Source-linked AI summary
WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
Stephan Rasp, Boris Babenko, Dominic Masters, Andrew El-Kadi, Samier Merchant, Guy Shalev, Ilan Price, Fred Zyda, Remi Lam, Sasha Shysheya, Matthew Willson, Stratis Markou, Shreya Agrawal, Suhani Vora, Mohammed Alewi Hassen, Sunny Mak, Tom R. Andersson, Megan Bela, Akib Uddin, Nofar Peled Levi, Ben Gaiarin, Ferran Alet, Aaron Bell, Peter Battaglia, Alvaro Sanchez-Gonzalez
TL;DR
AI weather models are limited by lower resolution and reliance on analysis data rather than direct observations. WeatherNext 3 ingests low-latency satellite data and sparse observations to produce hourly, 0.1° forecasts, achieving state-of-the-art probabilistic forecasting skill across diverse evaluations.
Problem
AI weather models trained exclusively on analyses inherit analysis biases, stale update cycles, and information loss from indirect observation assimilation.
Method
WeatherNext 3 directly ingests low-latency satellite imagery and trains prediction heads on satellite-derived precipitation, cyclone, and sparse station observations.
Results
WN3 achieves state-of-the-art probabilistic weather-prediction skill, outperforming WN2 and ECMWF’s AIFS ENS v2 on most analysis metrics while providing hourly 0.1° single-level output.
Takeaways & Limitations
Using information-dense observations in model inputs and outputs can alleviate shortcomings of analysis-only training for AI-based weather prediction.
Takeaways & Limitations
Individual WN3 samples exhibit spatial and temporal artifacts, including grid-shaped precipitation patterns and discontinuities across 6-hour boundaries in station predictions.
Abstract
from arXiv · showhide
State-of-the-art AI weather models have shown impressive medium-range forecast skill and computational efficiency, but suffer two key shortcomings: their forecasts have lower spatial and temporal resolution than the best physics-based models and they are exclusively initialized with and trained on analysis data. As a result, they cannot directly make use of observations, and any biases in the analysis are inherited by the forecast. WeatherNext 3 addresses these shortcomings and establishes a new state-of-the-art for probabilistic medium-range forecasting skill. First, WeatherNext 3 generates new forecasts every hour (rather than every 6 hours like traditional global models) by ingesting low-latency geostationary satellite data. Second, WeatherNext 3's temporal and spatial resolution are on par with physics-based global models, with hourly time steps and 0.1 degree resolution for single-level variables, including solar radiation and cloud cover. Third, WeatherNext 3 moves beyond traditional analysis variables by learning to predict satellite-derived precipitation estimates, as well as tropical cyclone and station observations. Modelling sparse station data allows WeatherNext 3 to make 2m temperature and dewpoint predictions at any location and time, conditioned on local geographical features, with substantially lower error than competing global models, even when evaluated against unseen stations. Together, WeatherNext 3's capabilities move operational AI-based weather forecasting beyond emulating the traditionally distinct stages of data assimilation, forecasting and post-processing, which helps to further push the frontier of performance and granularity for global weather prediction.
1. Introduction
WeatherNext 3 directly ingests and predicts raw observations, addressing biases and latency inherited from analysis-only AI weather models. It establishes a new state of the art on analysis and observation benchmarks while providing hourly, high-resolution global forecasts.
- Analysis-only models inherit analysis biases and six-hour production cycles, limiting their ability to use raw observations directly.The cited shortcomings include biases in precipitation and surface temperature and delayed availability of operational analyses.
- WeatherNext 3 directly ingests and predicts raw observations, setting a new state of the art on analysis and observation benchmarks.The model is introduced as extending prior efforts to leverage observations directly in AI weather forecasting.
- WN3 produces a new forecast every hour by ingesting the latest geostationary satellite imagery.Its satellite inputs are available hourly, with the latest frame having operational latency of just under one hour.
- WN3 produces hourly 0.1° output for all single-level variables, including solar radiation and cloud cover unavailable in WeatherNext 2.
- WN3 is trained to predict global hourly precipitation products derived from satellite data and makes geographically informed 2m temperature and dewpoint predictions at any global location.
2. WeatherNext 3
WeatherNext 3 extends probabilistic global forecasting to multimodal, hourly-initialized prediction by ingesting raw satellite and sparse station observations alongside analyses. Its second-order trajectory model supports mixed temporal resolutions, continuous local outputs, and 15-day 64-member ensembles at substantially finer spatial and temporal resolution than WN2.
- Modeling and uncertainty: The model jointly predicts autoregressive and target-only fields, with uncertainty represented through functional perturbation ensembles combining aleatoric noise and epistemic dropout.Autoregressive fields are fed into subsequent steps, whereas target-only fields are predicted without re-ingestion; WN3 uses two seeds plus epistemic dropout.
