The weather model runs. The ML model reads its output. The warning gets better.
Watch-to-warning severe weather guidance (arXiv:2603.20250): the gap between “conditions are favorable for severe weather” (watch, 2-6 hours ahead) and “severe weather is imminent” (warning, 0-60 minutes ahead). The Warn-on-Forecast System runs high-resolution ensemble weather models that produce detailed predictions. The predictions are rich but hard to interpret — 36 ensemble members, updated every 5 minutes, with dozens of fields per member.
ML post-processing: train models on 108 days of archived ensemble output to generate probabilistic forecasts of tornadoes, severe winds, and hail. Two architectures tested: histogram gradient-boosted trees (HGBT) and U-Net (convolutional deep learning). Both outperform a calibrated baseline using updraft helicity alone.
The architectures differ in useful ways. HGBT achieves better overall metrics but caps predictions at ~60% probability — it won't commit to certainty. U-Net extends to 100% and produces spatially smoother guidance. The gradient-boosted model is more reliable (calibrated) but less resolving (can't distinguish near-certain from merely likely). The deep learning model is more resolving but less reliable at the extremes.
The structural insight: the numerical weather model simulates the physics. The ML model translates the simulation into actionable probability. Neither alone is sufficient — the physics model can't produce calibrated probabilities, and the ML model can't simulate physics. The combination produces something neither can: physically grounded, probabilistically calibrated severe weather guidance at the timescale where decisions happen.