Frameworks, training and inference at scale, HPC, and compression of weather and climate data with ML.
Search and filter this topic Plain list on GitHub
AERIS is a 1.3 to 80B parameter Swin diffusion transformer reaching 10.2 ExaFLOPS on Aurora, stable to 90-day forecasts and outperforming IFS ENS.
Allen Institute for AI (Ai2, ACE)Argonne National LaboratoryUniversity of California, Los Angeles (UCLA)University of Chicago
FourCastNet uses adaptive Fourier neural operators to produce global medium-range forecasts about 80,000 times faster than NWP, scaling to thousands of GPUs.
AERIS is a 1.3 to 80B parameter Swin diffusion transformer reaching 10.2 ExaFLOPS on Aurora, stable to 90-day forecasts and outperforming IFS ENS.
Allen Institute for AI (Ai2, ACE)Argonne National LaboratoryUniversity of California, Los Angeles (UCLA)University of Chicago
Systematic study of design choices for DL weather models finds fixed-grid architectures beat grid-invariant ones, proposes a hybrid, and shows multi-step fine-tuning is essential.
CRA5 compresses 226 TB of ERA5 to 0.7 TB (over 300x) with a VAE transformer codec, and models trained on it match original-data forecast accuracy.
A minimally modified SwinV2 trained on ERA5 beats IFS, and ablations on loss, model size and multi-step fine-tuning show which training choices matter.
Lawrence Berkeley National Laboratory / NERSCUniversity of California, Berkeley
Autoencoder lossy compression reaches ratio 140 on benchmarks and 200 on 500 years of CESM climate output with negligible reconstruction error.
Position paper on Earth Virtualization Engines that pair high-resolution physics-based climate models with machine learning and federated access to exabyte-scale climate data.
CSCS - Swiss National Supercomputing CentreETH ZurichMeteoSwissMassachusetts Institute of Technology (MIT)Max Planck Institute for MeteorologyNSF NCAR / UCARNVIDIAUniversity of Reading
A DOE workshop report on AI architectures and HPC co-design for Earth system predictability, covering computing systems, data pipelines and sensor integration.
Machine learning tunes hardware and software parameters to reduce the I/O and runtime cost of the Low GloSea6 weather model.
Coordinate-based neural networks compress weather and climate grids by 300x to 3000x, beating SZ3 on error, and a 790x compressed WeatherBench loader raises RMSE by under 2%.
FourCastNet uses adaptive Fourier neural operators to produce global medium-range forecasts about 80,000 times faster than NWP, scaling to thousands of GPUs.
Presents XCast, an open Python toolkit that applies scikit-learn-style methods gridpoint-by-gridpoint to gridded climate data, simplifying statistical and ML climate forecasting workflows.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Columbia University (LEAP, Lamont-Doherty)
Tutorial for operational meteorologists covering linear and logistic regression, decision trees, random forests, boosting, naive Bayes and SVMs, with reproducible notebooks.
Conceptual proposal for FAIR, reproducible HPC-based workflows for ML weather and precipitation forecasting, using the DeepRain project as a case.
Deutscher Wetterdienst (DWD)Forschungszentrum Jülich / Jülich Supercomputing Centre
Talk slides on the challenges of preparing machine learning tools for operational weather prediction; abstract gives no quantitative results.
ML-based log parser extracts provenance from WRF model runs to build provenance graphs that trace faults in numerical weather forecast workflows.
Opinion piece arguing quantum artificial intelligence algorithms could provide breakthroughs for weather and climate forecasting and climate change science.
GeoDIP, an HPC data-integration platform, merges satellite and reanalysis inputs and raises 2-hour precipitation prediction F1 by 9% with a 97% speedup.
Characterizes compute, time and energy costs of training a UNet for precipitation nowcasting, showing mixed-precision and distributed training cut training time without losing accuracy.
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.
Web-based Image Labeler tool speeds labeling of Earth science images, such as hurricane imagery, to build shareable datasets for deep learning.
Describes a parallel HPC system generating probabilistic weather forecasts with the data-driven analog ensemble method, scaling ensemble generation across nodes.
AGU abstract-less entry applying machine learning to assess reproducibility of climate model simulations across hybrid HPC architectures.
Tiramisu and DeepLabv3+ segment extreme weather patterns and scale to 27,360 V100 GPUs at 325.8 PF/s sustained throughput.
Lawrence Berkeley National Laboratory / NERSCNVIDIAOak Ridge National Laboratory
ORBIT-2 is a 10-billion-parameter ViT downscaling foundation model reaching 0.9 km global resolution on 65,536 GPUs with R² of 0.98–0.99 on 7 km benchmarks.
ORBIT scales a vision transformer to 113 billion parameters for Earth system prediction, sustaining 1.6 exaFLOPS on 49,152 GPUs of the Frontier supercomputer.
An Arduino and Android weather station uses a backpropagation neural network on temperature and humidity, reaching a 78.6% forecast success rate on 14 trials.
A Fortran Torch Adaptor couples deep learning radiation emulators into WRF, achieving 50x to 300x GPU speedups and better forecasts than feedforward networks.
ResNet50 classifier on an edge device classifies ARM Doppler lidar cloud and clear-air images with 97.6% and 94.7% accuracy.
Federated prompt tuning of a frozen weather foundation model for on-device forecasting, using dynamic graphs to handle heterogeneity and cut communication cost.
WRF-ML coupler lets Python deep learning parameterizations run inside the Fortran WRF model, demonstrated with trained DL radiation emulators coupled to WRF.
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
Collaborative regression and neural forecasts from multiple Mauritius stations via OpenWeather API data reduce MAPE by about 5% versus non-collaborative setups.
Source code and weights for a neural network radiation emulator in WRF, with hyperparameters tuned by Sherpa to speed up the radiation scheme.
Hybrid physics-ML & parameterizationKorea Meteorological Administration (KMA)
DL4DS is a Python library for training configurable CNN downscaling models on gridded Earth data, demonstrated on CAMS air-quality fields over the Mediterranean.
ESEm, an open-source Gaussian-process emulation and calibration toolkit, reduces parametric uncertainty in a GCM and explores CMIP6 scenario uncertainty.
Data-driven global model forecasts atmospheric state and 0.25-degree precipitation without physics, with training scaled to 1024 GPUs for rapid iteration.
Proposes a streaming pipeline in which an LSTM continuously learns from live OpenWeatherMap feeds via Kafka to forecast temperature and precipitation for agriculture; no quantitative results reported.
LSTM and MLP models use city-bus sensor data to monitor and forecast temperature, humidity and pressure one day ahead.
Describes a deep-learning portal for estimating tropical cyclone intensity; no abstract is available, so the detail is limited.
Data-parallel CNN training across GPUs cuts precipitation nowcasting training time from 59 hours to about one hour.
Convolutional autoencoders compress CMIP5 pre-industrial surface temperature fields and beat PCA in reconstruction error at 40 dimensions, suggesting surrogate climate model use.
Deep learning precipitation nowcasting paired with HPC-optimized data pipelines; abstract unavailable, so the contribution is inferred from the title only.