- Trajectory modeling: WN3 generates 15-day, 64-member ensemble forecasts by modeling weather trajectories as a second-order Markov process over 6-hour outer windows with finer inner time steps.Each factor predicts a 6-hour window from the two preceding windows, while fields with finer temporal support are predicted at their native inner steps.
- Multiple data modalities and resolutions: WN3 handles modalities with different spatial resolutions, temporal cadences, and latencies through separate native-resolution encoders and decoders connected through a shared icosahedral processor mesh.The design avoids requiring all modalities to be regridded to one common input resolution.
- Observational input and hourly initialization: Direct geostationary satellite ingestion enables hourly forecast initialization instead of 6-hourly initialization, improving short-lead skill particularly for fast-evolving precipitation.The satellite data are available more frequently than HRES analysis, enabling more rapid forecast refreshes.
- Continuous output in time and space via latent interpolation: Continuous latent interpolation combines local station metadata with 0.1° latent representations to predict 2m temperature and dewpoint at arbitrary locations and times.The conditioning metadata include exact elevation, land-versus-sea status, and the observation’s offset within the outer 6-hour step.
3. Evaluation methodology
The evaluation identifies task-specific state-of-the-art baselines and independent ground truths, using CRPS as the main metric while adding task-specific measures. It separately evaluates unseen-station performance, precipitation across multiple products, and the operational benefit of hourly forecast updates.
- Evaluation design: Task-specific state-of-the-art baselines and independent ground truths are selected because WN3’s extended capabilities are not covered by a single model or reference.Different forecast resolutions are interpolated for comparison.
- Analysis and weather stations: Analysis predictions are evaluated against HRES-fc0, except AIFS ENS against its control fc0, while unseen METAR and Mesonet stations provide independent station evaluation via bilinear interpolation.The station evaluation also applies to wind speed, and WN3 station-head predictions are queried on a 0.05° grid before interpolation.
- Precipitation: Precipitation is evaluated against IMERG Final, MRMS Pass-2, and rain gauges to capture differing product trade-offs, although IMERG is not independent because it is a WN3 training target.MRMS is subsampled to 0.05° for computational reasons, and rain-gauge predictions are bilinearly interpolated to station locations.
- Metrics: CRPS is the main evaluation metric, with lower values indicating better forecasts; precipitation additionally uses Brier Scores, reliability diagrams, and evaluation capped at 4 mm/6h.Extreme precipitation is left for future work because true positives are too sparse for reliable evaluation.
- Latency-adjusted evaluation: Hourly WN3 forecasts are compared with an otherwise identical model using 6-hourly initializations to quantify the skill benefit of higher update frequency, assuming 7 hours of operational latency.The latency assumption includes 6 hours for raw-input availability and up to 1 hour for model execution and dissemination.
4. Results
WeatherNext 3 achieves state-of-the-art performance across analysis, station, precipitation, cyclone, and operational baseline evaluations, while hourly satellite-driven initialization and higher resolution improve early-lead skill. Its forecasts nevertheless exhibit spatial and temporal artifacts in individual samples despite strong marginal and spatial performance.
- Analysis scorecards: WN3 outperforms WN2 across most upper-level targets and all evaluated surface variables, with roughly 5% upper-level improvement translating to about 6 hours of additional lead time at equal skill.The largest surface-variable gains occur for 2m temperature, while upper-level humidity benefits particularly from geostationary satellite imagery.
- Observational targets: WN3 station-head predictions reduce short-lead 2m-temperature CRPS by up to 30% versus WN2 and 40% versus ENS, with similar relative-humidity gains versus ENS.The 0.1° analysis predictions also reduce early-lead 10m-wind-speed CRPS by 5%, while station-head predictions avoid unnecessary lapse-rate post-processing.
- Precipitation: WN3 precipitation heads reduce early-lead CRPS by up to 60% against IMERG, 30% against MRMS, and 10% against rain gauges, while remaining calibrated through 15 days.Hourly satellite-enabled updates gain 2–3 hours of lead time over latency-adjusted 6-hourly forecasts, and PARDIG has the lowest real-time MRMS CRPS by some margin.
- Distribution and artifacts: WN3 shows relatively small but consistent cyclone-track and intensity improvements, with larger extent gains at 1–3 days, but its ensembles are generally more under-spread than WN2.Spatially pooled precipitation skill shows no systematic degradation at larger pooling sizes, although individual samples contain hexagonal grid artifacts and 6-hour boundary jumps in station-head temperature.
- Operational comparison: WN3 outperforms AIFS ENS across all upper-level variables by roughly 10% during the first forecast week and beats it on most single-level variables, station metrics, and lead times.For real-time precipitation, PARDIG has the lowest MRMS CRPS, while PARDIG and IMERG are roughly on par against rain gauges.
