This experimental web site shows hourly probabilistic precipitation forecasts from several sources: The Weather Company's GRAF (Global high-Resolution Atmospheric Forecasting) model and the NOAA High-Resolution Rapid Refresh (HRRR) model, each post-processed into probabilities; the Google WeatherNext3 (WN3) ensemble; the older Weather Company WxMix probability of precipitation; and an ECMWF ensemble probability product; and a multi-model blend of HRRR, GRAF, WN3 and ECMWF.
GRAF is based on the National Center for Atmospheric Research (NCAR) Model for Prediction Across Scales (MPAS). It has run operationally at The Weather Company in some form for nearly a decade, and its validation (not shown here) is competitive with the NOAA High-Resolution Rapid Refresh (HRRR) for many variables, including surface precipitation, temperature, and wind. It has been optimized with help from NVIDIA to run efficiently on GPUs. Basic details of the GRAF configuration are included in the table below. The model runs with 50 vertical levels at 4-km grid spacing over the CONUS and Europe, relaxing to 15 km elsewhere.
The hourly precipitation probabilities were derived from an “Attention ResUNet” deep neural network described and validated in Hamill, High-resolution Calibrated Probabilistic Hourly Precipitation from a Deterministic Forecast (arXiv:2608.12685). Later changes to the method (conditioning on season and lead time, extension to 72 hours, and added predictors) are documented in Hamill, Methodological Changes to the Attention ResUNet Hourly Precipitation Postprocessor (arXiv:2609.38609). This form of convolutional neural network uses gridded GRAF deterministic precipitation forecasts and terrain relief information to estimate probability distributions for precipitation at each model grid point and each hour. The probability distributions are a blend of the probability of zero precipitation and a weighted mixture of two Gamma distributions for positive precipitation amount. The Attention ResUNet was trained against MRMS data in the United States. The forecasts are generally reliable and skillful with respect to climatology.
| Model | MPAS 6.3 (using GPUs, with TWCo tuning), 50 levels, ~ 4.8 M grid cells |
| Physics | WSM6 cloud microphysics; nTiedtke scale-aware convective parameterization; YSU planetary boundary layer and gravity wave drag; surface layer via Monin-Obukhov option. RRTMG radiative transfer scheme; NOAH land-surface model; Thompson cloud fraction |
| Initialization | MPAS 7.0 |
| Terrain | Global Multi-resolution Terrain Elevation Data 2010 |
| Albedo | MODIS Land, 30 arcsec grid spacing. |
| Land use | MODIS-derived predominant land type, veg fraction. |
| Ocean state | ECMWF analysis. |
| Atmospheric and land state | Semi-cycled. 3D-Var hourly updates to recently available observations that have accumulated since the most recently available ECMWF initial condition. |
Table: Configuration of the GRAF model.
The 3-km NOAA High-Resolution Rapid Refresh system is more widely known and used. It is a NOAA high-resolution model that is updated hourly, with non-synoptic cycles making forecasts to +18 h lead time and synoptic cycles making forecasts to +48 h lead time. We focus on the CONUS domain here. More extensive documentation on the HRRR model may be found at https://rapidrefresh.noaa.gov/hrrr/.
Hourly probabilistic precipitation forecasts were generated with a similar Attention ResUNet as used for GRAF (arXiv:2608.12685, arXiv:2609.38609). As a time series of HRRR forecasts exists at cloud providers, the network for HRRR was trained with additional features (predictors), including the dot product of water vapor flux and terrain gradient, climatological precipitation, CAPE, and climatologically normalized precipitable water.
The training protocol was somewhat different as well. All lead times were trained together using Feature-wise Linear Modulation (FiLM). Training data were pooled by season: Dec-Jan-Feb, Mar-Apr-May, Jun-Jul-Aug, Sep-Oct-Nov. More specific documentation on the HRRR calibration procedure is in progress. Forecasts are generally reliable and skillful, as with GRAF; results forthcoming.
