Abstract We present AI-DUST, a deep learning model for dust forecasting directly driven by AI-generated weather forecasts. Integrating a Multiple Stacked Graph Attention Network with physical constraints and a physics-based emission scheme, AI-DUST captures key atmospheric physical processes without relying on traditional numerical dust modeling chains. The model demonstrates exceptional accuracy in reproducing a traditional dust model, with correlations >0.99 (one-step) and >0.61 (80-step). In real-time forecasts of 2025 spring sand and dust storms (SDS) over East Asia, AI-DUST outperformed operational models, achieving a 27% higher Threat Score (TS) in 48-hour predictions across 14 strong events. Its 10-day forecast TS exceeds 0.22, demonstrating strong long-term capability. The model generalizes well to unseen regions like the Sahara, enabled by its architecture and standardized preprocessing. This work demonstrates the feasibility of building atmospheric environmental forecasting systems directly driven by AI-generated weather forecasts, paving the way for new, efficient AI-driven chemistry and transport models.
Wang et al. (2026) studied this question.