5. Discussion
WeatherNext 3 improves forecast skill and operational capability by achieving state-of-the-art probabilistic medium-range performance, 0.1° single-level resolution, and hourly initialization. It also trains directly on observation data, including satellite-derived precipitation reanalysis and station observations, rather than relying purely on analysis data.
- Forecast skill: WeatherNext 3 sets a new state-of-the-art for probabilistic weather prediction, outperforming WN2 and ECMWF’s AIFS ENS v2 on a vast majority of analysis metrics.This establishes the paper’s main forecast-skill result against both its predecessor and a leading operational AI model.
- Resolution and initialization: WeatherNext 3 produces single-level variables at 0.1° resolution, matching ECMWF ENS, and uses low-latency satellite input to initialize forecasts every hour.These capabilities bring its spatial and temporal output granularity in line with the best traditional global numerical weather prediction system.
- Observation-based training: Direct training on high-quality observations helps WeatherNext 3 circumvent known biases and inaccuracies in analysis data.For precipitation, the model uses PARDIG, a precipitation reanalysis derived from global space-borne radar data, and its forecasts are evaluated against multiple ground truths.
- Beyond analysis data: WeatherNext 3 uses information-dense, low-latency observations rather than relying purely on analysis data, while recognizing that analyses provide dense, uninterrupted training data but represent only a best guess of the atmosphere.ERA5 is identified as an important contributor to the advancement of AI weather models, despite the limitations of analysis-only data.
Disclaimer
WeatherNext is an experimental AI forecasting system whose predictions are informational rather than official warnings; users should defer to emergency authorities and national meteorological services.
- Disclaimer: WeatherNext forecasts are experimental, provided as-is, and not official severe weather warnings; users should rely on local emergency authorities and national meteorological services for life and property protection.Use of WeatherNext and its outputs is subject to the Terms of Service.
- Disclaimer: This document is based on data and products from the European Centre for Medium-Range Weather Forecasts (ECMWF).
Appendix · A.1. Model details · A.1.1. Model formulation
WeatherNext 3 represents atmospheric spatio-temporal fields with a probabilistic operator that encodes multiresolution inputs into a refined mesh and decodes both gridded and continuous predictions. Its end-to-end composite objective supports autoregressive training across heterogeneous, sparse, and irregular observations.
- A.1.1. Model formulation: The atmosphere is represented at each timestep and location by N spatio-temporal physical fields, each defined over its own domain, time subset, and value space.
- A.1.1. Model formulation: A single noise vector parameterizes aleatoric forecast uncertainty through conditional normalization layers, while the operator builds on FGN (Alet et al., 2025).
- A.1.1. Model formulation: The encode–process–decode operator uses two-resolution input grids, pointwise latent encoding, and a GNN that maps each to a 6-times-refined icosahedral mesh with 40962×1024 latent features.
- A.1.1. Model formulation: Gridded fields are decoded pointwise at native latitude–longitude resolutions, whereas station predictions query a 0.1° latent grid continuously.
- A.1.1. Model formulation: Continuous querying combines bilinear interpolation, a 4-layer width-768 CNN, local surface features, and query time, enabling sparse-observation training and arbitrary spatial-temporal inference.
- A.1.1. Model formulation: Training is end-to-end over R autoregressive rollout steps using fair CRPS from two sampled trajectories, with per-modality weights controlling relative importance independently of grid size and sampling density.
- A.1.1. Model formulation: Loss averaging over available station reports and non-missing cyclone targets prevents sparse modalities from being overwhelmed, while input-output variables generally use residual targets relative to preceding frames.
A.1.2. Additional model modifications
Additional modifications improve training calibration, handle cumulative and missing inputs, and support operational robustness and scalable processing. WN3 also applies specific sea-surface-temperature preprocessing and spatial sharding for implementation at scale.
- Global mean loss: A globally pooled CRPS term weighted at 0.3 of the corresponding base-variable loss reduces per-sample ensemble-member bias for precipitation and cloud cover.The modification addresses bias observed when training used only local point-wise marginal CRPS.
- Cumulative variables: Cumulative variables are represented as interval accumulations and de-accumulated from HRES inputs and targets, with HRES-fc0 yielding a constant zero after de-accumulation.The variables are cdir, fdir, ssrd, and tp; the interval is defined from initialization time to the corresponding lead time.
- Input dropout for operational robustness: Input dropout trains WN3 to produce skilful forecasts despite unavailable autoregressive inputs, eliminating separate initialization-time handling.This covers unavailable first-step cumulative inputs, operationally unavailable IMERG and PARDIG, and absent inputs in older training years.
