135 papers in the archive · tier-1 watchlist.
Reviews MJO predictability limits and sources, arguing dynamical, statistical and machine-learning forecasts should be combined into a unified subseasonal prediction framework.
Sub-seasonal to seasonalECMWFFudan University (FuXi)NSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Washington
A source-conditioned adapter maps operational GFS, HRES and CMA initial states to ERA5-like states, cutting FuXi 6-240h RMSE by a median 10.2% without retraining.
Joint-Gaussian post-processing of dynamically downscaled FuXi-ENS forecasts beats quantile mapping, extending useful temperature skill from one to two weeks over South Korea.
A learned CNN combines Pangu, GraphCast, FuXi and IFS forecasts with state-dependent weights, beating the best baseline on most variables out to 10 days.
WeatherDiagFlow adds motion, heavy-echo risk and uncertainty heads to radar nowcasting, plus a multi-agent bulletin and audit layer, on FJRADAR data.
FuXi-RTM embeds a differentiable ML radiative transfer surrogate in a global forecast model, improving shortwave radiation and cloud forecasts against stations and satellite products.
A CNN trained on ENSO-teleconnection-selected CMIP6 models extends skilful Northern Tropical Atlantic SST prediction to 8 lead months, versus 7 for dynamical ensembles.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Fudan University (FuXi)Met Office
FourCastNet-based hybrid deep learning downscales 1901–2023 CRU fields to 1/30° humidity, temperature and pressure over the Tibetan Plateau, validated against in-situ sites.
MAST-GAN fuses environmental fields and storm attributes with parallel attention to improve tropical cyclone wind field forecasts, especially for extreme typhoons.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A Fudan University perspective reviews AI for atmospheric and oceanic forecasting and research, with no abstract available to assess details.
Controlled Burgers-system experiments with 600 regimes show AI generalization fails when training data miss physical regimes, implying reanalysis-trained weather AI may struggle with extremes.
Explainability, uncertainty & physicsChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
A ResNet predictor with MLP corrector forecasts Yangtze maximum temperature with RMSE similar to ECMWF, and gives heatwave onset warnings 9 days earlier.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
FuXi-ENS gives smaller TC track errors than ECMWF-ENS for 90 storms in 2018 via better steering flow, but remains underdispersive.
Green's function ensemble with Pangu-Weather reproduces the optimal forcing of the western North Pacific anticyclone while revealing systematic biases.
Explainability, uncertainty & physicsFudan University (FuXi)
3DMergePrec, a 0.25° daily Tibetan Plateau precipitation dataset fused from 12 products with graph attention and 3D CNNs, cuts MSE by 30–40%.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)
A conditional generative emulator trained on CMIP6 projects temperature, precipitation and sea-surface height under arbitrary CO2 pathways, generalising to held-out SSP2-4.5 and new scenarios.
Climate emulationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence LaboratoryUniversity of Chicago
Introduces conditional nonlinear optimal perturbation as an XAI method that finds sensitive regions in AI typhoon and ENSO forecasts more faithfully than gradient saliency.
Explainability, uncertainty & physicsFudan University (FuXi)
Interpretable 3D-residual GRU model detects tropical cloud clusters that will undergo genesis, reaching Heidke skill 0.80 and 94% detection; SHAP shows dynamics dominate.
Extremes & hazardsChina Meteorological Administration (CMA)Colorado State University (incl. CIRA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
HITS is a regional AI typhoon model constrained by AIFS that cuts 120 h intensity errors by up to 48.3% versus AIFS, using a new 9 km typhoon reanalysis.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
Earth-o1 is a grid-free world model that learns atmospheric dynamics directly from ungridded observations, reaching surface forecast skill comparable to ECMWF IFS in hindcasts.
Global medium-range NWPChina Meteorological Administration (CMA)Columbia University (LEAP, Lamont-Doherty)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesShanghai Artificial Intelligence Laboratory
GAN-based post-processing of ECMWF S2S hindcasts over Southern China extends skillful probabilistic precipitation forecasts to week 3, improving CRPS and Brier score.
