ML inside physical models: subgrid parameterizations (convection, clouds, radiation, turbulence, gravity waves), online coupling, differentiable dynamical cores.
Search and filter this topic Plain list on GitHub
A NeuralGCM-based hybrid model trained on satellite precipitation cuts climate-model precipitation biases and beats the ECMWF ensemble at mid-range forecasts.
A U-Net emulator of embedded cloud-resolving models gives a stable 5-year hybrid climate run with realistic cloud condensate and a 0.96 mm/day precipitation RMSE.
NeuralGCM couples a differentiable dynamical solver with ML components, matching ML and ECMWF ensemble skill for 1-15 day forecasts and stable multi-decade climate runs.
ECMWFGoogle DeepMindGoogle ResearchMassachusetts Institute of Technology (MIT)
Machine-learned nudging tendencies correct a 200-km FV3GFS model toward 3-km X-SHiELD, gaining 18 hours of 3-7 day skill and 30% better precipitation.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Washington
A neural network trained on cloud-resolving output replaces subgrid convection in a global climate model, giving stable multi-year runs that capture precipitation extremes and approximately conserve energy.
A blueprint proposes Earth system models that learn parameterizations from observations and high-resolution simulations via data assimilation and machine learning, illustrated on a toy system.
Neural network emulates CAM longwave radiation 50-80 times faster with near-identical 10-year climate simulations, an early ML physics emulation.
A NeuralGCM-based hybrid model trained on satellite precipitation cuts climate-model precipitation biases and beats the ECMWF ensemble at mid-range forecasts.
Review of ML across climate physics, covering observation reconstruction, subgrid parameterization, causal inference and climate and weather prediction.
ECMWFInriaLawrence Berkeley National Laboratory / NERSCUniversity of Chicago
Hybrid NWP-AI system spectrally nudges GEM toward GraphCast large scales, improving prediction skill and tropical cyclone tracks while keeping physically consistent fine scales.
A U-Net emulator of embedded cloud-resolving models gives a stable 5-year hybrid climate run with realistic cloud condensate and a 0.96 mm/day precipitation RMSE.
Review of ML for climate physics and climate simulations, arguing physics knowledge and interpretability are essential in the small-data, non-stationary regime of long-term climate projection.
Massachusetts Institute of Technology (MIT)Stanford UniversityUniversity of Chicago
NeuralGCM couples a differentiable dynamical solver with ML components, matching ML and ECMWF ensemble skill for 1-15 day forecasts and stable multi-decade climate runs.
ECMWFGoogle DeepMindGoogle ResearchMassachusetts Institute of Technology (MIT)
Trains neural subgrid parametrizations with a posteriori losses via a neural emulator that propagates gradients through non-differentiable solvers, tested on Lorenz-96 and quasi-geostrophic systems.
An adversarially trained deep learning parameterization of subgrid vertical wind variability reproduces observed statistics and can run online in large-scale atmospheric models.
Neural network emulates the Met Office nonorographic gravity wave scheme using only two of six inputs, and reproduces QBO and ENSO-linked gravity wave variability in a climate model.
Ensemble of 32-layer residual CNNs emulates superparameterized convection and clouds, runs stably for 5 years in CAM5, and generalizes to a +4K warmer climate.
Builds a semi-automated pipeline for O(100)-member ensembles of ML subgrid parameterization runs, showing offline error reductions can have opposite effects on online error and stability.
MEDIDA uses sparse equation discovery with random Fourier features to learn Earth system model errors from sparse analysis increments in a quasi-geostrophic test case.
Multi-fidelity randomized prior networks combine physical parameterizations and storm-resolving runs to extrapolate to +4K warming, yielding more skillful climate projections with uncertainty.
Online re-training of two layers of a gravity-wave CNN via ensemble Kalman inversion yields realistic QBOs in a 1D model when offline small-data training fails.
Nature Climate Change perspective on combining AI and computing for climate modelling and prediction; abstract absent, judged from title and venue.
California Institute of Technology (Caltech)Massachusetts Institute of Technology (MIT)NSF NCAR / UCARPrinceton University
U-Net++ and U-Net3+ networks emulate full shortwave and longwave RRTM radiation in GFSv16, running 7510 and 90 times faster with reliable heating rates.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Hybrid model uses an ANN trained on analysis increments to correct errors in a coupled ocean-atmosphere model (MAOOAM), improving prediction skill across lead times.