- Sea surface temperature (sst): For sea-surface temperature, missing HRES-fc0-5 land values are imputed with the dataset’s minimum value using ERA5 and HRES land masks, as in Alet et al. (2026).ERA5’s SST validity mask is at 0.25° and HRES’s land-sea mask is at 0.1°.
- Spatial sharding: Spatial sharding partitions icosahedral mesh nodes by longitude by default, making grid-to-mesh and most expensive sharding operations local while retaining global communication for mesh-to-grid scatter.The design reduces cross-partition communication for large-scale WN3 processing.
A.1.3. Training
WeatherNext 3 uses a multi-stage curriculum that progressively increases temporal and spatial resolution while adding observation modalities. Specialized finetuning, normalization, and inference-time clipping support cyclone, station, autoregressive, and bounded-variable prediction.
- Progressive increase of resolution: The curriculum begins with ERA5 or HRES-fc0-5, geostationary satellite imagery, IMERG and PARDIG precipitation, and hourly sub-step analysis targets before adding further modalities.Training progressively increases spatial resolution and introduces additional observation modalities.
- Cyclone finetuning: Cyclone finetuning first trains new output heads with the backbone frozen, then unfreezes the full model for joint optimization.This two-step process introduces cyclone attribute maps while preserving a separate warmup phase.
- Autoregressive and frozen finetuning: Autoregressive finetuning increases rollouts from 2 to 8 steps, while frozen high-resolution finetuning trains station heads and interleaves them with autoregressive stages.The final schedule uses a 75%-of-steps frozen stage after 7 autoregressive steps, an 8-step stage, and a second frozen stage using merged backbone and station weights.
- Output clipping: Clipping bounded variables during inference prevents stationary pixel artifacts caused by rollout values leaving the training distribution.Cloud cover is clipped to [0, 1], while specific humidity, solar radiation, and precipitation are clipped at ≥0 before autoregressive feedback.
A.2. Additional information on evaluation … A.3.8. IBTrACS gridded fields
The appendix specifies evaluation choices and dataset construction for WeatherNext 3, including operationally relevant CRPS variants, baseline data handling, observation preprocessing, and specialized precipitation and cyclone targets. It also documents dataset limitations and implementation details that affect comparisons and model outputs.
- A.2. Additional information on evaluation; A.2.1. CRPS: Biased CRPS is used for WN3 multi-seed ensembles because they violate the i.i.d. assumption, while ENS uses an exception because its full ensemble was unavailable.The biased metric better reflects operational forecast skill; differences from unbiased CRPS are small for ensembles of approximately 50 members.
- A.2.2. Pooled CRPS: Pooled CRPS applies dimensionality-preserving spatial average or max pooling over approximately equal-area latitude–longitude patches before latitude-weighted spatial averaging.A pooled value is computed for a patch centered at every grid point.
- A.3. Additional information on datasets: The appendix provides supplementary dataset descriptions and a variable-abbreviation table, including accumulation-period suffixes for cumulative variables.These materials cover the additional datasets used throughout the evaluation.
- A.3.1. ECMWF ENS and AIFS ENS baselines: ECMWF ENS and AIFS ENS baselines are downloaded from TIGGE, MARS, or ECMWF open data, interpolated to evaluation grids, and adjusted for incomplete ensemble availability.ENS evaluations use fair CRPS because only 48 members, without the control forecast, were downloaded; TIGGE–MARS resolution differences introduce small error spikes at 6-hour lead times.
- A.3.2. Geostationary satellite mosaic: WN3 training uses a real-time geostationary satellite mosaic combining multiple instruments and 11 visible and infrared channels regridded to 0.1°.The historical mosaic extends to 2016 and spans several satellite and instrument generations.
- A.3.3. In-situ surface observations; A.3.4. Elevation and land-sea mask: The hourly in-situ dataset combines METAR, Mesonet, and ICOADS observations from June 2001, with quality filtering, temporal selection, and physical bounds used to remove unreliable records.Station predictions additionally require latitude, longitude, elevation, and land/sea metadata derived from global elevation and land-cover datasets.
- A.3.5. IMERG; A.3.7. MRMS: IMERG Final supplies 0.1° precipitation estimates aggregated from 30-minute data to hourly accumulations, while MRMS provides independently evaluated radar- and gauge-calibrated hourly estimates.MRMS evaluation excludes low-radar-quality areas and restricts the spatial domain.
- A.3.6. PARDIG; A.3.8. IBTrACS gridded fields: PARDIG is a separately trained satellite- and reanalysis-based precipitation estimator that preliminary evaluations find substantially more skillful than IMERG Final against radar or rain gauges; IBTrACS is converted into gridded cyclone targets with modified wind-radius imputation.PARDIG uses a four-hour geostationary-satellite, pressure-level, surface-variable, and microwave-sounder input window, while the cyclone targets use thresholds of 34kt, 50kt, and 64kt.