Post-processing & calibrationChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesTsinghua University (NowcastNet)
Review of AI weather prediction status and pathways toward multi-scale Earth system prediction; abstract unavailable, so classification rests on the title alone.
Global medium-range NWPChina Meteorological Administration (CMA)Fudan University (FuXi)Shanghai Artificial Intelligence Laboratory
Adding radiation fields and land fraction to a neural moist-physics emulator in CAM5 reduces high-latitude biases, but tropical land precipitation bias persists.
Wide and Deep LSTM and ConvGRU network with coordinate attention predicts typhoon tracks using CMA best-track data and ERA5 fields for 2020–2024.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A lightweight 2D deep learning lightning scheme using only CAPE raises global lightning R² to 0.89, a 24% gain over prior ML schemes, and better captures extremes.
Hybrid physics-ML & parameterizationChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesTsinghua University (NowcastNet)
Diffusion residual emulator downscales AIFS to 9-km typhoon fields, cutting 72-hour typhoon intensity error by about 39.5% relative to AIFS.
Downscaling & super-resolutionChina Meteorological Administration (CMA)Fudan University (FuXi)
Physics-informed VI-CNOP initial perturbations applied to FuXi produce typhoon track ensembles with better spread-skill relationships and reveal the role of subtropical high uncertainty.
Probabilistic & ensemble forecastingChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
Uses the FuXi AI model to compute low-cost 5-day tropical cyclone CNOP perturbations, showing AI and NWP models share rapid, physically consistent error growth.
Explainability, uncertainty & physicsFudan University (FuXi)
Flow-matching generative model reconstructs 3-km 3D tropical cyclone structure from FuXi forecasts, cutting maximum wind RMSE 36.5% versus FuXi.
Extremes & hazardsFudan University (FuXi)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
FuXi-Energy produces 1-hourly global ML forecasts of renewable-energy variables and beats ECMWF HRES for wind, solar and tropical cyclone tracks.
FuXi-CFD generates 30 m 3D near-surface wind fields from coarse inputs and terrain, matching CFD accuracy while cutting inference from hours to seconds.
Downscaling & super-resolutionChina Meteorological Administration (CMA)Fudan University (FuXi)
Orthogonal nonlinear optimal perturbations give an ensemble for FuXi tropical cyclone track forecasts that outperforms the operational IFS ensemble.
Probabilistic & ensemble forecastingChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
A scale-adaptive deep neural network parameterization of cloud vertical overlap, usable in climate models, keeps accuracy above 0.9 across resolutions and beats traditional schemes.
Hybrid physics-ML & parameterizationChina Meteorological Administration (CMA)Fudan University (FuXi)
RePPIC-Net fuses FuXi atmospheric fields with geostationary satellite data for surface precipitation phase, beating 4-hour-latency operational snowfall scores (CSI 0.157 vs 0.100).
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)
Lightweight KAN-based network estimates tropical cyclone maximum wind from FY-4 satellite data, cutting parameters 94.8% and running on edge hardware.
SATcast cascaded diffusion model forecasts satellite cloud fields up to 24 hours using FuXi-predicted fields as conditioning, outperforming conventional methods on fractions skill score.
Review of AI for atmosphere–ocean science covering forecasting, data assimilation, downscaling, explainable AI, hybrid physics-AI models, and society-facing applications.
Global medium-range NWPAllen Institute for AI (Ai2, ACE)China Meteorological Administration (CMA)Columbia University (LEAP, Lamont-Doherty)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesNanjing University of Information Science & Technology (NUIST)Potsdam Institute for Climate Impact Research (PIK)Shanghai Artificial Intelligence LaboratoryTsinghua University (NowcastNet)Technical University of Munich (TUM)
DecGAN, a GAN for bias correction and downscaling of CESM output, cuts 12-month Pacific SST error by 77.5% and Asian precipitation error by 18.8%.
YanTian, a Transformer with Earth-geometry-aware shifted attention and relay autoregressive fine-tuning, beats ECMWF HRES at 1-degree resolution and extends Z500 skill to 10.3 days.