A Fortran Torch Adaptor couples deep learning radiation emulators into WRF, achieving 50x to 300x GPU speedups and better forecasts than feedforward networks.
Compares random forest and neural network emulators of ICON's ecRad radiation; neural nets show top-of-atmosphere bias, and a recurrent design with flux normalisation performs best.
CAMNet, a physics-informed neural operator trained on high-resolution simulations, accelerates 2-D gravity wave and secondary wave simulations on GPUs, capturing key dynamics.
A GAN-based convective boundary layer parametrization, trained on DNS with Deardorff self-similarity, reproduces buoyancy and vertical velocity statistics and thermal organization.
Argues machine learning in climate model parameterization still requires expert judgment, so it is no more objective than conventional tuning.
Proposes memory-based neural network closures trained with a differentiable solver for Lorenz '96, giving more stable long-horizon forecasts than instantaneous parameterizations.
Physically constrained neural network with a 2D latent space predicts boundary-layer turbulent fluxes from LES data, outperforming the eddy-diffusivity mass-flux scheme.
Random forest emulators of CAM6 physics skill declines as model complexity rises across dry, moist, and extra-convection configurations, indicating limits for RF physics emulation.
Tests a machine-learning radiation emulator in an NWP model across 0.25–100 km resolutions over Korea, finding larger heating-rate errors at fine scales.
A neural-network precipitation parameterization fails at extremes unless informed by a novel sub-grid organization metric, which raises variance R² from about 0.4 to 0.95.
Argues that neural closures for atmospheric dynamics struggle with rare events and multiscale high-frequency waves, illustrated with the Lorenz 1980 model.
Neural networks trained on Mie-computed data replace E3SM's aerosol optics parameterization; randomly wired deep networks outperform MLPs and the existing scheme.
Neural networks emulate the full shortwave and longwave RRTM radiation scheme for GFSv16, running 6579x and 96x faster than the original with near-zero bias.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Compares three methods to extract 3D subgrid gravity wave fluxes and drag from 3 km tropical simulations, showing lateral fluxes can matter for data-driven parameterizations.
Physics-incorporated network enforces flux-heating rate energy conservation in radiative transfer; bidirectional RNNs and transformers beat FC nets and CNNs.
Nanjing University of Information Science & Technology (NUIST)Alibaba (DAMO Academy)
Causal discovery selects inputs for neural convection and radiation parameterizations that replace superparameterization, giving stable climate simulations matching the high-resolution model.
Builds a deep-learning westerly wind burst parameterization intended to improve ENSO prediction; the abstract is missing, so gains are unknown.
Symbolic regression yields an 11-parameter interpretable cloud cover equation matching neural network skill (R²=0.94), beating Xu-Randall and transferring better to ERA5.
Two-layer CNNs with about 1000 parameters learn subgrid stresses for forced 2D beta-plane turbulence, giving stable online LES at 16x downscaling. Idealized geophysical flow.
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.
Fudan University (FuXi)Nanjing University of Information Science & Technology (NUIST)
A neural-network subgrid momentum parameterization trained on high-resolution aquaplanet output conserves momentum and gives stable coarse-model runs with reduced wind biases.
Boosted trees and random forests map satellite SIF to GPP to calibrate E3SM land photosynthesis parameters, reducing simulated GPP uncertainty after surrogate-based benchmarking.
Oak Ridge National LaboratoryColorado State University (incl. CIRA)
Multiscale networks that pass information fine-to-coarse and coarse-to-fine correct chaotic models, demonstrated on a climate subgrid parameterization task for unresolved fine-scale dynamics.
Supermodels with machine-learned connections synchronize imperfect atmospheric models, reproducing blocking and ITCZ criticality absent from the constituent models.
Machine learning emulates the systematic error of a reduced-order coupled ocean-atmosphere model, and the resulting hybrid model beats the dynamical model alone in twin experiments, best when correcting both components.
A U-Net replaces EOF models for precipitation response to SST anomalies in hybrid ENSO modelling, outperforming EOF over the western-central tropical Pacific.
Nanjing University of Information Science & Technology (NUIST)
A two-step hyperparameter search tuned on a tiny data fraction gives up to 135x speedup, demonstrated on neural emulators of aerosol activation in climate models.
A DeepONet with autoencoder replaces the nudging bias-correction module in the E3SMv2 climate model, matching nudging tendencies with a learned surrogate for online use.
WRF-ML coupler lets Python deep learning parameterizations run inside the Fortran WRF model, demonstrated with trained DL radiation emulators coupled to WRF.
