~/aixsci
weekly updates

climate-weather/ai produced the result/Nature 2024 · v2

Diffusion model trained on past weather produces 15-day global ensemble forecasts

Researchers trained a generative machine learning model, GenCast, on decades of archived weather analyses, then used it to generate 50-member ensembles of 15-day global forecasts and compared them with the forecasts of a conventional physics-based ensemble system.

1. Build ERA5 training dataset2. Train diffusion denoiser in two resolution stages3. Generate 50-member ensemble forecasts4. Train deterministic ML baseline5. Generate perturbed-initial-condition ensembles6. Extract tropical cyclone tracks7. Derive regional wind power8. Verify against analysis and compare with ENS

spectrum · one line per step, placed by what the step does · bright lines used AI

Probabilistic weather forecasting with machine learning
Nature, 2024

doi:10.1038/s41586-024-08252-9 · record aix-00295 v2 · checked 2026-11-05

ai-resultrole of AI
AI was for
Forecasting
Model family
Diffusion model, Transformer, Graph neural network
Checked by
Held-out1320 tested
Code
not reported

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes, never the evidence.

Introduction by AIxSci · plain language

What this research was about

Weather forecasting is uncertain by nature. Small errors in the description of today's atmosphere grow quickly, so a single prediction says little about what will actually happen. Forecasters deal with this by running an ensemble: many forecasts from slightly different starting points or model settings, treated together as a spread of possible futures. If a tenth of the members show a storm hitting a city, that is read as roughly a one-in-ten chance. Producing such ensembles has traditionally meant repeatedly solving the equations of atmospheric physics on a supercomputer, which is costly, and the resulting spread has to be neither too narrow (overconfident) nor too wide (vague).

The researchers set out to produce these probabilistic forecasts a different way: by learning from the historical record rather than by simulating physics. They built a training set from ERA5, the European Centre for Medium-Range Weather Forecasts' reconstruction of past global weather, covering 1979 to 2019 on a grid of about a quarter of a degree, with six surface and six atmospheric variables at 13 heights in the atmosphere. They then assessed the resulting forecasts for 2019 against the record, and against ENS, the centre's own ensemble forecast.

Where AI came in

The forecasts themselves are the machine learning system's output. GenCast is a diffusion model: a network trained to turn random noise into a plausible sample of the data it learned from, here the next global weather state 12 hours ahead, given the two previous states. The network combines a graph neural network, which passes information between points on a mesh wrapped around the globe, with a transformer, which weighs how strongly those points should influence each other. Training ran in two stages, a coarse pass followed by fine-tuning at the finer grid. At forecast time, the model is applied again and again to reach 15 days, and because each run starts from fresh random noise, repeating it gives the ensemble.

So the learned model stands in where the numerical physics simulation normally sits, both the forecast itself and the source of the ensemble spread. The authors also trained a version of the same architecture without the noise machinery, running it as an ensemble by nudging the starting conditions instead, to isolate what the diffusion part contributes. Conventional, non-learned software did the work downstream: an existing algorithm to trace tropical cyclone tracks through the forecast fields, an idealised turbine curve to convert forecast winds into wind power at recorded wind farm sites, and standard statistical tests for the comparisons.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.

The work

Technical · from the record

GenCast is a conditional diffusion model trained on ERA5 reanalysis from 1979 to 2018 that produces global 15-day weather forecasts at 12-h steps and 0.25° resolution, sampling each step autoregressively so that different noise draws give an ensemble of trajectories. The authors generated 50-member ensembles for 2019 initialization times and compared them with ENS, the ECMWF ensemble forecast, using CRPS, spread/skill, rank histograms, Brier skill scores and relative economic value. GenCast had significantly lower CRPS than ENS on 97.2% of the 1,320 variable, lead-time and level combinations evaluated, and the paper also reports comparisons on spatially pooled CRPS, regional wind power CRPS at wind farm groupings, and tropical cyclone ensemble-mean position error and track probability. A deterministic version of the same architecture, run as an ensemble by perturbing initial conditions, served as an ablation of the diffusion component.

