Fast, accurate flood warnings are crucial for public safety, but the most detailed computer simulations are too slow for real-time alerts. While artificial intelligence (AI) models are much faster, they typically fail when applied to a new river system different from the one they were trained on.

We developed a new AI called FloodForecaster that overcomes this limitation. We created a special training technique that teaches the AI to transfer its knowledge by learning the fundamental physics of water flow, not just the quirks of one specific location.

When we tested our model, we found that standard methods cause the AI to suffer from 'catastrophic forgetting'—it forgets its original training when learning about a new river. Our technique avoids this, allowing the model to become an expert on both rivers. This resulted in predictions that were nearly 80% more accurate on the new river compared to the standard approach.

This breakthrough makes it possible to rapidly deploy reliable flood warnings for communities where data is scarce, saving critical time and resources and making life-saving forecasts more widely accessible.