Physics-informed neural network surrogate for shallow water equations cuts solver compute time by over 80% while keeping RMSE near 0.07 on normalized variables.
DeepBL, a small 1D convolutional tropical cyclone boundary-layer parameterization trained on LES, outperforms the WRF YSU scheme in predicting turbulent fluxes.
ClimSim-Online releases a 5.7-billion-sample multiscale dataset plus a containerized workflow for integrating ML emulators into a host climate model for hybrid ML-physics testing.
Deep neural network whitecap-fraction parameterization for Earth system models reduces error by about 15% versus a recent scheme and ranks drivers by permutation importance.
A knowledge-guided LSTM encoding diel hysteresis and Birch effect outperforms Q10 models and CLM in dryland soil respiration (correlation 0.7 vs under 0.4).
A differentiable solver trains an ML subgrid turbulence model inside a barotropic vorticity model, extending effective forecast lead time by up to 50%, or 73% with transfer learning.
A one-class SVM novelty detector disables ML corrections on out-of-sample inputs, stabilizing coarse climate runs and improving temperature and precipitation over standard ML corrections.
NVIDIAAllen Institute for AI (Ai2, ACE)Columbia University (LEAP, Lamont-Doherty)
Neural network corrections trained on a 3 km storm-resolving model reduce a 200 km climate model's land temperature and precipitation pattern errors by 6-25%.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
First neural emulator of a full cloud microphysics parameterization, including fast phase changes, performs well offline and online but shows Antarctic biases.
A Bayesian neural-network closure with memory, sampled by Hamiltonian Monte Carlo, gives skillful Lorenz-96 forecasts with calibrated uncertainty from noisy, sparse data.
Releases code and training data for a neural network radiation scheme in the WRF model, developed with stochastic weight averaging for numerical weather prediction.
Neural networks learn FV3-GFS systematic error from data assimilation analysis increments and apply online corrections, improving 10-day forecast skill at all lead times.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Columbia University (LEAP, Lamont-Doherty)
Random forests reconstruct nonorographic gravity wave variability at source levels from ERA5 columns with R2 near 0.85, finding local background wind most informative.
Stochastic weight averaging improves a neural radiation emulator in a Korean NWP model, giving a 60-fold radiation speedup and 12.6% week-long flux gains.
Neural network and random forest estimate surface momentum, heat and moisture fluxes from tower observations, outperforming Monin-Obukhov similarity parametrizations used in atmospheric models.
Machine-learned corrective tendencies from 25 km simulations improve a 200 km FV3GFS model's land precipitation pattern error by 8–28% across four SST-perturbed climates.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Argues that making Earth system models automatically differentiable enables objective calibration and hybrid ML components, targeting climate sensitivity, spatial biases and abrupt transitions.
Potsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)
Physics-incorporated network enforces energy conservation between fluxes and heating rates for fast radiative transfer emulation; Bi-LSTM performs best among tested architectures.
Nanjing University of Information Science & Technology (NUIST)Alibaba (DAMO Academy)
A 3D U-Net emulates rising-cloud direct numerical simulations over larger domains, matching DNS score metrics without solving the dynamical core, aiming at cloud parameterization.
Random forest surrogate for particle number concentration in GISS-ModelE2.1 reduces aerosol-cloud radiative forcing from -1.46 to -1.11 W m-2.
Replaces Monin-Obukhov surface-layer flux schemes with neural networks in large-eddy simulations, showing modified boundary-layer structure and highlighting risks of unphysical predictions.
Gradient-free Kalman inversion trains hybrid subgrid turbulence and convection parameterizations from large-eddy simulation data, with uncertainty estimates from unscented Kalman inversion.
Uses Calibrate-Emulate-Sample Bayesian learning to tune the Cloudy bulk microphysics scheme with far fewer model runs, learning a collision kernel parameter from particle-model output.
Hybrid multiscale neural operator learns grid-independent closure parameterizations for coarse dynamics, demonstrated on the chaotic multiscale Lorenz96 model.
Machine-learned corrective tendencies from nudged 25 km runs improve land precipitation pattern errors by 17-30% in a coarse FV3GFS across four climates.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Implements a neural-network gas optics scheme in ECMWF's IFS, speeding radiation by about one third with a similar model climate in 1-year coupled runs.
Residual neural networks replace moist physics and radiation in a realistic GCM, giving stable simulations over 10+ years with improved tropical land precipitation.