How AI was used

A dataset of ERA5 best-estimate analyses on a 0.25° grid (six surface and six atmospheric variables at 13 pressure levels, subsampled to 12-h sequences) was assembled for 1979-2018. A denoiser network (an encoder/decoder graph neural network with a graph-transformer processor of 16 blocks operating on a six-times refined icosahedral mesh, with conditional layer norm for noise-level conditioning) was trained to remove noise, using isotropic Gaussian noise on the sphere, area- and variable-weighted mean squared error, and two previous weather states as conditioning; training ran in two stages, 2 million steps at 1° followed by 64,000 fine-tuning steps at 0.25°. At forecast time, a candidate residual state initialized as pure noise is refined by 20 solver steps of a second-order DPMSolver++2S probability-flow ODE solver with stochastic churn, and the process repeats autoregressively for 30 steps; repeating with new noise samples yields an ensemble. A deterministic variant of the same architecture, trained without noise conditioning to minimize single-step MSE and initialized with Gaussian-process-perturbed analyses, was trained and run as an ML baseline. Non-learned software was used downstream: the TempestExtremes tracker for cyclone trajectories, an idealized turbine power curve for wind power, spatial pooling, and bootstrap significance testing.

The shape of the work

Structural · the record, drawn

PREPARATIONTRAININGSIMULATIONTRAININGSIMULATIONPREPARATIONPREPARATIONVALIDATION12345678AIAIAIAIBuild ERA5training datasetTrain diffusiondenoiser in tworesolution stagesGenerate50-memberensemble forecas…Traindeterministic MLbaselineGenerateperturbed-initial-conditionensemblesExtract tropicalcyclone tracksDerive regionalwind powerVerify againstanalysis andcompare with ENS↤ simulation↤ simulation↤ simulation↤ simulation
AI stepNo AI↤ what the AI stood in for

Build ERA5 training dataset

Cleaning, filtering, normalising or labelling data already obtained.

Our dataset contains the best-estimate analyses of ERA5 for a subset of the available variables, on 13 pressure levelswhere the paper describes this · verbatim
in the paper

Train diffusion denoiser in two resolution stages

Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.

The GenCast results reported in this paper were generated by a model that was trained in a two-stage process.where the paper describes this · verbatim
in the paper

Generate 50-member ensemble forecasts

Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.

GenCast generates an ensemble of stochastic 15-day global forecasts, at 12-h steps and 0.25° latitude–longitude resolutionwhere the paper describes this · verbatim
in the paper

Train deterministic ML baseline

Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.

GenCast-Perturbed is trained by taking the GenCast architecture for fθ described abovewhere the paper describes this · verbatim
in the paper

Generate perturbed-initial-condition ensembles

Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.

We used this model to generate ensemble forecasts (denoted as GenCast-Perturbed) by initializing it using ERA5 analysis perturbed by Gaussian Process noisewhere the paper describes this · verbatim
in the paper

Extract tropical cyclone tracks

Cleaning, filtering, normalising or labelling data already obtained.

we use the TempestExtremes v2.1 cyclone tracking algorithmwhere the paper describes this · verbatim
in the paper

Derive regional wind power

Cleaning, filtering, normalising or labelling data already obtained.

10 m wind speed of forecasts and analysis targets are interpolated at all 5,344 wind farm locations from the Global Power Plant Databasewhere the paper describes this · verbatim
in the paper

Verify against analysis and compare with ENS

Testing outputs against ground truth.

We compare GenCast with ENS on several verification metrics (detailed in Supplementary Information section A.5) computed on our 2019 evaluation set.where the paper describes this · verbatim
in the paper

What the record says

Technical · every part carries its own basis

+ in the paper~ our reading− not reported

How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.

~Role of AI
AI produced the resultour reading

The forecasts the paper's claims are about are produced entirely by the learned diffusion model; without it there is no result

+What the AI was for
Forecastingin the paper
GenCast is implemented as a conditional diffusion model, a generative ML method that can model the probability distribution of complex datawhere the paper describes this · verbatim
+How it was taught
+How results were checked
Held-out1320 testedin the paper
It has greater skill than ENS on 97.2% of 1,320 targets we evaluatedwhere the paper describes this · verbatim
+Code · weights · data
code not reportedweights not reporteddata availablein the paper
This training stage takes a little over 3.5 days using 32 TPUv5 instances.where the paper describes this · verbatim
+Compute
Pre-training took a little over 3.5 days on 32 TPUv5 instances (2 million steps); fine-tuning just under 1.5 days on 32 TPUv5 instances (64,000 steps); a single 15-day forecast takes about 8 min on a Cloud TPUv5 devicein the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated: 4 items
  • CodeWhether the code is available is not stated.
  • Trained model weightsWhether the trained model is available is not stated.
  • Version of GenCastWhich version of the model was used is not stated.
  • Version of GenCast-PerturbedWhich version of the model was used is not stated.

About this article

This paper was chosen for the collection rather than drawn at random, so it is left out of the statistics this site gives about the field. Record aix-00295, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error

Cite this record: AIxSci (2026). Diffusion model trained on past weather produces 15-day global ensemble forecasts. Record aix-00295, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00295/