Researchers built Prithvi-Precip, a global AI system that aims to give better forecasts of precipitation up to 96 hours (four days) ahead. The main idea is simple: train the AI on satellite-derived rainfall estimates instead of only on reanalysis products, and also let the model directly consume raw satellite observations. The team shows these steps can make precipitation forecasts more accurate, especially at short lead times and in the tropics and subtropics.

The work starts from an existing foundation model called Prithvi-WxC, a transformer-based atmospheric AI that was pretrained on the MERRA-2 reanalysis. The authors attach a small precipitation prediction head to that model and then fine-tune it for six-hourly precipitation fields up to 96 hours ahead. They explored different training styles and found that autoregressive rollout training — where the model learns to take a sequence of short steps and uses its own earlier predictions to build longer forecasts — produced substantially better results than conditioning the model directly on the requested lead time.

To improve the quality of what the model learns, the team tried replacing reanalysis precipitation targets (MERRA-2) with satellite-derived precipitation estimates from the IMERG product (Integrated Multi-satellite Retrievals for GPM). They also developed a sensor-agnostic way to feed raw satellite observations into the model. This encoding is designed to accept many types of satellite data, such as visible and infrared measurements from geostationary satellites and microwave observations from low‑Earth‑orbit sensors, without needing a different input format for each instrument.

They evaluated the system against independent precipitation measurements: gauge‑corrected ground radar over the contiguous United States and direct gauge observations over Brazil. Training on satellite-derived precipitation targets yielded better forecast accuracy than training on MERRA-2 precipitation fields. Ingesting raw satellite observations provided further improvements at short lead times. According to the paper, the biggest gains from direct satellite ingestion appeared in tropical and subtropical regions. The authors also report that the combined changes let Prithvi-Precip improve upon comparable forecasts from the Goddard Earth Observing System forward processing (GEOS FP) system.