This viewer also includes probabilities derived from the Google WeatherNext3 model (WN3; https://deepmind.google/science/weathernext/). WN3 is an ensemble of 64 members that provides hourly precipitation out to +90 hours on a 0.1° (about 11 km) grid. The probabilities shown here start as simple ensemble relative frequencies: the fraction of members whose hourly precipitation is at or above each threshold. A terrain-roughness blend, the same one used for GRAF and HRRR, is then applied: over flat terrain the probabilities are spatially smoothed, and over rugged terrain they are left as computed. The mean and median precipitation are plain, unblended statistics of the members. Unlike GRAF and HRRR, the WN3 probabilities are not passed through the Attention ResUNet, so there is no statistical calibration against observations. Both the CONUS and Europe domains are shown.
The older Weather Company multi-model mix process also generates a probability of precipitation (> 0.01 inches, or 0.25 mm/h). It is included here for comparison of old vs. new. It is the only field available for this product. It is supplied on a ~4 km grid and regridded onto the 3-km (CONUS) and 4-km (Europe) grids used elsewhere in this viewer. Forecasts run to +90 hours from the 00, 06, 12 and 18 UTC cycles. The CONUS field has a thin strip of missing data (about 1% of the grid) along the far western edge, where the source data have no coverage.
The viewer also shows probabilities from the ECMWF ensemble (IFS ENS), which has 51 members (a control run and 50 perturbed members) on a 0.1° (about 11 km) grid, with forecasts out to +90 hours from the 00, 06, 12 and 18 UTC runs. As for WN3, the probability of exceeding each hourly precipitation threshold is the fraction of members at or above it, and the ensemble mean and median are shown. ECMWF provides accumulated precipitation, so the hourly amount is the difference between consecutive forecast hours. The data currently cover only the contiguous United States and nearby areas (0–52°N, 130°W–57°W), so there is no ECMWF product for Europe yet. These probabilities are raw: they have not been calibrated or bias-corrected, and ECMWF’s raw ensemble tends to forecast light precipitation too widely. A statistical adjustment against MRMS is planned once enough forecast and observation pairs have been collected.
The multi-model blend combines the freshest forecast each model has for every valid time. A blend is issued every 6 hours (00, 06, 12 and 18 UTC). For the hours after an issue time, it uses HRRR and GRAF from that cycle (lead time k) and the newest available WN3 run, which is older because WN3 arrives later (for example, WN3 from 00 UTC contributes its 13-h forecast to the blend's 1-h forecast at 12 UTC, and the blend's 1-h forecast is valid at 13 UTC). The WN3 run is chosen dynamically: if the expected one is late, the blend waits a few hours and then uses whatever is newest. The blend runs to +78 h, where the oldest model run runs out. Over the contiguous United States, it also includes the newest ECMWF ensemble run, which receives somewhat less weight than WN3 at first because its raw probabilities are expected to be less reliable.
The weights are inversely proportional to each model's Brier score for exceeding 0.25 mm per hour, at the lead time that model is used at, so a model is weighted less where its forecasts are less accurate, as in inverse-variance weighting. The Brier scores come from the verification statistics for roughly the last week, restricted to the forecast cycles for which every model has statistics; they are pooled in proportion to the number of verified points and smoothed so they cannot decrease with lead time, and are refreshed daily. The blend is a linear opinion pool: for each exceedance threshold, the blended probability is the weighted average of the individual models' probabilities. Where a model has no data (HRRR beyond +48 h, GRAF beyond +72 h, or a model outside its domain, such as HRRR over Europe), it is dropped and the remaining weights are rescaled to sum to one. The CONUS blend is on the 3-km HRRR grid and the Europe blend is on the 4-km GRAF-Europe grid, with GRAF and WN3 interpolated onto them. The blended mean precipitation is available; a blended median is not, since it cannot be computed from the individual medians. Because the blend's forecast at a given lead time can use a different mix of model runs than the models' own, its cycles are issue times, not model initialisation times.
On the verification page, two models are compared only on the forecast cycles and lead times for which both have verification statistics, so differences in archive length or forecast range (for example, WN3 and the blend have a shorter archive than GRAF, HRRR ends at 48 h and the blend at 78 h) do not distort the comparison.