50 papers in the archive.
A Wasserstein GAN emulates 10 km Himalayan extreme precipitation from ERA5, reducing relative bias by about 35% versus a CNN baseline.
SciExam for ENSO benchmarks LLM agents building stochastic ENSO models; six of twelve agent systems beat the published model.
FlexCast uses one set of parameters to forecast arbitrary subsets of 69 ERA5 fields via metadata-conditioned adapters and a U-Transformer, with lower errors when cross-field context is compatible.
Global medium-range NWPNanjing University of Information Science & Technology (NUIST)Princeton University
S3N uses an L2 projection between latitude-longitude and HEALPix grids and spiral state-space scanning to forecast weather at 4, 7 and 10 days with slower error growth.
Global medium-range NWPNanjing University of Information Science & Technology (NUIST)Princeton University
A 1000-member ACE2 emulator ensemble and physics models show seasonal Atlantic TC counts are underdispersed versus Poisson, favouring a binomial model.
Deep learning estimate of solar energy reduction during hurricanes. Abstract missing, so the classification is low confidence and based on the title.
Applications & impactsPrinceton UniversityUniversity of Oxford
Convolution-based multivariate Gaussian process jointly corrects climate model temperature and precipitation biases, using surface energy flux relationships to keep outputs physically consistent.
Post-processing & calibrationETH ZurichNSF NCAR / UCARPrinceton University
HRadNN uses a neural radiative transfer emulator by default and falls back to physics when energy is not conserved, stable under 9 K SST warming and 4-5x faster.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Graph transformer emulates station storm surge from atmospheric forcing with peak-aware losses, beating a GNN baseline and running in seconds.
GraphCast inherits and amplifies ERA5's premature afternoon peak and weak nocturnal boreal-summer rainfall, while a 5-km Met Office model improves night-time precipitation.
Global medium-range NWPMet OfficePrinceton UniversityUniversity of Reading
Applies a data-driven ZB20 eddy momentum parameterization in an idealized coupled CESM, strengthening ocean heat transport and shifting the Southern Hemisphere mid-latitude jet.
Hybrid physics-ML & parameterizationNew York University (M2LInES)Princeton University
Machine-learning-guided site calibration of Noah-MP at tropical forest and urban sites improves latent and sensible heat flux and soil moisture, but nighttime fluxes remain hard.
Data assimilationInstitute of Atmospheric Physics, Chinese Academy of SciencesNSF NCAR / UCARPrinceton University
Random forest severe-convection forecasts from C-SHiELD convection-allowing ensembles show skill only through day 7, with a modest gain over parameterized-convection ensembles.
Extremes & hazardsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton UniversityUniversity of Oklahoma (AI2ES)
A deep learning model driven by observed SSTs generates 1,000 samples, placing the record 2020 Atlantic hurricane season at roughly a 0.5% event.
ACE2 emulator past-millennium runs link El Nino-like SST patterns to fewer global and Atlantic tropical cyclones via stronger wind shear and drier mid-troposphere.
Climate analysis & attributionColorado State University (incl. CIRA)Princeton University
Global CLM simulations analyzed with CatBoost and SHAP show urban heat island intensity during heatwaves is more positive at night, driven mainly by net longwave radiation.
DeepTCEye detects tropical cyclone eyes in infrared imagery with about 96% accuracy, and an eye persistence index cuts intensity-estimation RMSE for intense storms by up to 15.65%.
CNNs trained on SST and OLR maps predict Atlantic hurricane activity (ACE) categories four months ahead, with higher accuracy than past May and August seasonal forecasts.
Sub-seasonal to seasonalNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Neural ODEs fit 290 ice crystal growth series, and symbolic regression yields an extended growth equation that matches chamber data in early growth stages.
Hybrid physics-ML & parameterizationColumbia University (LEAP, Lamont-Doherty)Karlsruhe Institute of Technology (KIT)Princeton UniversityUniversity of Chicago
Tests ACE2, NeuralGCM and cBottle under uniform SST warming against GFDL AM4: precipitation response is reproduced, but radiative and land warming responses depart from physical behaviour.
SamudrACE is a coupled atmosphere-ocean ML emulator giving stable centuries-long runs at 1 degree with realistic ENSO variability absent in uncoupled mode.
Climate emulationAllen Institute for AI (Ai2, ACE)Columbia University (LEAP, Lamont-Doherty)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)New York University (M2LInES)Princeton University
Self-organizing maps link ENSO-driven large-scale sea level variability to West Coast coastal sea level, improving seasonal dynamical predictions.
Sub-seasonal to seasonalNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Shows two deep learning climate models, a hybrid NGCM and data-driven DLESyM, reproduce 1900-1960 land heat and cold wave frequencies with skill comparable to HiRAM, though DLESyM overestimates them.