Tests a neural radiation emulator across 15 microphysics schemes; errors rise 8.6–41.3% in idealized cases, reducible by 26.5–50.4% with a compound parameterization.
Neural network learns unresolved-scale error in a shallow water model mimicking convection, extending forecast skill to one day; a mass-conservation penalty delays drift.
Automatic learning-rate tuning with Sherpa improves a neural radiation emulator in a Korean NWP model, reducing one-week flux and precipitation errors.
Variational encoder-decoder compresses superparameterized aquaplanet convection into five latent nodes that separate convective regimes, offering an interpretable subgrid parameterization.
Neural networks emulate a non-orographic gravity wave drag scheme and generalize to reproduce the QBO and the climate response to increased CO2 when coupled online.
Proposes ML parameterizations with parameters as inputs, calibrated online against long-term statistics, which significantly reduces long-term biases in two simple dynamical systems.
Deep learning models transpiration stress from in situ and satellite data, then couples it bidirectionally into a process model to improve global terrestrial evaporation estimates.
End-to-end differentiable training of subgrid energy-backscatter parameterizations for 2D quasi-geostrophic turbulence yields models more stable and accurate than a priori-trained schemes.
Source code and weights for a neural network radiation emulator in WRF, with hyperparameters tuned by Sherpa to speed up the radiation scheme.
A probabilistic physically informed recurrent network learns stochastic parameterizations in the Lorenz 96 atmosphere model, outperforming autoregressive and GAN baselines on temporal patterns.
The Alan Turing InstituteUniversity of Cambridge (Aardvark)University of Oxford
A DNN surrogate replaces the process-based wildfire model in E3SM, matching observed burned area with R2 0.93 and far lower parameterization cost.
A bidirectional RNN emulates shortwave radiation with fewer parameters than feed-forward nets, achieving under 1% flux error and a 4-fold CPU speedup.
A neural-network-driven stochastic cloud population model parameterizes cumulus convection in WRF, reproducing realistic precipitation statistics, MJO propagation and diurnal cycle in regional tests.
Machine-learned nudging tendencies correct a 200-km FV3GFS model toward 3-km X-SHiELD, gaining 18 hours of 3-7 day skill and 30% better precipitation.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Washington
Neural networks correct the ecRad 1D radiation scheme for 3D cloud effects, recovering 3D signal with 20-30% error at about 1% extra runtime versus the 5x costlier SPARTACUS.
Proposes tunable neural-network climate parameterizations calibrated online against long-term statistics, reducing long-term biases of the ML model in a simple dynamical system.
Symmetry-preserving CNN subgrid closures for 2D Kraichnan turbulence respect translation, Galilean and rotation invariance and stay stable in coarse simulations.
Neural-network subgrid parameterizations using non-local 3x3-column inputs improve offline skill, especially for momentum transport, with vertical velocity or divergence as key inputs.
Neural network emulates the M7 aerosol microphysics scheme in ECHAM-HAM with 77.1% mean R² and up to 64× GPU speed-up, with physical mass constraints.
Uses GANs to learn a stochastic distribution of column diabatic heating profiles from MERRA2 reanalysis, offering a probabilistic alternative to deterministic physics parameterizations.
A neural network emulates an ensemble of urban land surface models and, coupled to WRF, is more accurate and cheaper than the reference TEB scheme.
Neural network cloud cover parameterizations trained on ICON storm-resolving simulations generalise across regional and global data, with SHAP used to explain their feature use.
Columbia University (LEAP, Lamont-Doherty)Max Planck Institute for Meteorology
A neural subgrid closure trained jointly with a quasi-geostrophic solver using an a posteriori loss gives stable, realistic 2D turbulence simulations.
Neural network and random forest predict tropical Pacific rain type from MERRA-2 large-scale state, but neither beats a GLM and all miss extreme rates.
Two compound parameterizations correct errors of a neural-network radiation emulator in a Korean NWP model, keeping its 60-fold speedup with lower error.
Variational multiscale neural closure for atmospheric turbulence is shown to need training on intra-step nonphysical states to stabilise implicit time stepping.
A neural-network emulator replaces KMA's radiation scheme in an operational model, giving a 60-fold speedup and 22–34% lower skin temperature and flux errors.
Neural network emulates the M7 aerosol microphysics scheme in ECHAM-HAM, reaching R² of 89% and a 120x GPU speed-up for climate model runs.
Releases code and training data for a neural network radiation emulator in an NWP model, tied to improved forecast accuracy.
