climate-weather/ai produced the result/Nature 2021 · v2
Generative AI model forecasts UK rainfall 90 minutes ahead from radar images
Researchers trained a generative neural network on UK weather radar records from 2016 to 2018, then used it to produce sets of possible rainfall maps for the next 90 minutes, judged against existing methods and by 56 Met Office forecasters.
spectrum · one line per step, placed by what the step does · bright lines used AI
Skilful precipitation nowcasting using deep generative models of radar
Nature, 2021
doi:10.1038/s41586-021-03854-z · record aix-00298 v2 · checked 2026-11-05
- AI was for
- Forecasting
- Model family
- Generative adversarial network, Convolutional neural network, Recurrent neural network, Transformer
- Checked by
- Held-out66851 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Rain radar sweeps the sky and returns a picture of where water is falling, updated every few minutes. Turning those pictures into a short-range forecast is called nowcasting: predicting the next hour or two, the window that matters for deciding whether to delay a flight, close a road or warn about flash flooding. Full physics-based weather models struggle here, because they take time to spin up and work on grids coarser than a shower. The traditional alternative is to assume rain simply drifts along with the wind, which works briefly but cannot invent the growth and decay of individual storms. Rain is also patchy and uncertain, so a single map is rarely enough.
The team set out to build a nowcasting system directly from the radar archive itself. They used observations from the Met Office RadarNet4 network, maps of surface rainfall rate at five-minute intervals, cropped into sequences and deliberately reweighted so that heavier rain appeared more often in training. The aim was to produce not one forecast but an ensemble, several plausible rain maps for the same moment, out to 90 minutes, and then to ask both statistical measures and working forecasters how the results compared with existing methods.
Where AI came in
The forecast itself is the AI's output. The researchers trained a conditional generative adversarial network, called DGMR, a pair of networks in competition: one invents future radar fields, while others try to tell invented fields from real ones. Here there were two such critics, one judging single maps and one judging whole sequences for realistic motion, plus a term penalising departures from the observed rainfall. The model reads four past radar fields, draws random numbers to represent the many possible futures, and emits 18 future fields covering the next 90 minutes. Training ran on 16 tensor processing unit cores for one week.
What the AI stood in for is the conventional nowcasting algorithm. The comparison, PySTEPS, works by tracking and advecting rain patterns with the estimated wind. Two further neural networks were trained as baselines, a UNet and a model using axial attention adapted from earlier work. All were scored against held-out radar from 2019 using standard forecast measures, and their outputs were shown anonymously and in random order to the forecaster panel for ranking. Code, trained weights and the processed data were released.
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
A conditional generative adversarial network, DGMR, was trained on UK radar composites from 2016–2018 to sample ensembles of future precipitation fields out to 90 minutes, and was evaluated on 2019 data against PySTEPS, a UNet and a radar-only axial attention model adapted from MetNet. Verification used critical success index, continuous ranked probability score, radially averaged power spectral density and a relative economic value analysis, alongside a ranked-preference study with a panel of 56 operational Met Office meteorologists. In that study the generative nowcasts were ranked first by 89% of meteorologists for medium-rain cases (above 5 mm h−1) and 90% for heavy-rain cases (above 10 mm h−1). The paper reports that at T+90 min the effective resolution was 32 km for UNet and 16 km for axial attention, while DGMR and PySTEPS matched the observed spectral characteristics; on CRPS at grid resolution DGMR, PySTEPS and axial attention performed similarly.
How AI was used
Radar composites from the Met Office RadarNet4 network, covering 2016–2019 at 1 km resolution and five-minute intervals, were cropped to 256 × 256 sequences and rebalanced by importance sampling toward heavier precipitation. A conditional generative model was trained on these sequences: a conditioning stack encodes four past radar fields, a latent conditioning stack reshapes Gaussian samples, and a convolutional GRU sampler emits 18 future fields. Training combined a spatial discriminator (a convolutional network over individual fields), a temporal discriminator (a 3D convolutional network over sequences), and a grid-cell regularization term penalizing deviation between the predictive mean and the observed sequence, optimized with Adam on 16 TPU cores. Two deep learning baselines (a residual UNet trained with a precipitation-weighted mean absolute error loss, and an axial attention encoder–decoder adapted from MetNet to radar only) were trained by the authors, and the advection-based PySTEPS system was run in its default configuration. At prediction time the fully convolutional latent module was applied to full 1,536 × 1,280 fields to sample ensembles, with an optional post-processing step that widens the latent distribution and applies a gamma-distributed multiplicative perturbation to the input radar. Outputs were scored against held-out 2019 radar with paired permutation tests and were also presented, anonymized and randomly ordered, to operational meteorologists for ranked preference assessment.
The shape of the work
Structural · the record, drawn
Assemble radar composite archive
Obtaining raw data, whether by measurement, download or retrieval.
we use a collection of radar composites from the Met Office RadarNet4 networkwhere the paper describes this · verbatim
Crop, rebalance and split dataset
Cleaning, filtering, normalising or labelling data already obtained.
the dataset is rebalanced to include more data with heavier precipitation radar observationswhere the paper describes this · verbatim
Train generative nowcasting model
Fitting model parameters, including fine-tuning an existing model.
The nowcasting model is a generator that is trained using two discriminators and an additional regularization term.where the paper describes this · verbatim
Train deep learning baselines
Fitting model parameters, including fine-tuning an existing model.
The model is trained with batch size eightwhere the paper describes this · verbatim
Sample ensemble nowcasts
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
The sampler makes predictions of 18 future radar fields (the next 90 min).where the paper describes this · verbatim
Post-process for reliability
Cleaning, filtering, normalising or labelling data already obtained.
we apply a stochastic perturbation to the input radar by multiplying a single constantwhere the paper describes this · verbatim
Quantitative verification on held-out year
Testing outputs against ground truth.
models are trained on years 2016–2018 and evaluated on 2019where the paper describes this · verbatim
Expert meteorologist preference study
Testing outputs against ground truth.
We conducted a two-phase experimental study to assess expert judgements of value, involving a panel of 56 experts.where the paper describes this · verbatim
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.
The nowcasts that the paper's claims are about are produced by the trained generative model; the verification and expert-preference results all describe its outputs
Learning is framed in the algorithmic framework of a conditional generative adversarial network (GAN), specialized for the precipitation prediction problem.where the paper describes this · verbatim
The subsampling reduces the size of the test set to 66,851where the paper describes this · verbatim
we have also released a pretrained version of the generative modelwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Version of DGMRWhich version of the model was used is not stated.
- Version of UNet baselineWhich version of the model was used is not stated.
- Version of Axial attention baseline (radar-only adaptation of MetNet)Which version of the model was used is not stated.
- Version of PySTEPSWhich version of the model was used is not stated.
- What step 3 replacedThe paper gives no basis for what the AI stood in for.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
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-00298, 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). Generative AI model forecasts UK rainfall 90 minutes ahead from radar images. Record aix-00298, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00298/