ML models that produce ensembles or full forecast distributions: generative models (diffusion, flow matching), CRPS-trained models, learned perturbations.
Search and filter this topic Plain list on GitHub
AIFS-CRPS trains an ML ensemble model on a CRPS-based loss, outperforming the IFS ensemble for most medium-range variables and competing at subseasonal lead times.
ArchesWeather is a transformer deterministic model, and ArchesWeatherGen a flow-matching ensemble that beats IFS ENS and NeuralGCM on WeatherBench headline variables at academic compute cost.
A 1.1-billion-parameter SFNO ensemble with bred vectors and perturbed parameters passes spectral tests and is calibrated, setting up ML huge ensembles of 1,000+ members.
Lawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
FuXi-ENS, a VAE-based ensemble trained with CRPS and KL loss at 0.25°, beats ECMWF ENS in CRPS on 98.1% of 360 variable and lead time combinations.
China Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Lagged ensembles give a parameter-free probabilistic comparison showing GraphCast and Pangu tie on CRPS, and that multi-step losses hurt probabilistic skill.
Lawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
GenCast, a diffusion ensemble trained on reanalysis, beats ECMWF ENS on 97.4% of 1320 targets, with better tropical cyclone and wind power prediction, in 8 minutes.
SEEDS uses diffusion models to emulate operational weather ensembles, generating thousands of realistic members cheaply with improved reliability and extreme-event probabilities after bias correction.
FourCastNet 3 is a spherical-CNN probabilistic ensemble model that rivals diffusion models, forecasts 8-60x faster, and stays calibrated and spectrally realistic to 60 days.
FGN generates ensembles through learned model perturbations trained on per-location CRPS, achieving state-of-the-art probabilistic skill and skillful ensemble tropical cyclone tracks.
AIFS-CRPS trains an ML ensemble model on a CRPS-based loss, outperforming the IFS ensemble for most medium-range variables and competing at subseasonal lead times.
ArchesWeather is a transformer deterministic model, and ArchesWeatherGen a flow-matching ensemble that beats IFS ENS and NeuralGCM on WeatherBench headline variables at academic compute cost.
Continuous Ensemble Forecasting samples diffusion ensemble trajectories in parallel without autoregressive rollout, achieving competitive global forecasts at arbitrary temporal resolution.
A 1.1-billion-parameter SFNO ensemble with bred vectors and perturbed parameters passes spectral tests and is calibrated, setting up ML huge ensembles of 1,000+ members.
Lawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
FuXi-ENS, a VAE-based ensemble trained with CRPS and KL loss at 0.25°, beats ECMWF ENS in CRPS on 98.1% of 360 variable and lead time combinations.
China Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Lagged ensembles give a parameter-free probabilistic comparison showing GraphCast and Pangu tie on CRPS, and that multi-step losses hurt probabilistic skill.
Lawrence Berkeley National Laboratory / NERSCNVIDIAUniversity of California, Berkeley
Graph-EFM is a hierarchical graph latent-variable model that generates arbitrarily large spatially coherent ensembles in one forward pass per step, matching deterministic error levels.
GenCast, a diffusion ensemble trained on reanalysis, beats ECMWF ENS on 97.4% of 1320 targets, with better tropical cyclone and wind power prediction, in 8 minutes.
SEEDS uses diffusion models to emulate operational weather ensembles, generating thousands of realistic members cheaply with improved reliability and extreme-event probabilities after bias correction.
Score-based diffusion models super-resolve coarse NWP to satellite-resolution day-ahead solar irradiance, giving probabilistic forecasts with many samples.
Explores replacing precipitation ensemble prediction systems with machine learning; no abstract is available, so the approach and results are unclear.
Sinh-arcsinh-normal output distributions give neural networks asymmetric, heteroscedastic uncertainty, demonstrated on tropical cyclone intensity forecast errors.
Deep learning precipitation forecasting that jointly models input, target and model uncertainty via error propagation and weight sampling, giving tighter predictive uncertainty than prior joint approaches.
