Up to 24 hours of maturity, FitzRoy equals or exceeds the state of art without the blurred rain fields typical of AI, and reports record cumulations of the Texas and Missouri floods.
Predicting precipitation remains a puzzle: physical models do not solve mesoscale convection well, while machine learning often produces fuzzy, physically inconsistent maps that underestimate extreme events. FitzRoy, designed for the continental United States, claims to be the first system to reconcile accuracy and realism.
Its principle: a conditional discrete diffusion model, then refined by reinforcement learning, which makes it possible to no longer sacrifice the fidelity of rainfall structures to point-by-point accuracy. Over timeframes of up to 24 hours, it equals or surpasses the best models while maintaining realistic rainfields, and its low latency makes it usable in real time.
On concrete cases, such as the Texas Hill Country floods in 2025 and those in Missouri in 2026, FitzRoy announced record accumulations that existing models had largely missed.
Read the full article on the original source
Go to the source →