WeatherNext Cyclones Learns Storm Intensity From Labels Its Weather Grid Cannot Resolve

A global weather grid can help predict where a cyclone will travel while understating how strong it will become. The large atmospheric currents that steer a storm extend far beyond its core. Its peak winds are far more localized, so a coarse representation can smooth away the one quantity an intensity forecast most needs.

WeatherNext Cyclones addresses that problem without moving to a finer global grid. It trains on both atmospheric fields and expert-curated records of individual storms, so the model receives explicit targets for properties the atmospheric grid does not represent well.

This article reads the August 2026 sources after the fact. Google DeepMind's blog post of 6 August 2026 presented the system as an advance in track, intensity and wind-structure forecasting. The accompanying Nature paper, Operational tropical cyclone forecasting with AI, published the same day, explains the more interesting design choice: information needed for a prediction does not have to appear as a resolved physical field in the model's input. It can also enter through the training target.

That distinction matters to developers building models for sparse, consequential events. Improving the input grid is one possible investment. Improving what the model is taught to predict is another, with different data and evaluation requirements.

A storm's position and its strength need different evidence

Conventional cyclone forecasting faces a scale dilemma. Global numerical models represent the large-scale environment that governs a storm's track, and the paper describes the ECMWF ensemble as the strongest track guidance for that reason. Regional models such as NOAA's HAFS devote their computation to the fine processes near a storm's core, which helps intensity, but the paper notes this comes at the expense of track accuracy. Simulating everything at fine resolution across the globe is too expensive.

The paper describes a related weakness in earlier AI weather models. Because they learn to predict atmospheric analyses on a 0.25-degree grid, their intensity estimates inherit the low peak winds represented in those analyses. A model can improve at matching its training fields while still missing the observed maximum wind of a real storm.

WeatherNext Cyclones keeps that 0.25-degree grid, roughly 28 kilometres at the equator, and adds a second source of supervision: the International Best Track Archive for Climate Stewardship, or IBTrACS. This archive covers nearly 5,000 observed tropical cyclones over 45 years and records each storm's position, maximum sustained wind, minimum pressure and wind radii. These measurements describe an individual cyclone rather than the average weather field around a grid cell. The paper puts the imbalance between the two data sources in bytes: roughly 20 terabytes of gridded atmospheric analysis against roughly 10 megabytes of tabular storm records.

The crucial distinction is between learning from those historical observations and receiving the future answer at inference. The paper states that the cyclone records are not model inputs. The model receives atmospheric analysis states and learns to predict both subsequent atmospheric states and cyclone properties. The storm records shape that mapping only during training.

Turn sparse observations into targets the network can learn

A table holding a storm's position and intensity does not line up with a dense latitude-longitude weather grid. The authors bridge the two representations by placing the cyclone targets onto the grid, a step they call griddification.

For location, they place a Gaussian peak with a 120-kilometre scale around each observed cyclone centre. This gives the model a graded signal about where the storm should be, including nearby positions, instead of withholding feedback until a prediction hits one exact coordinate. The resulting map of likely storm centres, the existence field, is trained across the whole globe, because the model's prediction of it is used everywhere.

For properties such as maximum sustained wind, they copy the storm's scalar value into every grid cell within a 120-kilometre disc around that centre. The model's error on that property is counted only inside such discs. Elsewhere, that part of the loss is masked. The network therefore learns a storm's intensity where a storm exists, rather than treating every cyclone-free grid cell as another intensity example.

A copied intensity value is not a claim that the physical wind is constant throughout the disc. It is a training representation for a storm-level property, and that reading is ours rather than the paper's wording. Confusing the two would bring back the original problem: trying to recover a peak wind from a coarse physical wind field alone.

After the network predicts the existence and property fields, a separate heuristic tracker converts them back into tabular trajectories. The paper's methods describe advancing each active centre by momentum, refining it against the predicted existence field, ending tracks whose existence probability has faded, detecting new storms, and reading scalar properties at each centre by interpolation. Several tracker parameters were tuned on the 2022 validation set. The authors also checked that the tracker recovers tracks from ground-truth gridded records to within a few tens of kilometres in most cases. The tracker remains an algorithm outside the network; the system has not eliminated the work of identifying and following a storm.

The authors report that joint training improves both atmospheric and cyclone predictions. Their explanation is that a shared model learns from the abundant atmospheric data while receiving direct supervision for relatively sparse storm events. In their ablation, co-training on cyclone data improved track forecasts by around 6 hours at long lead times, and they argue a two-step design that predicts the atmosphere first and derives storms afterwards would lack that transfer. The paper's own interpretation is that coarse analyses contain more intensity signal than previously recognized. The blog post adds that fully understanding why remains an open research question. Neither source claims that arbitrary fine-scale detail can be reconstructed from any coarse input, and this article does not either.

