AI has won Nobel Prizes, but can it save lives and businesses with flash flood warnings?

When all you have is a hammer, the entire world looks like a nail, they say, and if you have a screwdriver instead, the world is screwed. Training in the hypothesis-guided sciences suggests that designing solutions to look for problems may lead to a dead end. However, transformative solutions can upend this common wisdom. Precipitation nowcasting, such as short-term and spatially distributed forecasting of rainfall amounts, especially in a way that provides adequate warning for flash floods, has been known to be a hard problem for decades. When some of us (e.g., Robert Kuligowski and Ana Barros, Rafael Bras and I) tried in the late 1990s and early 2000s to adapt Artificial Intelligence (AI) methods, our efforts were limited by the AI tools of the day, all of which would probably belong to a cyber museum today, much like dinosaurs in a museum of the animal kingdom. However, there was a clear promise. The advent of the AI winter nearly put a stop to this line of work. The AI methods of today trace their origins to the old AI (much like living birds today may trace their lineage to dinosaurs), but are different beasts altogether. From physics and chemistry to transportation, materials, and medicine, and indeed from Turing Awards to Nobel Prizes, AI seems to be ubiquitous and winning. Could precipitation nowcasting be far behind?

The causes of flash flood hazards and the consequent damages are broader than what anthropogenic global warming alone can explain

Just a few years back, Google DeepMind reported in Nature that their physics-free deep generative model of radar improved precipitation nowcasting. A couple of years later, Tsinghua University led an article in Nature, where they reported that incorporating (well-understood even if relatively simple) physics into the representation of generative AI models improved forecasts of convective storms and extreme precipitation, which are critical for flash floods. This hybrid physics-AI approach made us reflect on our related prior work. We began to explore, with stakeholders and collaborators, whether these newer approaches can be made trustworthy for water resources and flash flood managers. Our NASA-funded RAIN (“Remote-sensing data driven Artificial Intelligence for precipitation-Nowcasting”) project led to a paper in a Nature Partner Journal along with the following testimonial from our stakeholders at the Tennessee Valley Authority: “The hybrid physics-AI based precipitation nowcasting method and the corresponding metrics developed by the NASA-RAIN team hold promise for near-term river and flash flood management and is being tested currently on TVA's operational river forecast system.” Our research in this area has also been discussed in United Nations resilience workshops and in follow-on publications. However, preliminary success stories in the literature and anecdotes apart, major barriers remain, such as in improving extremes, reducing bias, generating uncertainty, enhancing interpretability, and adding explainability. The promise is clear, but what is clearer is that there is quite a distance to go.

Flash floods worldwide have been causing thousands of deaths and hundreds of billions of dollars in economic damage annually. As the flash floods during July 2025 in Texas (USA) and October 2024 in Valencia (Spain) suggest, even richer nations are not necessarily resilient. Climate change is often blamed for these devastations, and there is evidence (including in our own work) for warming-induced intensification of precipitation extremes. However, the causes of flash flood hazards and the consequent damages are broader than what anthropogenic global warming alone can explain. Thus, holistic risk management solutions are motivated. Early warning of flash floods is crucial, and precipitation nowcasting is critical, especially (but not only) in situations where the amount of precipitation overwhelms the infiltration capacity. This is where AI, especially when effectively combined with physics and knowledge, is beginning to show promise.