WeatherPre

One line to forecast. One line to compare.
Weather (≤ 15 days) and sub-seasonal-to-seasonal (> 15 days) forecasts from 29 AI and NWP models (+ 2 baselines) (GraphCast, AIFS, Aurora, FourCastNet 3, GEFS, ECMWF extended range, CFSv2 …), scored against ERA5 with WeatherBench-X.

English · 中文 · Examples · Survey · GPU routes · Contributing

import weatherpre as wp

ds  = wp.forecast("graphcast", "z500", "7d", init="2020-10-03")            # model, variable, horizon -> xarray
cmp = wp.compare(["graphcast", "pangu", "gefs", "ifs"], "z500,t2m", "7d", init="2020-10-03")
cmp.ranking()                                                               # who wins, scored vs ERA5

✨ Why WeatherPre

Research universe update (19 Oct 2026): PBC calibration, GenFocal downscaling and gray-swan ERF diagnostics are available. wp.load_research(...) defaults to official author code/weights (lite=True is explicit), with optional weatherpre[research]. Their catalog status is research; untrained lite outputs do not enter forecast rankings.

• Model × variable × horizon → result. No GRIB plumbing, no regridding, no per-model APIs: one xarray.Dataset with the same names and units for every model.
• Two forecast scales, one interface. "48h", "7d", "15d" → weather (fields at lead hours); "6w", "45d", "week3-4" → S2S (weekly means and anomalies).
• Accuracy built in. wp.compare(...) scores every model on the same grid with WeatherBench-X (area-weighted RMSE and ACC; ERA5 for the past, IFS analysis for recent cycles), averages over many inits, ranks the models and draws the figures.
• No GPU needed for 13 models + 2 baselines. Hosted forecasts are streamed from WeatherBench 2, ECMWF Open Data and NOAA's open buckets. GPU models (AIFS v2, Aurora 1.5, FourCastNet 3, Atlas, U-CAST, WeatherNext 2, FuXi-S2S …) run through NVIDIA Earth2Studio or a Hugging Face ZeroGPU Space with the same call.
• Nothing re-implemented. Every route wraps an upstream product or package, so you get the official numbers.
• Publication-ready figures. Robinson maps, skill curves and S2S anomaly panels in one call.

🚀 Install

pip install "weatherpre @ git+https://github.com/GISWLH/WeatherPre"      # Python >= 3.11
pip install "weatherpre[gpu] @ git+https://github.com/GISWLH/WeatherPre" "earth2studio[fcn3]"   # optional: GPU models

⚡ Quick start

import weatherpre as wp

# weather scale (<= 15 days): fields at lead hours
ds = wp.forecast("aifs", "t2m", "15d")                                   # latest ECMWF AIFS cycle, 0.25°
ds = wp.forecast("gefs", ["z500", "tp"], "48h", init="2020-10-03")       # NOAA GEFS ensemble mean
ds = wp.forecast("aurora", "msl", "7d", init="2020-10-03")               # GPU: HF ZeroGPU (HF_TOKEN) or Earth2Studio
ds = wp.forecast("e2s:FCN3", "t2m", "7d", init="2020-10-03")             # any Earth2Studio model by class name

# S2S scale (> 15 days): weekly means, week k = days 7(k-1)..7k after init
s = wp.forecast("ifs-ext", "t2m", "6w", init="2020-10-01")               # ECMWF extended range, weeks 1-6
s = wp.forecast("cfsv2", "tp", "week3-4")                                # latest CFSv2, weeks 3-4
s = wp.forecast("canglong", "t2m", "6w")                                 # latest CAS-Canglong run (canglong.earth), weeks 1-6
wp.anomaly(s)                                                            # vs ERA5 1990-2017 weekly climatology

# accuracy comparison
cmp = wp.compare("auto", "z500", "15d", init="2020-10-03")               # best hosted models for that date
cmp = wp.compare(["ifs-ext", "gefs", "cfsv2", "persistence", "climatology"], "t2m", "6w",
                 init="2020-10-01..2020-10-29")                          # every Mon/Thu init, scores averaged
cmp.table("acc"); cmp.ranking(); cmp.plot("scores.png"); cmp.plot_maps("week3.png", "t2m", 3)

Same thing from the shell:

weatherpre models --scale s2s                                 # what can I run? (--all adds GPU models)
weatherpre forecast graphcast z500 7d --init 2020-10-03 --plot
weatherpre compare ifs-ext,gefs,cfsv2,climatology t2m 6w --init 2020-10-01   # table + ranking + figures

lead=	scale	what you get
"48h" · "7d" · "15d"	weather	dims (lead, latitude, longitude); leads every 6 h to 48 h, 12 h to 7 d, 24 h to 15 d
24 · [6, 12, 24] · "24,48,72"	weather	exactly these lead hours
"6w" · "45d" · "2m" · "s2s"	S2S	dims (week, latitude, longitude), weeks 1..N, weekly means (tp in mm/day)
"week3-4" · "w3"	S2S	only those weeks

Variables: z500 [m], t850 [K], t2m [K], msl [hPa], tp [mm, accumulated since init; mm/day for S2S] (aliases like "precip", "mslp", "T2M" work). init: date, "YYYY-MM-DDTHH", "latest", a list, or "START..END" for compare.