Extremes & hazardsPrinceton UniversityUniversity of Washington
Bi-LSTM bias correction reduces the SST forecast error in TCWAGFS by 97% at day 10 (0.38 K to 0.01 K), improving extended-range forecasts tied to MJO and ENSO.
ML sea ice bias correction embedded online in the GFDL SPEAR coupled model cuts Antarctic winter ice extent forecast errors more than twofold at 4-6 month leads.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)New York University (M2LInES)Princeton University
Reinforcement learning derives adaptive seawall strategies for Manhattan under sea-level rise, cutting expected cost by 6–36% versus prior methods under SSP2-4.5.
CNNs with layer-wise relevance propagation predict western US cloud-to-ground lightning days (median AUC 0.8), usable for climate-model lightning parameterization.
Extremes & hazardsColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Neural network classifies temperature and precipitation maps by radiative forcing scenario using SPEAR output, with attribution identifying regional fingerprints of climate change.
Climate analysis & attributionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
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
CNNs trained on CESM2 large ensemble using tropical Pacific SST beat regression and dynamical models for US seasonal temperature at 1–6 month leads.
Ola is an AI coupled ocean-atmosphere model using Spherical Fourier Neural Operators at 0.25° that produces realistic ENSO and offers initial seasonal ENSO skill comparable to SPEAR.
Sub-seasonal to seasonalNVIDIAPrinceton UniversityUniversity of Washington
Perturbation-based ensembles from the Pangu AI model generate thousands of tropical cyclone scenarios, tracking ECMWF ensemble forecasts up to seven days before landfall.
A U-Net predicts flow-dependent localized covariances for the Ensemble Kalman Filter, matching or exceeding 3DVar and EnKF in a quasi-geostrophic model.
Data assimilationNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Explainable neural networks date emergence of anthropogenic summer temperature signals over the US, finding minimum temperature signals emerged by the early 2000s, earliest in the East.
Climate analysis & attributionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
A physics-informed neural network reconstructs full 2D and 3D tropical cyclone wind and pressure fields from sparse simulated data and reconstructs Hurricane Ida from real observations.
Data assimilationCalifornia Institute of Technology (Caltech)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton UniversityStanford University
PCA with quantile regression forests yields probabilistic over-ocean tropical cyclone rainfall maps for the North Atlantic, reaching storm-total correlation of 0.96 with satellite data.
U-Net predicts flow-dependent covariances for an Ensemble Kalman Filter in a quasi-geostrophic model, matching 3DVar and EnKF for ensemble sizes up to 80.
Data assimilationNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
Bayesian deep learning forecasts long-term daily solar irradiance from climate simulation outputs for energy applications.
Nature Climate Change perspective on combining AI and computing for climate modelling and prediction; abstract absent, judged from title and venue.
Hybrid physics-ML & parameterizationCalifornia Institute of Technology (Caltech)Massachusetts Institute of Technology (MIT)NSF NCAR / UCARPrinceton University
Bayesian deep learning downscales daily wind speed for climate monitoring; the abstract is missing, so the scope is inferred from the title.
Supplementary figures only for a machine learning study of tropical cyclone precipitation; the record gives no method or results.
Logistic regression and neural networks show stratospheric aerosol injection signals in regional temperature emerge within a decade, while precipitation signals are harder to detect.
Climate analysis & attributionColorado State University (incl. CIRA)Princeton University
Software accompanying a study that uses neural networks to predict hurricane storm surge and assess its sensitivity to storm characteristics.
An overview of ML downscaling of GCM output, with a deep-learning case study producing local weekly temperature projections for a New Jersey watershed.
Convection-permitting simulations with ML damage models project 128% higher typhoon losses for Southeast China under RCP8.5 warming, with 50-year losses up 20–170%.
Applications & impactsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
A GAN generates realistic cloud fields conditioned on reanalysis, aiming to address cloud-physics uncertainty in climate models; large-scale patterns are good, fine textures are not.
Climate emulationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)NVIDIAPrinceton University
A deep autoencoder detects anomalous heavy precipitation events in Japan and attributes their increase since 1977 partly to anthropogenic forcing, with more expected this century.
Climate analysis & attributionNSF NCAR / UCARNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Nanjing University of Information Science & Technology (NUIST)Princeton University
A random forest learns nudging tendencies and is coupled to FV3GFS, reducing 500 hPa height forecast error and annual-mean precipitation RMSE by about 20% in year-long runs.
Hybrid physics-ML & parameterizationAllen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton UniversityUniversity of Washington
Perspective proposing Neural Earth System Modelling, which fuses deep networks and Earth system models into interpretable hybrids that might eventually replace them.
Hybrid physics-ML & parameterizationColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Potsdam Institute for Climate Impact Research (PIK)Princeton UniversityUniversity of Washington
Prospectus arguing that the end of Dennard scaling will push climate science toward combining machine learning pattern recognition with physical equations, reshaping how weather and climate are simulated.
Hybrid physics-ML & parameterizationNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Princeton University