Echo state network trained on direct numerical simulations reproduces convective boundary-layer buoyancy flux statistics and generalizes to unseen flux ratios without solving the equations of motion.
Knowledge-guided U-Net++ emulates the RRTM shortwave radiation scheme for NWP, running about 10^4 times faster with good tropical generalization.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Data-driven nonlocal turbulence model whose covariance kernel is partly a neural network, calibrated to 1968 Kansas field data and more accurate than the IEC standard.
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.
Allen Institute for AI (Ai2, ACE)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of WashingtonPrinceton University
Compressed sensing learns a stochastic parameterization of fine-scale processes in the Lorenz-96 atmospheric model, combined with ensemble Kalman filtering and compared with polynomial regression.
Develops an ANN subgrid-scale model for LES of atmospheric-type wall turbulence, tested on channel flow, but finds instability when coupled into simulations.
A residual U-Net trained on reanalysis, with cubed-sphere inputs, learns global precipitation that doubles grid-point skill versus the operational IMD NWP model, as a proof of concept.
Physically regularized ML emulators of aerosol activation outperform many existing parameterizations, enabling hybrid models for Earth system models.
Neural network emulates shortwave radiative fluxes trained on ERA5 to diagnose Arctic albedo, water vapour and cloud feedbacks, capturing albedo nonlinearity that linear kernels miss.
XGBoost-based convection trigger for GCM deep convection schemes beats CAPE triggers, reaching 91-93% F1 and curbing overpredicted convection occurrence.
Calibrates cloud geometry parameters for 3D radiation using ML surrogate emulators, reducing top-of-atmosphere and surface flux errors from about 10 to 5 W/m2.
A single shallow neural network emulates the full atmospheric physics suite in NCEP GFS, remaining stable for medium-range forecasts and a year-long AMIP run with realistic output and speedup.
Shows shallow-network emulators of radiative transfer, built for an older GCM, stay stable and realistic when transplanted into a newer state-of-the-art GCM under AMIP-like runs.
Neural-network emulators replace bin-microphysics warm-rain processes in a GCM, reproducing its climate behaviour at the control run's computational cost.
Perspective proposing Neural Earth System Modelling, which fuses deep networks and Earth system models into interpretable hybrids that might eventually replace them.
Colorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Potsdam Institute for Climate Impact Research (PIK)Princeton UniversityUniversity of Washington
Neural emulators of the non-orographic gravity wave drag scheme run stably to seasonal timescales, improve medium-range forecasts and run ten times faster on GPUs than the CPU scheme.
Copula-generated synthetic data augments training of a neural emulator of longwave radiation, reducing mean absolute error by up to 62% in a toy setting.
Machine-learning-constrained autoconversion and accretion parameterizations, fitted to Azores cloud-probe data, cut uncertainty to about 15% and 5% and reveal drizzle number's role.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of WashingtonUniversity of ReadingColorado State University (incl. CIRA)
AGU 2020 presentation on blending machine learning with numerical weather models; a conference abstract with no detailed results.
Gaussian process calibration tunes boundary-layer parameterization parameters against large-eddy simulations in single-column mode, before global model tuning.
Emulator-based calibration of an atmospheric model's parameterizations against large-eddy simulations matches hand-tuned global performance while keeping process-level behaviour physically sound.
Neural networks replace gas optics in the RRTMGP radiation scheme, running 1–6 times faster with heating-rate errors below 0.1 K/day, giving 3.5x faster longwave fluxes.
Compares random forest and neural network moist-physics parameterizations trained on a 3 km global simulation; NN wins offline but crashes within 7 days when coupled.
Feed-forward DNNs emulate cloud superparameterization with real geography, explaining over 70% of mid-to-upper tropospheric temporal variance at 15-minute scale.
Physically constrained neural network subgrid parameterization trained on high-resolution simulations stays stable in coupled runs and tolerates reduced numerical precision.
Neural network radiation emulators in a cloud-resolving model run 20–100 times faster and, run frequently, improve 6 h forecast accuracy over infrequent original radiation calls.
A diagonal-structured VARX process parameterizes small-scale Lorenz '96 features, matching unimodal and trimodal invariant distributions with few parameters.
Machine-learned preconditioners for the semi-implicit linear solver in a shallow-water model are competitive with conventional preconditioners and generalize beyond the training range.
Machine-learning replacement for the autoconversion cloud microphysics process in a two-moment scheme, offered as a faster parameterization for atmospheric models.
Neural network emulators replace longwave radiative transfer in a climate model, giving up to 370x GPU speedup with under 0.1 K/day heating-rate error.