SwinVRNN learns VAE-based distribution perturbations within a Swin transformer recurrent model, giving better ensemble accuracy and spread than other perturbation methods on WeatherBench.
Conditional GANs predict ensemble spread of 500 hPa geopotential from a single deterministic forecast, avoiding costly operational ensemble runs.
Chains dimensional reduction with normalizing flows to produce conditional probabilistic forecasts of surface wind from 22 years of WRF simulation data.
Generative networks trained with prequential scoring rules, avoiding adversarial training, improve probabilistic calibration on chaotic models and a global weather benchmark.
Neural networks predict full probability density functions for WeatherBench geopotential and temperature at 3 and 5 days, stacked into a combined forecast that beats some NWP models.
Bayesian LSTMs give reliable probabilistic forecasts of the Lorenz 84 atmospheric analogue, staying near the attractor and beating VAR and persistence.
A conditional convolutional GAN forecasts 500 hPa height, 2 m temperature and precipitation over Europe for 24 h, and MC dropout yields a deep learning ensemble.
Compares four ensemble generation methods for a neural weather model; retraining gives the largest gain, but ensembles remain less skillful than operational NWP.
A conditional VAE replaces the historical archive in analog ensemble wind speed correction for NAM, matching AnEn skill with up to 25x less memory.
Deep 3D U-Net variants estimate weather forecast ensemble spread from a small set of numerical simulations, reducing the compute needed for uncertainty quantification.
A CNN trained on past forecasts assigns a confidence score to medium-range forecasts without running an ensemble; it is cheaper but less skilful than ensembles.
Deep learning framework for quantifying uncertainty in weather forecasts; abstract unavailable, so only the title is available.
Operational AI ensemble WeatherNext Cyclones predicts tropical cyclone track, intensity and size to 15 days, gaining about 1 day of lead time over leading operational models.
Extremes & hazardsColorado State University (incl. CIRA)Google DeepMindGoogle ResearchMet OfficeNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
HRRRCast is a ResNet-based emulator of the HRRR convection-allowing model over CONUS, with DDIM ensembles that beat HRRR at the 20 dBZ reflectivity threshold.
Regional & km-scale modelsColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Diffusion-LAM is a conditional diffusion model that generates probabilistic limited-area forecasts from boundary conditions, demonstrated on the MEPS regional ensemble dataset.
DYffusion combined with a spherical Fourier neural operator emulates a coarse FV3GFS climate model, producing stable 100-year probabilistic ensemble simulations at low cost.
Operational multi-task neural network post-processes HRRR forecasts with a CRPS loss, improving RMSE by 3–25% and CRPS by 31–45% over HRRR.
Conditional GANs generate synthetic ensemble members from HRRR convection-allowing forecasts, and a CNN turns them into severe weather probabilities with up to 20% Brier skill score gain.
Score-based data assimilation with a memory-efficient score network is tested on a two-layer quasi-geostrophic atmosphere-like model, demonstrating scalability to high-dimensional geophysical systems.
Trains a conditional GAN on ERA5 to generate downscaled 1 km precipitation ensembles over UK and US regions, competitive with IFS ensembles at 8–16 h lead.
Multi-fidelity randomized prior networks combine physical parameterizations and storm-resolving runs to extrapolate to +4K warming, yielding more skillful climate projections with uncertainty.
Stochastic, task-independent atmosphere model learned by self-supervised representation learning; skillful for nowcasting, interpolation, model correction and downscaling without task-specific training.
Generative ensemble diffusion produces precipitation scenarios over Central Europe from ERA5 data, outperforming U-Net baselines after a learned post-processing step.
Neural field compactly represents non-linear mutual-information dependencies across a 1000-member weather forecast ensemble, enabling interactive volume-rendered visualization with lower memory and compute.
Analog ensemble variant that applies past analog errors reduces dry bias and improves heavy 3-hourly precipitation reliability over British Columbia, and can forecast record values.
FuXi is a cascaded ML system giving 15-day global forecasts at 0.25 degrees, the first ML model to match the ECMWF ensemble mean at 15 days.
A 20-member SRCNN ensemble downscales 5 km surface winds to 1 km around Tokyo, with the ensemble mean beating the input winds all year.
