Google DeepMind says WeatherNext model adds a day of lead time to cyclone forecasts
A model trained on 20 terabytes of atmospheric data and 5,000 storms now beats physics-based forecasts, a case study in what learned models can do.
Google DeepMind and Google Research on 6 August 2026 published results for WeatherNext Cyclones, a machine-learning model that forecasts tropical cyclone track, intensity and wind structure. According to the post, the model delivers more than 24 hours of additional lead time compared with existing systems: its three-day forecasts match the accuracy that earlier models achieved only at two days. Track refers to the path of the storm centre, intensity to its peak winds, and wind structure to how those winds are distributed around the centre.
The model was trained on close to 20 terabytes of global atmospheric data together with the IBTrACS historical archive of roughly 5,000 storms, a compiled record of past tropical cyclone observations maintained for research use. DeepMind describes it as co-trained on general weather dynamics and on expert-curated cyclone observations. In benchmarking across the 2023 to 2025 seasons, the company says it outperformed the ECMWF ensemble on track forecasts and the HWRF model on intensity forecasts.
WeatherNext Cyclones runs on a 28 km grid, which the post says is about 100 times coarser than conventional physics-based cyclone models, and can produce a 1,000-member ensemble in under a minute on a single TPU. DeepMind characterises the gain as roughly a decade's worth of progress in forecast skill.
The US National Hurricane Center used the model during the 2025 hurricane season, including for advance predictions of Hurricane Melissa, which the post says gave communities extra preparation time. DeepMind also said WeatherNext 2, WeatherNext Cyclones and WeatherNext 2-mini are now freely available on GitHub for academic research and operational forecasting.
Why it matters
Operational forecasting has long belonged to numerical models that solve physics equations on national supercomputers. A learned model that runs on one accelerator and beats those systems on a high-stakes target shows machine learning becoming an operational tool in the physical sciences. Open release also lowers the barrier for agencies without large computing budgets.
Operational forecasting has long belonged to numerical models that solve physics equations on national supercomputers.
Ingest the global atmospheric state
Current observations and analysis fields on a 28 km grid form the starting point.
Perturb the starting conditions
Many slightly different initial states are generated to represent uncertainty.
Run the learned model on a TPU
A 1,000-member ensemble is produced in under a minute, per DeepMind.
Read track, intensity and wind structure
The spread across members gives a probability cone rather than a single line.
Forecasters interpret and issue guidance
The US National Hurricane Center used the model during the 2025 season.
What you can learn from this
- Learned forecasters replace equation solving with pattern mapping. A numerical weather model divides the atmosphere into a grid and steps the physics forward in small increments, which is accurate but slow. A machine-learning model instead learns a direct function from past atmospheric states to future ones, so a forecast is a single forward pass through a network. That is why the same prediction that takes hours on a supercomputer can take seconds on a TPU. The trade-off is that a learned model can only reproduce behaviour represented in its training data, whereas a physics model generalises by construction.
- Ensembles turn one forecast into a distribution. Small errors in the starting conditions grow over days, so forecasters run many slightly perturbed copies and read the spread as uncertainty. When each member is cheap to compute, running 1,000 of them becomes practical, and the cone of possible tracks gets a much better-defined probability. Cheap inference is therefore not just a convenience; it changes what kind of forecast can be issued. Agencies typically publish the central estimate alongside the spread so that planners can prepare for the tail as well as the most likely outcome.
- Rare events need curated labels, not just raw data. Cyclones are a tiny fraction of all weather, so a model trained only on general atmospheric fields sees few examples of them. Co-training on an expert-maintained storm archive gives the network a concentrated signal for the exact quantities that matter, such as the storm centre and peak winds. This is the general principle that specialised, well-labelled data improves skill on the tail of the distribution.
- Grid resolution and forecast skill are not the same thing. A coarser grid stores less detail, yet a learned model can implicitly capture the effects of sub-grid processes because it has seen how they played out in years of observations. That is why evaluation against independent benchmarks, rather than nominal resolution, is the honest measure of a model. It also explains why the comparison here is stated in terms of lead time and error against named baselines. Nominal resolution still matters for local detail such as rainfall over hills, which is why coarse global models are usually paired with finer regional ones.
- Temporal hold-out sets prevent leakage. Weather from consecutive days is highly correlated, so randomly shuffling samples into train and test sets would let the model peek at near-duplicates. Evaluating on whole seasons that were never seen in training, as this work does for 2023 to 2025, gives a realistic estimate of performance on future storms. The same rule applies to any time-series model you build.
We teach this
How to use this in practice
- Load a public weather dataset and plot one variable. Download a small sample of reanalysis or station data from a national meteorological archive, read it into a notebook with pandas or xarray, and plot surface pressure or wind speed over a month. Keep the file small; a single station or grid point over one month is enough to see the structure without waiting on large downloads. Note the sampling interval and any gaps, since missing data is the first thing a forecasting model has to cope with. Done looks like a saved chart with labelled axes and a sentence describing one pattern you can see.
- Build a persistence baseline and try to beat it. For the same series, predict that tomorrow equals today, compute the mean absolute error, then fit a simple regression on the previous few days and compare. Done looks like a two-row table showing the error of each method, which is exactly how forecasting papers frame their claims. If the regression cannot beat persistence, that is a useful result too, because it shows how strong the trivial baseline is at short lead times. Try a three-day lead time as well and watch how quickly the error grows.
- Clone the WeatherNext repository and read its documentation. Note what inputs the model expects, what it produces, and which hardware it targets, without trying to run it. Look for an evaluation script and note which baseline models it compares against, since that tells you how the authors define success. Done looks like a one-page summary in your own words of the input shape, output variables and the licence terms.
- Draw the ensemble pipeline. On paper, sketch an initial atmospheric state, an arrow fanning out into many perturbed copies, a forward pass for each, and a final distribution of cyclone tracks. Label where uncertainty enters and where it is summarised. Done looks like a diagram you could use to explain to a non-technical friend why forecasts are shown as a cone rather than a line. Add a second sketch showing how the spread widens as lead time grows from one day to five, and label the point where the cone becomes too wide to be useful for decisions.
Sources
- WeatherNext: AI model achieves breakthrough in forecasting cyclones — Google DeepMind
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report