Random forests and polynomial regression trained on Perdigão sonic anemometer data remove the MYNN bias in boundary-layer TKE dissipation and cut average error by up to 30%.
Gaussian process parameter estimation for the Lorenz-96 climate benchmark outperforms deep learning and linear regression under small-data constraints, with uncertainty quantification.
ResCu, a residual CNN trained on SPCAM output, parameterizes moist convection and clouds with moist-static-energy conservation, outperforming conventional schemes in single-column tests.
A VAE trained on about 6 million high-resolution vertical velocity samples reconstructs convection and clusters organization regimes, supporting stochastic convection parameterizations.
Data assimilation corrects neural-network subgrid closures during online runs, improving long-term prediction of chaotic Lorenz 96 and Kraichnan turbulence dynamics.
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.
Neural networks emulate the RRTMGP gas optics scheme, running up to 4x faster while keeping radiative flux errors within 0.5 W/m2.
Deep neural networks emulate subgrid viscosity and diffusivity for turbulent stratified shear flows, giving a 2-2.5x simulation speedup while preserving kinetic and density variance budgets.
Eos commentary on a study finding GANs can produce stochastic parameterizations that improve how weather and climate models represent unresolved subgrid processes.
Uses neural-network probabilistic classification of reference data to build a stochastic parameterization for unresolved scales, validated on Lorenz 96 under two challenging settings.
Introduces tools to interpret neural convection parameterizations and diagnose gravity-wave instability when coupled to climate models, using SPCAM and cloud-resolving training data.
Allen Institute for AI (Ai2, ACE)Columbia University (LEAP, Lamont-Doherty)University of Washington
Data-driven super-parameterization with RNNs matches super-parameterization long-term statistics at much lower cost on Lorenz 96, with transfer learning improving generalization.
Enforces physical constraints in neural convection emulators via loss or architecture, and rescales data so models generalize to unseen climates.
Conditional GAN generates realistic cloud reflectance fields from sea surface temperature and humidity, a step toward stochastic cloud parameterization.
Uses machine learning on Darwin C-band radar data to derive transition functions for a stochastic convective-stratiform cloud model aimed at global climate model parameterization.
Overview of machine learning for model physics, covering emulating existing parameterizations, developing new ones, and enforcing physical constraints in atmospheric models.
Random forests, linear regression and CNN-based networks emulate CAM6 physical forcings offline; tuning is needed to capture detailed patterns, with online coupling left for future work.
A random forest subgrid parameterization learned from 3D high-resolution output yields stable coarse-resolution climate simulations that reproduce precipitation extremes.
Statistical models predict tropical rainfall type and amount over the next 6 hours from MERRA-2 atmospheric profiles, trained on TRMM data, capturing time-mean spatial patterns.
Proposes deep-learning data-driven super-parametrization of turbulent flow for weather and climate models; the abstract is unavailable, so details are limited.
Deep neural networks learn subgrid stresses from DNS of the atmospheric boundary layer, achieving higher correlation than Smagorinsky-based models across stability regimes and resolutions.
Domain-aware deep neural network emulates a WRF planetary boundary layer scheme over the full diurnal cycle and transfers to nearby locations.
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.
A neural network trained prognostically on ERA5 emulates northeast Pacific boundary-layer tendencies in single-column runs over 7-day periods, outperforming the underlying forecast model in tracking reanalysis.
Proposes coupled online learning, where a neural parameterization runs alongside a high-resolution simulation with nudging, to reduce instabilities and biases; illustrated on Lorenz 96.
Convolutional networks recover Lorenz-96 parameters from chaotic dynamics, a proxy for objectively inferring cloud parameterization constants in climate models.
Architecture-level enforcement of linear energy and mass conservation in neural cloud emulators reaches numerical precision and improves generalization to warmer climates.
A one-hidden-layer neural network replaces Monin–Obukhov stability functions for surface-layer fluxes, matching the standard method at seven flux sites.
Karlsruhe Institute of Technology (KIT)Forschungszentrum Jülich / Jülich Supercomputing Centre
CNN predicts moist convection effects on the environment at grey-zone resolution, using horizontal gradients, as a potential replacement for conventional cumulus parameterization.
Conditional GAN generates realistic MODIS marine low cloud reflectance fields from large-scale variables, proposed as a stochastic cloud parameterization for sub-grid cloud distributions.
Deep neural networks emulate SP-E3SM shortwave and longwave radiation with 90–95% accuracy at 8–10 times lower cost than the original scheme.
