The forecast is better than ever. The physics is not.

The forecast wins

The 2026 hurricane season is the one where the AI forecast stopped being an experiment. Brightband's analysis of Google DeepMind's performance finds it outperformed operational GFS, HAFS-A, HAFS-B and National Hurricane Center forecasts in track forecast errors for the East Pacific season, with the highest frequency of superior performance for track forecasts. Performance in the Atlantic was similarly strong, with DeepMind the best-performing model for track forecasts at most lead times. The institutional seal is the last line of the analysis: Google DeepMind is now consistently used in the Hurricane Center forecast process. The AI model is no longer competing with the forecasters. It is part of how they forecast. m:…2357

The same analysis carries the caveat that keeps the story honest. Despite its overall accuracy, Google DeepMind displayed a consistent negative intensity bias in the East Pacific season — it underestimates how strong the storm will be. Atlantic conclusions are limited by the season's low activity and smaller sample size. The model is better at where the storm goes than at how strong it will be when it gets there. For anyone whose decisions depend on intensity — evacuation orders, insurance, port closures — the track record is not the whole record. m:…2357

The simulation scales

The second paper of the week scales the same bargain up to the whole planet. STRATA achieves 48 simulated days per megawatt-hour — a 50 times energy efficiency gain over the SCREAM physics model — and 741 simulated days per wall-clock day on 512 H100 GPUs. Kilometre-scale emulation still requires roughly ten times more floating-point operations per grid point than coarse-resolution AI weather models, and large-scale biases develop in STRATA rollouts with lead time. The simulation is fifty times cheaper to run than the physics, and it drifts as it runs. Both facts are true at once, and the second is the price of the first. m:…6585

The physics check

The third paper exists because of that drift. PhysXAI introduces a diagnostic framework to identify what its authors call physical hallucinations: outputs that violate fundamental laws such as mass conservation in modern weather foundation models. The framework decouples forecast error from physical plausibility, so a forecast can be scored on two axes at once — how close it is, and whether it obeys physics. It supports detection and attribution across GraphCast, Pangu-Weather and NeuralGCM, and its latent representation is approximately isometric to the physical residual space, which means the model's internal geometry tracks the physics it is supposed to respect. The point of the tool is not to replace the forecasters. It is to tell them which parts of a good-looking forecast are physically impossible. m:…6583

Put the three together and the state of AI weather is clear. The forecast is now good enough to sit inside the Hurricane Center's process, the simulation is cheap enough to run at planetary scale, and the field has built the instrument that catches the models breaking the laws of physics. The bargain of AI weather is accuracy without obedience, and the work this week is about closing that gap — measuring the bias, pricing the drift, and mapping the hallucination. The forecast you can trust is the one that has been checked twice: once against the storm, and once against physics. m:…2357 m:…6583