Google's GenFocal turns coarse climate projections into local, multi-day weather, Nature Machine Intelligence paper says

Researchers at Google Research, one of them also at Caltech, have published GenFocal, a generative AI method that turns coarse global climate projections into fine-grained, time-consistent weather for a chosen region. The paper, "Regional climate risk assessment from climate models using probabilistic machine learning", appeared online in Nature Machine Intelligence on 28 September 2026 and is open access.

What it does

Global climate models run on coarse grids, so they say little about the local, multi-day weather that drives risk. The paper's abstract says GenFocal "generates statistically accurate, fine-scale weather from coarse climate projections without requiring paired simulated and observed events during training". Free-running climate simulations never line up day-for-day with real weather, so the paired training most machine-learning downscalers rely on is unavailable.

GenFocal works in two stages:

• Debiasing: a rectified-flow model maps daily climate-model output at 1.5° onto the statistics of coarse-grained real weather.
• Super-resolution: a conditional diffusion model adds detail, taking the data to 0.25° and from daily to 2-hourly steps within a regional patch, with overlapping time windows stitched together so long sequences stay consistent.

It was trained on 1980–1999, using the ERA5 reanalysis as the high-resolution target and the CESM2 Large Ensemble (LENS2) as the coarse source, tuned on 2000–2009 and evaluated on 2010–2019.

What the paper reports

The main-text comparisons are against two statistical methods: BCSD, which the authors describe as routinely used for downscaling CMIP ensembles, and STAR-ESDM, which they call "a state-of-the-art method recommended for use in the US Fifth National Climate Assessment". The authors say in the supplement they picked these because "Most existing ML-based downscaling methods are inapplicable for climate risk assessment" without time-aligned data. Results are scoped as follows:

• Summer heat, conterminous US (June–August, 2010–2019): for the 99th percentile of the heat index, the paper reports "an average bias reduction over 35% with respect to the statistical downscaling baselines, which systematically underestimate risk". On 5-day "extreme caution" heat streaks, "GenFocal reduces average bias by 44% and 57% compared to BCSD and STAR-ESDM, respectively." On the tail dependence of temperature and humidity extremes, it reports an average error reduction of 44% against the two.
• Single-variable statistics: the supplementary Table SI 3 (summer, CONUS) has 12 rows (four variables, three error metrics) across five methods. GenFocal has the lowest error on all six temperature and humidity rows; on wind speed and sea-level pressure, at least one of BCSD and STAR-ESDM scores lower than GenFocal on every row, and a quantile-mapping variant of GenFocal scores lowest on two of them.
• Tropical cyclones, North Atlantic (2010–2019): the paper says GenFocal produces storm tracks, genesis locations, landfalls, frequency and intensity consistent with ERA5, including storms "largely absent" from the coarse input. The supplementary information adds that pressure depressions are "systematically underestimated in downscaled projections", so storm output is calibrated to observed frequencies over the training period. That step is applied to all methods compared, and the authors say GenFocal needs the smallest adjustment.
• Future projections: for cities in the western US over 2020–2080, GenFocal's regional warming pattern is compared with physics-based downscaling (WRF, from the WUS-D3 dataset) and matches it "more consistently than other methods", which the paper says predict near-uniform warming. For 2050–2059 it projects more North Atlantic landfalls on the US east coast, which the authors say "aligns with forecasts from other downscaled climate projections, such as the Risk Analysis Framework for Tropical Cyclones model" and describe as contributing to "the ongoing scientific investigation".

The abstract's wider claim, that GenFocal "samples high-impact, rare events more accurately than leading methods", rests on these comparisons.

Cost, code and disclosures

• Compute: training each stage took about three days on TPU v5p hardware. By the supplement's own estimate, debiasing the full 100-member ensemble over 140 years took about nine hours on 100 nodes of eight H100 GPUs, roughly US$36,000 at an assumed $5 per GPU-hour.
• Code and data: source code is on GitHub under Apache 2.0 and on Zenodo (CC BY 4.0). Model weights are on Zenodo (CC BY 4.0) and Google Cloud, and training data and downscaled outputs are on Google Cloud. The repository describes itself as "not an official Google product".
• Authors: Zhong Yi Wan, Ignacio Lopez-Gomez, Robert Carver, Tapio Schneider, John Anderson, Fei Sha and Leonardo Zepeda-Núñez, all listed at Google Research. Schneider is also at Caltech. Anderson is now at General Motors and Sha is now at Meta.
• Funding: the paper's funding statement reads, in full: "This work was supported by Google LLC."
• Competing interests: "The authors declare no competing interests."
• This piece is based on the published version and its supplement, not the 2024 arXiv preprint.

Why it matters

The paper says the resolution gap limits risk assessments for infrastructure design, energy planning, flood forecasting and insurance pricing. On our reading, the results so far cover two regions (CONUS for heat, the North Atlantic for cyclones) and one climate model's ensemble, CESM2's LENS2. The main-text baselines are statistical methods rather than other AI downscalers, and the model was built and funded by Google. Independent tests on other regions and climate models would show how far the gains carry over.