DGMR explained
Nature · Vol 597 · 30 Sep 2021 · Open access

Teaching an AI to imagine the next 90 minutes of rain

A plain-English guide to "Skilful precipitation nowcasting using deep generative models of radar" by Ravuri, Lenc, Willson et al. (DeepMind and the UK Met Office).

In short: the team trained an AI on years of UK rain-radar maps so it can imagine several realistic versions of the next 1.5 hours of rain. When 56 Met Office forecasters compared it with two other methods, they picked it as the most accurate and useful in 89% of cases.
drag to rotate
A radar rain map, drawn in 3D. Height and colour show how hard it's raining. Blue is light drizzle, yellow to pink is a downpour. (Illustration, not real data.)
1 · The problem

"Will it rain on my street in the next hour?"

Forecasting rain 0–2 hours ahead, street by street, is called nowcasting. Flood warnings, air-traffic control, power grids, outdoor events and emergency services all rely on it.

The big physics-based weather simulators (numerical weather prediction, or NWP) are weak in this window. They need time to "spin up", so the first couple of hours is exactly where they struggle.

Analogy: a weather supercomputer is like a cruise ship. It's great for crossing an ocean (forecasting days ahead) but slow to turn. Nowcasting needs a speedboat.

So nowcasters work directly from weather radar. In the UK a new rain map arrives every 5 minutes at 1 km detail. The task: given the last few maps, draw the next ones.

The nowcasting task. The last 4 radar snapshots (20 minutes of history) are known. The empty frames are the future, up to 90 minutes ahead, that a model must fill in.
2 · How it's done today

Method A: slide the rain along the wind

The standard tool, PySTEPS, works out which way the rain is moving and pushes the current rain map forward like it's on a conveyor belt, with some random wobble added.

Analogy: predicting where a puff of smoke will be by assuming it just drifts with the breeze.

That works for steady rain. But thunderstorm cells are born, grow and die within an hour, and simple drifting can't capture that. In the paper's Scottish storm case study, PySTEPS overestimated the intensity and missed where the rain actually spread.

drag to rotate
Left: reality. Storms drift but also appear and fade. Right: advection only slides the starting pattern along the arrows, so new storms never appear and old ones never die.

Method B: earlier AI, which hedges with blur

Earlier deep-learning models (UNet, and an "axial attention" model adapted from Google's MetNet) learn from radar history and predict rain amounts directly.

They're trained to minimise average error. When they're unsure, the safest answer is a smeared-out average of every possible future, so the further ahead they look, the blurrier the forecast gets.

Analogy: asked where a dropped marble will roll, you shade in the whole floor. You're technically "not wrong", but it doesn't help anyone.
Top: reality. Bottom: UNet. By 90 minutes the paper measured UNet's effective detail at 32 km instead of 1 km (axial attention: 16 km). Heavy downpours, the ones that matter most, get washed out.
3 · The new idea: DGMR

A forger and two art critics

DGMR (Deep Generative Model of Rainfall) is a type of AI called a GAN, short for generative adversarial network. It doesn't predict one averaged answer. It generates a realistic future, like an artist painting the next 90 minutes.

drag to rotate
How DGMR learns. The generator turns 4 past frames plus random noise into 18 future frames. Two critic networks judge them: one checks single frames for realism, the other checks the whole sequence for realistic motion.
Analogy: the forger keeps practising until the critics can't tell its paintings from real photos, and the reality check makes sure it's painting this storm, not just any convincing storm.
~1.5 Mtraining clips from UK radar, 2016–18
1 weekof training on 16 TPU cores
~1.3 sper full-UK forecast on one GPU
2019held-out test year, never seen in training

Rain is rare: about 89% of radar cells are dry, and fewer than 0.4% have heavy rain. So the team deliberately oversampled stormy scenes when building the training set. The model also works on maps bigger than it was trained on, covering the whole 1,536 × 1,280 km UK area.

4 · Many possible futures

Change the noise, get another believable future

Change the random noise and DGMR paints a different but equally realistic future. Generate 20 of them and you have an ensemble, the same idea weather services already use for longer-range forecasts.

If 15 of 20 futures flood your town, that's a 75% risk. Probabilities like this are what forecasters and decision-makers actually need.

drag to rotate
One past, many futures. Each green-labelled tile is a separate DGMR sample for T+60 min. They differ in the details but all look like real rain. Together they show the range of what's likely.
5 · The verdict

56 professional forecasters voted

Met Office meteorologists were shown anonymised, shuffled forecasts of 2019 storms next to what really happened, and asked which was most accurate and useful.

"best envelope" · "representing the risk best" · "much higher detail compared to what [we] are used to at the moment". Forecasters on DGMR.

The forecasters weren't just swayed by nice-looking pictures. In follow-up interviews they described carefully weighing location, extent, motion and intensity. The numbers back them up:

6 · Try it yourself

Compare the forecasters

A simulation of the behaviours described in the paper. It isn't real model output. Drag the slider to look further ahead.

Same storm, four forecasters. Reality, DGMR, PySTEPS and UNet at the chosen lead time.
Lead timeT + 30 min
lightmoderateheavy rain

7 · Not perfect yet

What it still gets wrong

The deeper lesson: standard scores and expert judgment don't always agree. A blurry model can score well on paper and still be useless to a forecaster. The authors call for new ways to measure forecasts that reflect real-world usefulness.

One sentence: by learning from radar the way an artist learns from photos, and being judged for realism in both space and time, DGMR gives sharp, believable, probability-aware rain forecasts for the next 90 minutes, and the people who forecast for a living preferred it 9 times out of 10.

Paper: doi.org/10.1038/s41586-021-03854-z · PDF · Code and pretrained model: deepmind-research/nowcasting