Scientists from the University of the Philippines Diliman have created a new artificial intelligence model that provides a faster way to predict rainfall from typhoons.
The research was led by Cris Gino Mesias and Dr. Gerry Bagtasa from the Institute of Environmental Science and Meteorology. Their work is based on the observation that tropical cyclones often follow repeating patterns, where similar storms moving along comparable tracks tend to produce similar amounts and distribution of rain.
The AI model was designed to learn from this history by linking the tracks of past typhoons to their recorded rainfall data. This approach is fundamentally different from the complex dynamic models currently in use. Dr. Bagtasa explained that those traditional models require immense computing power and high-performance infrastructure, making them difficult and resource intensive to run.
In contrast, the new AI model can generate its predictions within minutes on a standard laptop. Despite its simplicity, its forecasting skill was found to be comparable to the conventional dynamic models. Dr. Bagtasa noted that the AI model even performed better at predicting episodes of extreme rainfall, which are critical for disaster preparedness.
The research identified that a storm’s distance from a specific area and how long it lingers are the most important factors for predicting its rainfall. This means a typhoon near Luzon is unlikely to cause heavy rain in Mindanao, and a slow moving system will generally produce more total precipitation over the places it passes.
Dr. Bagtasa clarified that this tool is not perfect but serves as a valuable addition to the existing suite of forecast models. It can provide disaster managers with quicker, additional information on impending hazards. The model can also be updated with new data, allowing it to learn and improve its accuracy over time.
He also distinguished this practical AI from the well known large language models like ChatGPT. He emphasized that not all AI is the same, noting that while some models like this one are efficient, others consume vast amounts of energy. The study was published in the journal Meteorological Applications and supported by two DOST agencies.