Machine-Learning Medium-Range Weather Forecasting
Global forecast models learned from reanalysis data that predict the next 10 to 15 days much faster than physics models.
Tier Grades
Tier A: Frontier AI labs
feasible
Google DeepMind and Huawei train on decades of ERA5 and beat operational physics models; Microsoft trains Aurora on over a million hours of data.
Tier B: Indonesian AI labs
stretch
A 256M-parameter model on public ERA5 is within reach, though training compute is not stated in the abstracts.
Tier C: Institution units such as ITB
stretch
Public ERA5 and open GraphCast code allow fine-tuning studies, though training compute is not stated.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Running a released model in under one minute per forecast fits plantation and energy forecasting proofs of value.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputePangu-Weather: about 256 million parameters in total; GraphCast: a 10-day 0.25 degree forecast in under one minute; training compute not stated in the abstracts.
- DataPangu-Weather: 43 years of hourly ERA5; FuXi: 39 years of ERA5; Aurora: over a million hours of geophysical data.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 18 authors on the GraphCast paper (arXiv 2212.12794); 6 on Pangu-Weather.
- Time horizonYears.
- Shelf-life signalsGraphCast beat the most accurate operational deterministic systems on 90% of 1380 targets; FuXi then matched the ECMWF ensemble mean at 15 days.
- Pipeline roleTier A trains global models; Tier B and Tier C fine-tune them for the Maritime Continent; Tier D runs released models for sector forecasts.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Auroraweatherexemplar
ECMWF AIFS (Single / ENS)weatherexemplar
Pangu-Weatherweatherexemplar
FuXiweatherexemplar
GraphCast (WeatherNext 1 Graph)weatherexemplarWeatherNext 3weatherproduced byFuXi-2.1weatherproduced by
FourCastNet 3weatherproduced by
Baguanweatherproduced by
Prithvi-WxCweatherproduced byExcarta (HEAL-ViT)weatherproduced by
NeuralGCM (0.7 and ENS)weatherproduced byWeatherNext 1 Graph (GraphCast)weatherproduced byFourCastNet SFNOweatherproduced by
FengWuweatherproduced byFourCastNetweatherproduced byKeisler (2022)weatherproduced byAIFS v2weatherproduced byFengWu-4DVarweatherproduced byFengWu-GHRweatherproduced byTianjiweatherproduced byWeatherMesh4weatherproduced byIFS HRESweatherused byLinks
R&D TopicMachine-Learning Medium-Range Weather Forecasting
ModelsGraphCast (WeatherNext 1 Graph)Google DeepMind, weatherPangu-WeatherHuawei (Huawei Cloud), weatherFuXiFudan University, weatherAuroraMicrosoft AI, weather+19 more
GraphCast (WeatherNext 1 Graph)Google DeepMind, weatherPangu-WeatherHuawei (Huawei Cloud), weatherFuXiFudan University, weatherAuroraMicrosoft AI, weather+19 moreUse CasesWeather ForecastingAgriculture
Models 23
WeatherNext 3Google DeepMind, Google, weatherproduced byECMWF AIFS (Single / ENS)ECMWF, weatherexemplarFengWuShanghai AI Lab, weatherproduced byFourCastNetNVIDIA, weatherproduced byFourCastNet 3NVIDIA, weatherproduced byNeuralGCM (0.7 and ENS)Google, weatherproduced byWeatherNext 1 Graph (GraphCast)Google DeepMind, Google, weatherproduced byAIFS v2ECMWF, weatherproduced byFengWu-4DVarShanghai AI Lab, weatherproduced byFengWu-GHRShanghai AI Lab, weatherproduced byFourCastNet SFNONVIDIA, weatherproduced byFuXiFudan University, weatherexemplar+11
FuXi-2.1Fudan University, weatherproduced byGraphCast (WeatherNext 1 Graph)Google DeepMind, weatherexemplarPrithvi-WxCIBM, NASA, weatherproduced byAuroraMicrosoft AI, weatherexemplarBaguanAlibaba DAMO, weatherproduced byIFS HRESECMWF, weatherused byPangu-WeatherHuawei (Huawei Cloud), weatherexemplarExcarta (HEAL-ViT)weatherproduced byKeisler (2022)weatherproduced byTianjiweatherproduced byWeatherMesh4weatherproduced byPapers
- FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural OperatorsJaideep Pathak et al., arXiv preprint, 2022 founding
- A Foundation Model for the Earth SystemCristian Bodnar et al., arXiv preprint, 2024 state of the art
- FuXi: A cascade machine learning forecasting system for 15-day global weather forecastLei Chen et al., arXiv preprint, 2023 state of the art
- WeatherBench 2: A benchmark for the next generation of data-driven global weather modelsStephan Rasp et al., arXiv preprint, 2023 state of the art
- Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather ForecastKaifeng Bi et al., arXiv preprint, 2022 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2212.12794 https://arxiv.org/abs/2212.12794
- arxiv.org/abs/2211.02556 https://arxiv.org/abs/2211.02556
- FuXi: A cascade machine learning forecasting system for 15-day global weather forecast arxiv.org https://arxiv.org/abs/2306.12873
- arxiv.org/abs/2405.13063 https://arxiv.org/abs/2405.13063
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Machine-Learning Medium-Range Weather Forecasting", https://aurion.semai.id/topics/machine-learning-medium-range-weather-forecasting/, records of October 2026.