet’s 4.27 μg/m 3 . However, AGATNet demonstrates greater stability at extended lead times (>+48 h) and better captures extreme events. Operator analysis reveals that coupled advection-diffusion systems are essential for capturing transport dynamics, with error increment by 153.5% at +1 h, while isolated advection or diffusion produced 72-73% error increment. TransNet represents a physics-informed framework with improved capability for operational PM 2.5 forecasting.
Dimri et al. (Thu,) studied this question.