Downscaling & super-resolutionJapan Meteorological Agency (JMA)
A CNN estimates state-dependent forecast error covariance from one model run for hybrid data assimilation, matching ensemble Kalman filter performance on Lorenz96 with small ensembles.
A latent diffusion model for radar precipitation nowcasting gives more diverse, well-calibrated ensembles than DGMR and PySTEPS, with more accurate predictions.
Diffusion model conditioned on monthly temperature or precipitation generates realistic daily fields matching Earth system model heat waves and dry spells at a fraction of the cost.
Deep-learning downscaling of ACCESS-S1 seasonal rainfall ensembles to 12 km, trained on CRPS, beats quantile mapping for lead times up to 217 days.
Machine learning gives probabilistic tropical cyclone intensity forecasts with intervals, and its point forecasts outperform an operational statistical-dynamical model over 2010–2020.
PoET, a hierarchical transformer, calibrates ensemble members directly and improves global 2m temperature skill by up to 20% over the raw ensemble.
Probabilistic Output Extreme Learning Machine builds multi-model ensembles for Indian summer monsoon rainfall, aiming to avoid overconfidence and Gaussian assumptions. No quantitative skill score given.
Sub-seasonal to seasonalNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Normalizing-flow neural post-processing jointly calibrates ensemble temperature forecasts across locations and lead times, reaching state-of-the-art on the EUPPBench station benchmark.
A U-Net postprocesses 4 km West-WRF precipitation reforecasts into censored shifted gamma distributions, beating analog, regression and meta-Gaussian benchmarks at all lead times.
Neural networks map reanalysis weather to PV, wind and load, generating 42-year synthetic ensembles for Hungary; Dunkelflaute events and a variance-correction step are analysed.
Proposes a probabilistic clear-sky benchmark and neural post-processing methods for NWP-based solar irradiance forecasts, with proper scoring and calibration checks.
Post-processing & calibrationKarlsruhe Institute of Technology (KIT)
Deep-learning latent space replaces hand-tuned similarity weights in analog ensembles for near-surface wind and solar forecasts, improving RMSE in Pennsylvania.
Conditional diffusion model for short-term rainfall nowcasting; the abstract is unavailable so the classification relies on the title.
PTCIF gives probabilistic 6-24 h tropical cyclone intensity forecasts from multimodal data, with prediction intervals comparable to forecast centres on deterministic scores.
PreDiff is a latent diffusion precipitation nowcasting model with a physics-based knowledge alignment step that enforces domain constraints on SEVIR radar forecasts.
Score-based data assimilation learns a generative prior over state trajectories and applies observations only at inference, enabling zero-shot observation scenarios for long dynamical-system trajectories.
A two-stage statistical downscaling method combines optimal-transport debiasing with diffusion upsampling, enabling unpaired 8x and 16x super-resolution on fluid flows relevant to weather and climate.
EasyUQ converts single-valued model output into calibrated predictive distributions via isotonic regression, with a smooth variant and weather forecast examples; the method is general across models.
Post-processing & calibrationETH ZurichKarlsruhe Institute of Technology (KIT)
3D Earth-Specific Transformer trained on ERA5 is the first AI model to beat IFS across variables and lead times from 1 hour to one week.
Adds a stochastic noise generation step to a deep network for ensemble urban radar rainfall nowcasts, matching optical flow skill over a full year at 5-minute resolution.
Generative neural network post-processing of ensemble forecasts captures multivariate dependencies better than copula methods for German station temperature and wind.
CNN post-processing of GEFS re-forecasts beats CSG EMOS and raw ensembles for 1–7 day China precipitation, especially for extremes above 50 mm/day.
Post-processing & calibrationNanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Deep-learning latent similarity metric for analog ensembles outperforms the original AnEn in RMSE, bias and CRPS for wind and solar forecasts in Pennsylvania.
A CNN gives skilful probabilistic MJO forecasts whose uncertainty tracks instantaneous predictability, using predicted standard deviation and Monte Carlo dropout.