📊 Accuracy at a glance

S2S scale: weeks 1–6, every Monday/Thursday init of October 2020 (9 inits), weekly means vs ERA5 — one call:

cmp = wp.compare(["ifs-ext", "gefs", "cfsv2", "persistence", "climatology"], "t2m,z500,tp", "6w", init="2020-10-01..2020-10-29")

t2m ACC (weekly mean anomaly)	wk 1	wk 2	wk 3	wk 4	wk 5	wk 6
ECMWF extended range (ens. mean)	0.84	0.64	0.41	0.36	0.41	0.41
NOAA GEFS (ens. mean, to 35 d)	0.77	0.58	0.42	0.40	0.44	–
NOAA CFSv2 (member 1)	0.74	0.45	0.20	0.22	0.24	0.17
ERA5 anomaly persistence	0.54	0.37	0.37	0.27	0.22	0.24

t2m RMSE [K]	wk 1	wk 2	wk 3	wk 4	wk 5	wk 6
ECMWF extended range	0.89	1.41	1.81	1.95	1.94	2.04
NOAA GEFS	1.21	1.59	1.86	1.97	2.00	–
climatology	1.89	1.91	1.97	2.04	2.10	2.22
NOAA CFSv2	1.31	2.01	2.52	2.57	2.65	2.82
persistence	1.83	2.16	2.19	2.39	2.52	2.58

Skill drops fast after week 2: from week 3 on the ensemble means keep a t2m ACC around 0.4 but their RMSE is barely below climatology, and z500 ACC falls to 0.1–0.3. A single, non-bias-corrected CFSv2 member does worse than persistence for t2m. All numbers (z500, tp, per init): results/examples/s2s_2020-10_scores.csv.

Weather scale, 2020-10-03 00Z, next 7 days (wp.compare(["graphcast","pangu","gefs","ifs-hres"], "z500,t2m", "7d", init="2020-10-03"), vs ERA5, 1.5°):

mean rank	model	z500 RMSE +72 h [m]	z500 RMSE +168 h [m]	t2m RMSE +168 h [K]
1.3	GraphCast	12.9	44.0	1.54
1.8	Pangu-Weather	12.7	42.7	1.59
3.2	IFS HRES	16.2	47.4	1.84
3.8	GEFS ens. mean (0.5° → 1.5°)	19.7	43.8	2.04

More (next 48 h / week / 15 days with 7 models, the latest live cycle, maps): docs/EXAMPLES.md.

🗂️ Model zoo

✅ hosted = runs anywhere, no GPU (streamed from open archives) · 🟡 GPU = Earth2Studio / HF ZeroGPU, same call · ⛔ = no open route. weatherpre models --all prints this table.

Weather scale (≤ 15 days)

