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).
"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.
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.
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.
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.
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.
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.
- Generator (the forger): looks at the last 4 radar frames (20 min) plus some random "inspiration" noise, then paints 18 future frames (90 min).
- Spatial critic: looks at individual frames and asks "does this look like real radar?" This is what stops blur.
- Temporal critic: watches the sequence and asks "does this move like real weather?" This stops jumpy, flickering rain.
- Reality check: a third rule keeps the average of several attempts close to what actually happened, with extra weight on heavy rain. Without it, the forecasts looked real but put the rain in the wrong places.
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.
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.
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.
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:
- Location accuracy (CSI): better than PySTEPS at every rain threshold tested.
- Realistic detail (power spectrum): it keeps the fine-scale texture of real rain, while UNet and axial attention blurred more over time.
- Economic value: the most useful for cost/benefit decisions like "should we deploy flood barriers?"
- It also held up on a separate US radar dataset.
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.
What it still gets wrong
- Heavy rain tends to fade too early by about 90 minutes ahead.
- Isolated pop-up showers are still hard to predict.
- Heavy rain at long lead times remains hard for every method.
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.
Paper: doi.org/10.1038/s41586-021-03854-z · PDF · Code and pretrained model: deepmind-research/nowcasting