Global medium-range NWPFudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesNanjing University of Information Science & Technology (NUIST)
RTCC-Net fuses convective core and polar-coordinate maps from infrared imagery with ResNet and self-attention to classify tropical cyclogenesis 24 h ahead, with 98.7% detection rate.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Integrates causal inference into deep learning to predict Niño3.4 up to 22 months ahead and identifies lead-time-dependent causal pathways for ENSO.
Deep learning combined with dynamical detection for compound drought and heatwave events across historical and future scenarios; abstract missing, so classification relies on the title.
An online interactive self-adaptive network that embeds domain knowledge predicts tropical cyclone intensity; the abstract is unavailable, so results are unknown.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
A lightweight GNSS-PWV multimodal model gives probabilistic 1-hour station precipitation nowcasts across 100+ Chinese stations, with AUC above 91% in the best settings.
Hybrid SHTM–FuXi typhoon system cuts 72-h track error by 16.5% versus FuXi alone and improves 2024 land precipitation threat scores.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
A probability-matching loss replaces MSE for ML precipitation nowcasting, reducing smoothing and heavy-rain underestimation and better matching the observed intensity distribution.
NowcastingChina Meteorological Administration (CMA)Fudan University (FuXi)
Deep learning bias correction of tropical cyclone track and intensity forecasts for the TianXing large weather model, aimed at operational use.
Post-processing & calibrationChina Meteorological Administration (CMA)Fudan University (FuXi)
Environment-conditioned deep learning using 3D atmospheric forecasts and radar beats operational NWP for East China reflectivity and precipitation up to 12 h ahead.
Neural cloud and convection parameterization runs stably for a decade in a real-geography GCM under +4 K warming, without warm-climate training data.
Machine learning corrects mesoscale model estimates to improve offshore wind resource assessment over the Yellow Sea. Abstract unavailable; assessment based on title.
Applications & impactsChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Tsinghua University (NowcastNet)
An LLM agent framework automates extreme-weather diagnostic analysis from raw meteorological data, with a 103-event benchmark and a step-wise evaluation metric.
LLMs & agentsFudan University (FuXi)Shanghai Artificial Intelligence Laboratory
Couples the Dayu radiative transfer model with WRF to simulate Himawari brightness temperatures for Typhoon Hato, finding WRF underestimates low clouds, with an ML cloud classifier used for evaluation.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Assesses CFS seasonal prediction of Western North Pacific DGPI and reports improvement from deep learning correction; abstract unavailable.
Sub-seasonal to seasonalFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Adding atmospheric stability to a neural cloud fraction scheme cuts RMSE by 6.7–15.5% for cloud types and improves stratocumulus simulation in the FGOALS-f3-L GCM.
Optimizes initial conditions of the FuXi subseasonal model to improve prediction of the 2022 Pakistan floods and identifies moisture and circulation processes behind the gain.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)University of Reading
Machine learning fuses ERA5 with multi-source data to map high-resolution urban climate in Shenzhen; abstract unavailable, inferred from title.
Downscaling & super-resolutionChina Meteorological Administration (CMA)Fudan University (FuXi)Tsinghua University (NowcastNet)
DAWP assimilates irregular satellite observation tokens with a masked autoencoder and forecasts in observation space, removing reliance on reanalysis for global weather prediction.
Data assimilationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Random forest feature selection feeds an LSTM that forecasts drought (PDSI) in the Yellow River basin, reaching R2 0.766 and projecting worsening drought under SSP5-8.5.
Compares the Shanghai Typhoon Model with leading ML weather models on Typhoon Danas, assessing relative track and intensity skill for a single case.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
A hybrid FuXi machine-learning and physical SHTM typhoon model with CNOP-targeted satellite assimilation improves track, intensity and rainfall for two 2024 super typhoons.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
Tropical heating experiments show Pangu-Weather reproduces the Matsuno-Gill response and Amundsen Sea Low changes, but overweights Pacific heating versus GCMs.
Explainability, uncertainty & physicsFudan University (FuXi)
KG2data pairs a knowledge graph with a ReAct LLM agent for meteorological API calls, reducing hallucination failures to 0% versus RAG and chat baselines.
