Deep Learning for Accurate Global Dust Aerosol forecasting

Dust aerosol forecasting is of significant scientific and societal importance. Currently, the most accurate forecasting systems rely on numerical weather prediction methods, which solve differential equations to simulate the physical and chemical processes of dust aerosols and predict dust concentrations. However, errors introduced by initial and boundary conditions, along with the complex nonlinear interactions between aerosol physical-chemical processes and atmospheric dynamics, result in uncertainties and high computational costs in numerical prediction methods. In recent years, artificial intelligence (AI) methods have demonstrated significant potential in the field of weather forecasting. However, AI-based approaches for addressing such challenging extreme weather events remain in their infancy. Here, we introduce DustWatcher, an AI-based forecasting method for global dust aerosols. DustWatcher integrates spatiotemporal Transformers with conditional generative networks to develop a neural network framework that optimizes forecasting errors in an end-to-end manner. Compared to current state-of-the-art global and regional aerosol forecasting systems, DustWatcher, trained on 41 years of global reanalysis aerosol data, delivers more accurate deterministic forecasts for most aerosol variables including surface dust concentration and aerosol optical depth. DustWatcher provides skillful forecasts every three hours for the next seven days at a resolution of 0.5°×0.625°. Our results demonstrate the potential of AI in improving the dust forecasts accuracy and advancing its application in dust aerosol forecasting field.