Predicting rain, especially heavy downpours, is hard. Physics-based weather models run on coarse grids and miss small, intense storms. Machine learning models often produce blurry forecasts that underestimate big events because they smooth out details to be more accurate on average. There has been a trade-off: you could have realistic storm shapes or correct average rainfall, but not both.
FitzRoy, a new system, tackles this by combining two techniques. First, a learns to generate realistic precipitation patterns, starting from random noise and gradually refining into plausible rain fields. Then, , trial and error guided by a reward, fine-tunes the model so that its forecasts also match observed rainfall amounts. This preserves the sharp edges and fine detail of real storms while improving accuracy.
For forecasts up to 24 hours ahead over the United States, FitzRoy beats both physics-based and other machine learning models in skill and realism. It correctly forecast record-breaking rains in flood events that other models missed. This means more reliable warnings for flash floods, giving emergency managers extra time to act.