Predict a set of storms, not one reassuring average

WeatherNext Cyclones generates ensembles of possible futures. Each member carries a complete trajectory with a location, intensity and wind structure, so the system can estimate how often a location encounters winds above a threshold within a forecast window.

The model varies its predictions through what the authors call functional generative networks. A low-dimensional noise vector is injected into the network's normalization layers, so each ensemble member runs a slightly different function rather than adding noise to pixels. Four independently trained models are combined to cover uncertainty about the model itself. Each forecast step needs one network call, which the paper reports makes sampling eight times faster than its diffusion-based predecessor GenCast; one 15-day forecast runs in under a minute on a single TPU. Training optimizes a proper probabilistic scoring rule rather than the distance between one predicted value and one observed value, so the model is pushed to fit a distribution, not only its mean.

That distribution changes the decision a forecast can support. A mean storm track does not say whether all members follow a similar route or whether two groups pass on opposite sides of an island. A moderate average intensity can conceal a subset of severe outcomes. A probability forecast keeps those possibilities visible. The paper reports that the ensemble's track spread is well calibrated, and that its intensity spread is slightly too narrow but much closer to calibrated than the baselines it compares against.

The paper compares ensembles of 50 and 1,000 members using relative economic value, a decision model in which a hypothetical user pays a preparation cost to avoid a larger loss. The larger ensemble raised that value particularly at low cost-to-loss ratios and at longer lead times. At a cost-to-loss ratio of 0.001, 1,000 members at seven days delivered more value than 50 members at five days. This is a model-based evaluation on the 2023–2024 test cases, not money saved by any organization using the forecasts. More members improve sampling of the model's distribution; they cannot correct a distribution that is systematically wrong.

The authors also identify a limit in the location-level wind probabilities. These are derived from approximate storm wind radii. Real winds need not decrease smoothly with distance from the centre, so the probabilities are not a detailed prediction of every local gust. The paper leaves more precise per-location winds, along with precipitation and storm surge, to future work.

Read the lead-time gain together with the operating conditions

The paper reports an average lead-time advantage of a day or more for track, intensity and wind-radius predictions relative to leading operational models. That means comparable accuracy at a longer forecast horizon in the evaluated comparisons. On the 2023–2024 test set, the ensemble-mean track error at five days was 230 kilometres against 370 kilometres for the ECMWF ensemble, which the paper translates into just over 30 hours of added lead time at that accuracy. Three-day intensity error was 3.75 knots lower than HAFS. These are averages over paired, homogeneous evaluations, in which every model is scored on the same storms at the same lead times. They are not a guarantee that every storm receives an extra day of warning.

The evaluation distinguishes global cyclone tests for 2023–2024 from North Atlantic and East Pacific tests for 2025, when a version of the model ran operationally. Design choices were frozen after validation on 2022, and each test-year model was trained only through the preceding year. Different baselines served different tasks: the global ECMWF ensemble for track and probabilistic comparisons, and NOAA's regional HAFS model for deterministic comparisons in the basins it covers. The result is more informative when those comparison boundaries stay visible.

Inference speed is only part of delivery time. The model starts from atmospheric analyses produced by an upstream data-assimilation process. The paper reports that its forecasts arrive around 6.5 hours after the analysis time. Its paired operational evaluations therefore use a forecast initialized six hours earlier and corrected with the latest real-time cyclone observations, a standard practice in operational verification. A fast neural network does not remove the delay before its input becomes available.

The strongest operational interpretation is as additional guidance. The paper evaluates adding WeatherNext Cyclones to the National Hurricane Center's track and intensity consensus ensembles with weights tuned on 2022 data. Track consensus error fell by 18 to 38 percent depending on lead time, and intensity consensus error by 5 to 14 percent. The authors attribute the gain to the model's skill combined with low error correlation with physics-based models, meaning it tends to fail differently. On intensity alone, the existing consensus beats the model at longer lead times, and the authors read their intensity results as evidence that conventional numerical models still carry substantial value. Those findings support combining complementary forecasts rather than assuming one model makes the others unnecessary. Official forecasts and warnings remain the responsibility of meteorological agencies.

The design offers a concrete lesson beyond the performance headline. When a model's input representation smooths away a target of interest, inspect the supervision before assuming the only remedy is a finer input. Direct, well-defined event labels can teach a shared model to predict that target separately, provided the input still carries signal for it and the labels can be anchored to a location the model can find. The evidence must then test the prediction at the level people will use: storm strength, location, uncertainty and timely delivery, rather than the visual detail of the weather grid. The paper suggests the same recipe could extend to heat extremes and extreme precipitation. Whether those phenomena have a comparably clean, expert-curated label archive and a comparably reliable anchor for griddification is the condition that would decide it.