AI improves storm forecasts without a supercomputer

An AI tool developed at Brazil’s University of São Paulo has improved short-term storm forecasting in tests, cutting false alarms by nearly 70% compared with a traditional method. Called TITAN-LSTM, it predicts a storm’s path, rain-covered area and rainfall intensity up to 30 minutes ahead, updating every five minutes.

The tool pairs established radar-tracking software with AI that learns how storms change over time. Rather than processing entire radar images, it works with numerical descriptions of each storm, such as its size, speed and rainfall intensity. This lighter computing workload means it can run on an ordinary laptop.

Researchers used radar observations collected near São Paulo between 2016 and 2019. From more than 32,000 tracked storms, they selected 439 whose full development could be followed, using 307 to train the AI and keeping 132 separate for testing. The findings were published in the Journal of Geophysical Research: Machine Learning and Computation.

There are important limits: the tool currently handles only storms that neither merge with others nor split apart, and uncertainty increases further into the forecast. The team is now working on more complex storms and hopes to adapt the system to other radars, with the aim of supplying forecasts to agencies responsible for disaster warnings.