Covariance-constrained GAN emulates Rayleigh-Bénard convection, an idealized atmosphere-like turbulent flow, as a closure model and cuts training time by up to 80%.
ML learns timestep-by-timestep error corrections to a physical model; tested on Lorenz-96, it gives stable runs, better initialised skill and improved long-term statistics.
A neural network parameterization trained on coarse-grained 4 km cloud-resolving simulations is coupled to a GCM, and removing upper-level inputs stabilizes spatially extended runs.
Encoder-decoder CNNs learn total precipitation from geopotential height alone in NWP reanalysis data, offering a data-driven precipitation parameterization with high accuracy and a method to identify influential levels.
Hybrid model pairs a physical shallow-water model with a neural network trained on observations to represent unknown forcing and dissipation terms, reproducing them with correlation near 1.
Physics-informed generative learning is proposed to emulate unresolved climate-model physics; the abstract is missing, so details are limited.
A neural network emulates a WRF planetary boundary layer scheme to speed up regional weather simulations; abstract-less AGU entry.
Linear ML regression predicts 3D ozone from temperature alone, improving stratospheric ozone variability over fixed climatology for preindustrial and 4xCO2 climate sensitivity runs.
University of Cambridge (Aardvark)Karlsruhe Institute of Technology (KIT)
A random forest moist convection parameterization trained on high-resolution output runs stably in a GCM, capturing precipitation extremes and climate change when trained on warm-climate samples.
Neural network unified physics parameterization trained on 4-km aquaplanet data gives stable prognostic single-column tests better than CAM's scheme.
CBRAIN, a deep network trained on superparameterization output, skillfully emulates convective heating and moistening in an aquaplanet, a proof of concept for ML convection schemes.
A neural network trained on cloud-resolving output replaces subgrid convection in a global climate model, giving stable multi-year runs that capture precipitation extremes and approximately conserve energy.
Neural network parameterizations of typhoon-induced sea surface cooling reduce maximum wind intensity errors by 60–70% in four atmosphere-only typhoon forecast cases.
Massachusetts Institute of Technology (MIT)China Meteorological Administration (CMA)
Eos news piece proposing that climate models learn and adapt on the fly with machine learning to reduce prediction uncertainty.
CBRAIN, a neural network trained on superparameterized aquaplanet data, replaces convection parameterization and reproduces its features with reduced stochasticity.
A blueprint proposes Earth system models that learn parameterizations from observations and high-resolution simulations via data assimilation and machine learning, illustrated on a toy system.
Discusses hybrid models combining deterministic physics and machine learning components for climate modeling and weather prediction; abstract unavailable.
Neural network emulates CAM longwave radiation 50-80 times faster with near-identical 10-year climate simulations, an early ML physics emulation.
Early neural network approach computes longwave radiative budgets quickly and accurately for atmospheric models; abstract unavailable.
A stochastic latent transformer emulates a stochastically forced zonal jet model with CRPS training, giving a five-order-of-magnitude speedup for large ensemble generation.
A simple ML model links PW, CIN and wind shear to tropical rain-regime transitions, with a PW above 60 mm threshold, proposed as a convection trigger.
Commentary on reservoir-computing hybrid models coupling ML to a coarse climate model and SST predictor, arguing ML could disrupt conventional climate modeling.
Trains reservoir-computing RNNs to enforce Lyapunov spectrum and fractal dimension, giving longer stable forecasts on Lorenz 96 and a quasi-geostrophic model.
Perspective reviewing machine learning combined with data assimilation for climate surrogate models and model error correction, with open challenges for operational use.
Hybrid AI surrogate links cloud radiative effects to large-scale circulation and regional climate, enabling low-cost scenario runs for cloud-based climate intervention.
A neural-network shallow-water emulator of MPAS with tangent-linear and adjoint code enables a continuous 4D-Var assimilation system demonstrated on twin experiments.
Machine learning combined with WRF improves western North Pacific tropical cyclone track simulation; the abstract is unavailable, so the gain is unquantified.
Calibrate-emulate-sample with a Gaussian process emulator quantifies gravity wave parameterization uncertainty, showing QBO parametric uncertainty comparable to internal variability under 2xCO2.
Online neural-network model error correction is added to incremental 4D-Var in ECMWF's OOPS, showing online learning beats offline learning on a quasi-geostrophic model.
An ensemble-based algorithm targets data acquisition to regions and times that most reduce parametric uncertainty in a GCM convection scheme, within calibrate-emulate-sample.