LSTM-parameterised stochastic differential equations for probabilistic Indian monsoon rainfall forecasting from 40 years of station data.
Bayesian neural network learns spatiotemporally varying weights to combine climate-model precipitation ensembles, outperforming three ensemble methods and quantifying epistemic uncertainty.
A convolutional VAE pretrained on imperfect quasi-geostrophic simulations and transfer-learned on noisy observations gives stable stochastic long-term rollouts with accurate climatology.
Climate emulationLawrence Berkeley National Laboratory / NERSCNVIDIA
Extends WeatherBench with probabilistic metrics, an IFS ensemble baseline and three probabilistic ML models; Monte Carlo dropout underestimates uncertainty and no ML model matches IFS.
Benchmarks, datasets & evaluationTechnical University of Munich (TUM)
Hybrid analog-ensemble and CNN post-processing of GEFS precipitation raises CRPSS about 10% over quantile-mapped baseline out to 7 days in British Columbia.
Convolutional conditional neural processes provide off-grid multisite temperature and precipitation downscaling over Europe, beating existing methods and improving extreme precipitation representation.
Downscaling & super-resolutionThe Alan Turing InstituteUniversity of Cambridge (Aardvark)
Conditional GAN and dense networks postprocess COSMO-E and IFS cloud cover ensembles over Switzerland, with cGAN producing realistic scenarios without copula or shuffle templates.
GAN generates spatio-temporal weather fields conditioned on extreme-event segmentation masks, tested on surface radiation and zonal wind data.
DGMR, a deep generative model of radar, produces skillful 5-90 minute precipitation nowcasts that Met Office experts ranked first in 88% of cases.
A CNN-based cubed-sphere ensemble of 32 deep-learning weather models gives 2-6 week forecasts that are only modestly worse than the ECMWF S2S ensemble over land.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)University of Washington
Proposes machine learning algorithms for uncertainty quantification in numerical weather prediction models; no abstract is available to confirm details.
Explainability, uncertainty & physicsArgonne National Laboratory
Wasserstein GAN with a CNN generator learns a 64-dimensional latent space of synthetic PLASIM climate states, reproducing leading EOF statistics and geostrophic balance.
Quantile regression forests give probabilistic icing-related production-loss forecasts for Swedish wind parks, improved by physical icing model inputs.
Proposes a neural conditional-density estimator with case-control sampling for tropical cyclone intensity forecasting, outperforming other ML conditional-distribution methods in simulations and Atlantic storm data.
Distilling an analog ensemble into a deep network postprocesses ECMWF stratospheric wind forecasts, cutting latency to near zero and beating ENS wind-speed error by 20% against balloon data.
Combines random feature map forecasting with ensemble Kalman filter data assimilation, learning sequentially from noisy observations to give long forecast skill, reliable ensembles and closure for multiscale chaotic systems.
Operational quantile regression forests and ensemble copula coupling improve station temperature and gridded hourly rainfall forecasts in France, better capturing cold and heat waves than raw ensembles.
Quantile regression forests post-process and combine multiple NWP models across lead times into calibrated probabilistic road surface temperature forecasts.
ML-tuned blending of multiple ensembles gives probabilistic thunderstorm forecasts from 9 hours to 4 days, clearly better than raw ensemble output.
Deep generative models replace stored archives in the analog ensemble method, cutting memory needs by orders of magnitude for probabilistic NWP post-processing at stations.
Trains a GAN as a stochastic sub-grid parameterization for the Lorenz '96 model; some configurations beat a bespoke scheme at weather and climate timescales.
Hybrid physics-ML & parameterizationNSF NCAR / UCARUniversity of Oxford
Neural networks and quantile trees nowcast precipitation growth and decay in Swiss Alps radar archive, with persistence inputs cutting RMSE by 20-30% in mountains.
Describes a parallel HPC system generating probabilistic weather forecasts with the data-driven analog ensemble method, scaling ensemble generation across nodes.
Neural network predicts probability distributions over discretised power bins for wind farm forecasts, matching deterministic MAE with lower CRPS and avoiding costly NWP ensembles.