Fuses infrared and water-vapor satellite imagery with convective features in a deep model (TCISP-fusion) for 24-hour western North Pacific tropical cyclone intensity, RMSE 10.87 kt.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Neural-network cloud-top heights from Himawari-8 combined with optical flow yield two-layer cloud motion vectors that resolve Typhoon Mulan's wind structure, matching dropsondes to 5.2 m/s.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)
Uses a ResUnet retrieval applied to GOES infrared data to reveal diurnal cycles of cloud properties in North Atlantic tropical cyclones, 2019–2023.
OmniEarth-Bench is a 29,855-annotation multimodal benchmark spanning six Earth spheres including atmosphere; none of 9 MLLMs exceeds 35% accuracy.
Benchmarks, datasets & evaluationFudan University (FuXi)Shanghai Artificial Intelligence Laboratory
Align-DA frames atmospheric data assimilation as score-based generation and aligns the background prior with reward signals for accuracy, forecast skill, and physical consistency.
Data assimilationColumbia University (LEAP, Lamont-Doherty)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A BP neural network retrieves total precipitable water from FY-3D MWHS-II over sea and land, beating seven baseline methods with RMSE of 3.5-4.0 mm against radiosondes.
FuXi-CFD downscales 10 m coarse winds to 30 m 3D wind fields over complex terrain, matching CFD accuracy while cutting inference time from hours to seconds.
Deep learning model for East Asian atmospheric circulation at S2S timescales that explicitly includes multi-sphere (ocean and land) coupling to improve forecast skill.
A random forest generates pseudo soundings for Tibetan Plateau initial conditions, reducing hourly precipitation simulation bias by 31.8%.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
Hybrid FuXi-SHTM typhoon model cuts 2024 western Pacific intensity errors by 60% at 72 h versus FuXi and improves track errors over ECMWF HRES.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
Adds a satellite-derived inner-core symmetry index and an ML ensemble for tropical cyclone rapid intensification, giving 24-h POD of 0.24 to 0.41 in real-time tests.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Physics-constrained perturbations from an AI model's own dynamics enable 2000-member tropical cyclone track ensembles that outperform ECMWF on track skill.
End-to-end neural radiative transfer observation operator for FY-4B GIIRS radiances, trained with maximum likelihood to handle label uncertainty; matches RTTOV Jacobians.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
SwinIR super-resolution downscales and bias-corrects FuXi-2.0 hourly precipitation to 1 km over China, outperforming ECMWF high-resolution forecasts in accuracy and cost.
Nowcastformer fuses radar with upper-air variables and satellite pretraining, and integrated gradients show atmospheric predictors gain importance at longer leads.
NowcastingChina Meteorological Administration (CMA)Fudan University (FuXi)
Satellite-guided conditional diffusion model downscales ERA5 fields to 6.25 km at arbitrary resolution, using GridSat and station guidance to better match observations.
Downscaling & super-resolutionFudan University (FuXi)Shanghai Artificial Intelligence Laboratory
Non-crossing quantile smoothing splines emulate humidity-temperature relations at 297 Chinese stations, projecting humid-heat extremes under 2.0 and 4.5 °C warming.
A multi-task TabNet corrects ECMWF-IFS offshore wind forecasts at 14 stations, cutting wind speed RMSE from 2.20 to 1.25 m/s.
Proposes a physics-guided learning approach for urban surface-atmosphere interactions to model urban climate dynamics; no abstract was provided, so this relies on the title.
Hybrid physics-ML & parameterizationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A Res34-Unet post-processing of NMME six-month forecasts of eastern China summer rainfall identifies WPSH and SST as key drivers and shows which models capture them.
Post-processing & calibrationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Assesses predictability of weather and climate events in AI forecasting models, examining how well data-driven models capture such events.
Explainability, uncertainty & physicsFudan University (FuXi)
Multi-task learning fuses satellite imagery with deviation-angle variance for short-term rolling prediction of tropical cyclone intensity.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Applies a Vision Transformer to satellite images to estimate tropical cyclone intensity directly.
DuoCast splits precipitation nowcasting into low- and high-frequency diffusion branches in latent subspaces, outperforming single-branch U-Net diffusion baselines on four radar benchmarks.
