Google DeepMind open-sourced inference weights Friday for a six-hour precipitation nowcasting model trained on U.S. radar composites, giving NOAA regional forecast partners a checkpoint they can run on local GPUs instead of only viewing Google-hosted demos. The release bundles validation notebooks, threshold guidance for flash-flood messaging, and a changelog that marks where learned radar fusion outperformed optical-flow baselines on Gulf Coast thunderstorm lines—artifacts evaluators can rerun without trusting a slide deck.

What shipped in the tarball

The package includes ONNX and PyTorch checkpoints, a slim inference server, and geographic masks so offices can clip outputs to county warning areas. Weights cover 0–6 hour lead times at 2 km resolution aligned to NOAA mosaic tiles; DeepMind said training used historical Level-II radar with quality-control scripts published alongside the weights. A separate calibration table maps predicted rain rates to probabilistic exceedance curves forecasters can compare with legacy QPE products.

DeepMind’s blog post emphasized reproducibility: random seeds, train/validation date ranges, and held-out hurricane seasons are documented so NWS science officers can dispute or confirm claims. Google did not release full training code for the largest cluster runs, citing internal infrastructure dependencies, but said fine-tuning scripts for regional adapters are included.

Why NOAA partners asked for local runs

Regional headquarters in Miami, Birmingham, and Lake Charles piloted cloud APIs last year during heavy convection seasons; latency and data-sovereignty concerns pushed teams to ask for on-prem inference when microwave backhaul falters during storms. NOAA’s Atlantic Oceanographic and Meteorological Laboratory helped craft acceptance tests comparing model guidance to forecaster-drawn polygons for marine warnings.

NCEI archives radar volumes that nowcasting models ingest; releasing weights lets offices keep inference inside NOAA firewalls while still using the same numerical core DeepMind validated. That matters when commercial weather vendors market similar AI nowcasts to utilities; public forecasters want parity without recurring API fees.

Claim versus benchmark

DeepMind reports a 12 percent relative improvement in critical success index for ≥1 inch/hour cores in Gulf Coast holdout years versus pySTEPS optical flow, with the largest gains on training lines ahead of cold fronts. The notebooks show failure modes too: false elongation of stratiform shields behind decaying systems, and under-forecasting of extreme cores when radar beams overshoot evolving cells.

NOAA scientists cautioned that machine-learning QPE is not a replacement for human warning meteorologists; it is another guidance layer that must be calibrated to local hydrology. Flash flood partners want timestamps proving when weights were generated so they can align with gauge bias corrections.

Operational guardrails

The inference server defaults to “advisory only” mode—outputs do not auto-publish to wireless emergency alerts. Forecasters must toggle overlays in their AWIPS workflows; DeepMind and NOAA jointly documented color tables that remain color-blind safe and avoid implying precision the model lacks.

Compute requirements are modest by AI standards—a single A100 can serve a region—but smaller offices may share clusters through NOAA’s WCOSS cloud. DeepMind offered a CPU fallback with fourfold latency increase, usable for post-event review rather than live warning desks.

Licensing and community

Weights ship under a research-friendly license that permits government and academic use; commercial redistribution requires a separate agreement. That split mirrors earlier GraphCast releases that academia embraced while insurers waited for vendor packaging.

University collaborators in Oklahoma and Hawaii said they will fine-tune adapters for mountain and island clutter patterns not abundant in Gulf training sets. Open-source contributors must submit regression tests against the published notebooks before merging architecture tweaks—an attempt to keep community forks from diverging silently from NOAA baselines.

What evaluators still cannot see

Training included non-radar features—lightning, satellite water vapor—that the open tarball omits for bandwidth reasons; offices must stream those feeds themselves. DeepMind has not published full ablation on tornado-spin-up signatures; tornado warnings remain outside model scope. Bias correction for winter precipitation types is labeled experimental.

Google’s corporate sustainability commitments do not change NOAA’s need to document energy use when scaling GPU inference during hurricane landfalls; DeepMind listed per-field energy estimates in an appendix, inviting agencies to cap concurrent runs.

Next for regional offices

NWS leadership set a 90-day evaluation window before any office cites the model in service change requests. Miami will stress-test hourly updates during October squalls; Birmingham will compare against human-machine mix experiments already running in the Southeast integrated warning team.

If weights perform, the win is institutional: forecasters keep authorship of warnings while AI sharpens short-fuse rain estimates they already sweat over. If not, the open checkpoints still give NOAA a common benchmark to beat—something evaluators can test without signing another black-box MOU.