Explainability, uncertainty & physicsMassachusetts Institute of Technology (MIT)California Institute of Technology (Caltech)
NeuroClim provides a hybrid climate modeling platform, open dataset, and stability and approximation metrics for learning climate simulation components.
Benchmarks, datasets & evaluationTsinghua University (NowcastNet)code
Calibrate-Emulate-Sample with a Gaussian process emulator quantifies gravity wave parameterization uncertainty; QBO parametric uncertainty is comparable to 2xCO2 internal variability.
Physics-informed neural network estimates 3D boundary-layer winds of extratropical cyclones over eastern North America, coupled with a synthetic track database for wind risk.
ML and causal inference reduce CMIP6 land model evaporative fraction biases at 64 FLUXNET sites, raising correlation from 0.47 to 0.80.
Climate analysis & attributionLawrence Berkeley National Laboratory / NERSC
Calibrating a convection parameterization with seasonally averaged statistics cuts calibration error up to tenfold and narrows posterior spread by two to five times in an idealized GCM.
Explainability, uncertainty & physicsCalifornia Institute of Technology (Caltech)Massachusetts Institute of Technology (MIT)
An ensemble-based experimental design algorithm, using calibrate-emulate-sample, chooses high-resolution simulation regions that most reduce uncertainty in GCM convection parameters, with 10^2 rather than 10^5 model runs.
Explainability, uncertainty & physicsMassachusetts Institute of Technology (MIT)California Institute of Technology (Caltech)
A parametric cyclone and neural network hybrid predicts typhoon storm surge at Longdong, Taiwan, with 33–40% lower RMSE than a pure network at 8 h lead.
ClimART offers over 10 million radiative transfer samples from climate conditions to benchmark neural emulators, showing shortcomings of prior datasets and architectures.
Benchmarks, datasets & evaluationEnvironment and Climate Change Canada (ECCC)code
A Gaussian process emulator couples the HadSM3 climate model to an ice sheet model, enabling efficient million-year ice sheet–climate simulations and sensitivity tests.
CHEESEHEAD19 land-atmosphere flux observations are presented as training data for machine learning to scale energy fluxes across heterogeneous surfaces to model grid cells.
Bayesian neural networks estimate convective-scale model parameters from observations within data assimilation in a 1D shallow-water perfect-model setup, reducing initial state errors.
Neural network emulator of gravity wave drag yields tangent-linear and adjoint models that succeed in 4D-Var assimilation, easing code maintenance.
Calibrate-emulate-sample with Gaussian process emulators and ensemble Kalman inversion quantifies convective parameter uncertainty in an idealized GCM at a fraction of MCMC cost.
Explainability, uncertainty & physicsCalifornia Institute of Technology (Caltech)
Neural networks trained on CESM2 output emulate 21st-century Greenland ice sheet surface melt across CMIP6 models, avoiding costly regional climate model runs.
Philosophical argument that neural-network cloud parameterizations lacking physical process representation undermine the reliability of climate model predictions.
An extended Lorenz-63 toy model reveals feedforward network instability, which training on Latin Hypercube samples removes, showing sampling design matters for stability.
Explainability, uncertainty & physicsMétéo-France (incl. CNRM)
Neural networks reproduce weak-constraint 4D-Var model error corrections in the operational ECMWF IFS and could extend correction to the whole atmospheric column.
Hybrid model combines a knowledge-based quasi-geostrophic model with deep learning, learned iteratively within 4D-Var, to correct model error and improve short- to mid-range forecasts.
Ensemble Kalman filter estimates model error from sparse noisy observations to train a neural parametrisation, yielding hybrid forecasts with better skill on Lorenz and a coupled ocean-atmosphere model.
Data assimilationLawrence Berkeley National Laboratory / NERSCUniversity of Reading
Spherenet convolution and hemisphere-specific network design improve WeatherBench CNN forecasts at 1.4 degrees for 500 hPa height and 850 hPa temperature up to day 10.
Fortran-Keras Bridge transfers Keras networks into Fortran; a climate case study tests over 100 DNN emulators of subgrid cloud and radiation physics inside a global model.
Compares logistic regression, random forests, boosted trees and neural networks predicting ERA5 thunderstorm occurrence, with about twofold skill gain over CAPE at 0–45 min.
CNN predicts daily precipitation at 14 US grid points from resolved dynamics, outperforming reanalysis and linear, forest and dense-network baselines.