DaYu, a transformer for geostationary thermal infrared imagery at 0.05° and 0.5 h, forecasts cloud images with correlation above 0.8 at 6 hours.
Deep Interpolation raises temporal resolution of historical in-situ meteorological observations in China, achieving 0.934 mean correlation and beating linear interpolation and ERA5-Land.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
FuXi-En4DVar is an ensemble 4DVar system built on the FuXi ML forecast model that produces physically balanced analyses using automatic differentiation gradients.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
Couples deep learning with data assimilation to estimate model error in geophysical forecasting; the abstract was unavailable, so this rests on the title.
TC-GEN uses an ML global weather model and an RNN bias correction to generate synthetic tropical cyclones with realistic track density, landfall and wind extent.
Extremes & hazardsFudan University (FuXi)Google ResearchMassachusetts Institute of Technology (MIT)Princeton UniversityStanford University
A deep-learning ENSO model uses conditional nonlinear optimal perturbations to identify fastest-growing initial errors for two El Niño types, validated in GFDL CM2p1.
FuXi-2.0 delivers 1-hourly global forecasts with more variables, beating ECMWF HRES for wind power and adding an atmosphere-ocean coupled component.
Transformer ENSO-PhyNet embeds heat-budget processes in self-attention, reaching Niño3.4 skill up to 22-month lead with attention maps showing physically consistent drivers.
Neural networks retrieve atmospheric temperatures from FY-3E MWTS-3 brightness temperatures with lower errors than linear regression and recover typhoon warm cores.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)
Pangu-Weather fields spectrally nudge WRF for Typhoon Doksuri, improving track and intensity over ECMWF IFS, and assimilating FengYun-4B water vapour channels reduces intensity error.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
Random Forest and Bi-LSTM bias correction of hourly IMERG-E precipitation, with WRF covariates raising CC from 0.43 to 0.60.
FuXi Weather is an ML system that assimilates multi-satellite observations and produces 10-day global forecasts, outperforming ECMWF HRES in observation-sparse regions like central Africa.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Tsinghua University (NowcastNet)University of California, Los Angeles (UCLA)
Physics-informed deep learning scheme predicts western North Pacific tropical cyclone intensity; abstract unavailable, so details are limited.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
A random forest surrogate of shortwave cloud radiative effect attributes GCM bias to cloud fraction, liquid and ice water paths, and diurnal concurrence in FGOALS-f3-L.
Climate analysis & attributionFudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
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.
Probabilistic & ensemble forecastingChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A Bi-LSTM replaces the multi-scale Kain-Fritsch convection scheme in WRF at 5 km over South China and runs stably when coupled.
CloudDiff uses a diffusion model to retrieve Himawari-8 cloud properties at 1 km from 2 km thermal infrared data, day and night, with ensemble-based uncertainty.
Observations & remote sensingChina Meteorological Administration (CMA)Fudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
FuXi-DA is a deep learning data assimilation framework that assimilates FY-4B AGRI satellite radiances, reducing analysis errors and improving forecasts.
Data assimilationChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
STAC-Pred dual-branch attention CNN improves western North Pacific tropical cyclone 3- and 6-h intensity forecasts over official baselines by 47.69% and 28.26%.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Commentary on the Pangu-Weather AI global forecasting breakthrough, placing it among China's top scientific advances of 2023.
Reviews a decade of quantitative causality analysis from atmospheric predictability and its extension to causal deep learning, with climate applications such as attribution of global warming and El Niño Modoki prediction.
Climate analysis & attributionChina Meteorological Administration (CMA)Fudan University (FuXi)
Framework uses incremental learning to replace sub-modules of an Earth system model with AI components without retraining from scratch.
Proposes three rules for large AI weather models, argues for integrating AI with numerical models, and demonstrates a global ocean wave forecast model.
Global medium-range NWPFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence Laboratory
Res34-Unet post-processes North American Multi-Model Ensemble seasonal precipitation forecasts, improving skill by up to 45% over raw global forecasts.
