Aerosols—tiny particles such as wildfire smoke, desert dust, sea salt, and volcanic sulfates—affect visibility, human health, and the climate by reflecting or absorbing sunlight and seeding clouds. But their diverse properties and ability to travel across continents make them difficult to represent in weather and climate models, adding uncertainty to forecasts.
Traditionally, models rely on preprocessed satellite data that indirectly capture aerosol properties, an efficient but less accurate approach. Directly assimilating raw satellite radiance has required too much computing power to be practical. Researchers now report a method that uses a pretrained AI component to quickly interpret how aerosols and sunlight reflected from Earth's surface influence satellite observations, making direct assimilation feasible.
In tests against real-world satellite observations in China, the new approach reduced aerosol prediction errors by about 50% compared with standard simulations. The technique, described in the Journal of Advances in Modeling Earth Systems, could offer a practical way to improve aerosol simulations and, in turn, boost the accuracy of weather and air-quality forecasts.