Post-processing & calibrationFudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
TCIP-Net predicts tropical cyclone intensity from infrared satellite imagery using radial structure evolution, with interpretable contributions and without reanalysis inputs.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Uses UNet++ with Himawari-9 satellite data to reconstruct radar reflectivity for offshore weather monitoring; abstract unavailable, so classification relies on the title.
Observations & remote sensingFudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of SciencesNanjing University of Information Science & Technology (NUIST)
Adas is an AI data assimilation model that, combined with FengWu, forms an end-to-end AI forecast system with skill comparable to IFS.
Data assimilationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)Shanghai Artificial Intelligence Laboratory
Machine learning S2S model outperforms ECMWF's S2S system for precipitation and OLR up to 42 days, extending MJO prediction skill from 30 to 36 days.
Sub-seasonal to seasonalChina Meteorological Administration (CMA)Fudan University (FuXi)
Replaces the multi-scale Kain-Fritsch convection scheme with a Bi-LSTM emulator coupled into WRF at 5 km over South China, achieving high accuracy for gray-zone heavy rainfall.
A denoising diffusion model restores fine-scale detail in FuXi 5-day forecasts, improving extreme precipitation and wind over HRES, but not TC intensity.
SwinIR and Uformer transformers downscale and bias-correct ECMWF HRES 2 m temperature and 10 m wind to 0.01°, outperforming UNet and RCAN baselines.
Downscaling & super-resolutionAlibaba (DAMO Academy)Fudan University (FuXi)
Explainable CNN on 1200-year simulations attributes North China summer precipitation variability to PDO and AMO sea surface temperature patterns, with 96-98% accuracy.
Climate analysis & attributionFudan University (FuXi)Institute of Atmospheric Physics, Chinese Academy of Sciences
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.
Mutual-information-based reweighting strategy improves precipitation nowcasting on heavy rainfall minorities without hurting majority performance or image quality.
Neural network sub-grid cloud fraction scheme trained on CloudSat data reaches r > 0.9 and reduces high- and low-level cloud biases in GCM simulations.
Hybrid physics-ML & parameterizationFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Mutual-information reweighting and time-superimposing training improve radar precipitation nowcasts for convective scenarios across several state-of-the-art models.
Deep learning reconstructs monthly surface air temperature over the Tibetan Plateau for 1901–2020, filling gaps in the historical station record.
A 3D DBSCAN clustering workflow detects the 2022 Yangtze heatwave and drought from ERA5 and tests how clustering parameters change event tracking.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
SwinVRNN learns VAE-based distribution perturbations within a Swin transformer recurrent model, giving better ensemble accuracy and spread than other perturbation methods on WeatherBench.
Probabilistic & ensemble forecastingAlibaba (DAMO Academy)Fudan University (FuXi)
A deep neural network corrects bias in daily satellite precipitation estimates over tropical regions, using ground observations as the reference.
A 3D CNN from one meteorology frame predicts daily CONUS precipitation, beating operational weather models up to 5-day leads and improving them when combined.
LightNet+ fuses radar, station and NWP data in a deep network for lightning forecasting, beating three established schemes and improving as data sources are added.
Extremes & hazardsChina Meteorological Administration (CMA)Fudan University (FuXi)
An end-to-end spatiotemporal scale-selection model improves ECMWF precipitation bias correction, especially for heavy rain, beating 11 published methods on two benchmarks.
An ordinal distribution autoencoder corrects ECMWF and SMS-WARMS precipitation forecast bias and outperforms classical ML and deep baselines.
MeteCNN classifies 11 weather phenomena in a new 6,877-image dataset, reaching about 92% test accuracy.
Observations & remote sensingFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
ConvLSTM with star-shape bridge and rain-oriented loss improves radar precipitation nowcasting over convolutional LSTM baselines and extrapolation.
Hybrid CNN with classification and BP network estimates tropical cyclone intensity from satellite images with RMSE of 8.91 kt.
Extremes & hazardsFudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
Expert-free denoising autoencoder with ordinal regression corrects ECMWF precipitation forecasts, outperforming classical ML and deep baselines on threat scores.
EEMD-decomposed ConvLSTM forecasts the NAO index and outperforms six benchmarks and NOAA GFS and MRF ensemble-mean forecasts in the reported tests.