Model	Type	Status	Route	Grid	Lead	Hosted period	Licence
ifs-hres ECMWF IFS HRES	NWP	✅ hosted	opendata, wb2	0.25° / 1.5°	≤15 d	latest ~4 d (Open Data) · 2016–2022 (WB2)	CC-BY-4.0
aifs-single ECMWF AIFS	AI	✅ hosted	opendata	0.25°	≤15 d	latest ~4 d (Open Data)	CC-BY-4.0
aigfs NOAA AIGFS	AI	✅ hosted	noaa	0.25°	≤16 d	2026-04-16 → now (NOAA S3)	public domain
gfs NOAA GFS	NWP	✅ hosted	noaa	0.25°	≤16 d	2022 → now (NOAA S3)	public domain
gefs NOAA GEFS ens. mean	NWP	✅ hosted	gefs	0.5°	≤35 d (00Z)	2020-09-23 → now (NOAA S3, GEFSv12)	public domain
graphcast GraphCast	AI	✅ hosted	wb2, Earth2Studio GraphCastOperational	1.5° (hosted)	≤10 d	2019-11 → 2021-01 (WB2)	weights CC-BY-NC-SA-4.0
pangu Pangu-Weather	AI	✅ hosted	wb2, Earth2Studio Pangu6	1.5° (hosted)	≤10 d	2018–2022 (WB2)	weights CC-BY-NC-SA-4.0
fuxi FuXi	AI	✅ hosted	wb2, Earth2Studio FuXi	1.5° (hosted)	≤15 d	2020 (WB2)	see upstream
gencast GenCast (ens. mean)	AI	✅ hosted	wb2	1.5°	≤15 d, 12 h	2020 (WB2)	weights CC-BY-NC-SA-4.0
neuralgcm NeuralGCM	AI	✅ hosted	wb2	1.5°	≤15 d, 12 h	2020 (WB2)	see upstream
ifs-ens ECMWF IFS ENS mean	NWP	✅ hosted	wb2	1.5°	≤15 d	2018–2022 (WB2)	WB2 terms
cfsv2 NOAA CFSv2	NWP	✅ hosted	cfs	1°	≤9 months	2020 → now (NOAA S3), member 1	public domain
aurora Aurora	AI	🟡 GPU	hf, Earth2Studio Aurora	0.25°	any (chunked)	any date (ERA5 / IFS ICs)	MIT
aurora-1.5 Aurora 1.5	AI	🟡 GPU	Earth2Studio Aurora1p5_6h	0.25°	6 h steps	any date	MIT
aifs2 ECMWF AIFS v2	AI	🟡 GPU	Earth2Studio AIFS2	0.25°	6 h steps	latest (IFS ICs)	CC-BY-4.0
aifs2-ens ECMWF AIFS-ENS v2	AI	🟡 GPU	Earth2Studio AIFS2ENS	0.25°	6 h steps	latest (IFS ICs)	CC-BY-4.0
fcn3 FourCastNet 3	AI	🟡 GPU	Earth2Studio FCN3	0.25°	6 h steps	any date	see model card
atlas NVIDIA Atlas	AI	🟡 GPU	Earth2Studio Atlas	0.25°	6 h steps	any date	see model card
ucast U-CAST	AI	🟡 GPU	Earth2Studio UCast	1.5°	12 h steps	any date	see model card
weathernext2 WeatherNext 2 (Cyclones)	AI	🟡 GPU	Earth2Studio WeatherNext2Cyclones	0.25°	6 h steps	any date	see model card
gencast-mini GenCast mini	AI	🟡 GPU	Earth2Studio GenCastMini	1°	12 h steps	any date	see model card
sfno SFNO	AI	🟡 GPU	Earth2Studio SFNO	0.25°	6 h steps	any date	see model card
fengwu FengWu	AI	🟡 GPU	Earth2Studio FengWu	0.25°	6 h steps	any date	unclear (keep private)
weathernext Google WeatherNext 2/3 (hosted)	AI	⛔	–	0.25°	≤15 d	allow-list only	experimental ToS

S2S scale (> 15 days, weekly means)

Model	Type	Status	Route	Grid	Lead	Hosted period	Licence
gefs NOAA GEFS ens. mean	NWP	✅ hosted	gefs	0.5°	≤35 d (00Z)	2020-09-23 → now (NOAA S3, GEFSv12)	public domain
ifs-ext ECMWF extended range (ens. mean)	NWP	✅ hosted	wb2-ext	1.5°	46 d, weekly means	2016–2022, Mon/Thu inits (WB2)	WB2 terms
cfsv2 NOAA CFSv2	NWP	✅ hosted	cfs	1°	≤9 months	2020 → now (NOAA S3), member 1	public domain
canglong CAS-Canglong (atmosphere)	AI	✅ hosted	canglong	1°	6 weeks, weekly means	latest live run (canglong.earth)	see canglong.earth
climatology ERA5 climatology	baseline	✅ hosted	baseline	1.5°	any	any date (1990–2017 clim.)	Copernicus
persistence ERA5 anomaly persistence	baseline	✅ hosted	baseline	1.5°	any	1959 – 2023-01 (ERA5 in WB2)	Copernicus
ucast U-CAST	AI	🟡 GPU	Earth2Studio UCast	1.5°	12 h steps	any date	see model card
fuxi-s2s FuXi-S2S	AI	🟡 GPU	Earth2Studio FuXiS2S	1.5°	42 d, daily	any date	CC-BY-NC-ND
dlesym DLESyM (atmos.+ocean)	AI	🟡 GPU	Earth2Studio DLESyMLatLon	~1° HEALPix	6 h → seasonal	any date	see model card
ace2 ACE2-ERA5	AI	🟡 GPU	Earth2Studio ACE2ERA5	1°	6 h → climate	any date (forcings)	Apache-2.0

Research methods in the forecast universe (usage)

Method	Role	Availability
pbc	calibration	fit historical CDFs, climatology and observable lags first
genfocal	downscaling	official JAX/Flax, published ReFlow and SR/EMA weights
translocation	diagnostic	official pretrained GraphCast + signed ERF

🧭 How it works

wp.forecast(model, variable, lead, init)
   │  catalog.py: scales, routes, periods, licences          leads.py: "7d" -> weather leads | "6w" -> S2S weeks
   ▼
 resolve()  ── hosted first ──►  WeatherBench 2 (GCS zarr) · ECMWF Open Data · NOAA S3 (GFS, AIGFS, GEFS, CFSv2) · canglong.earth
            ── then GPU ──────►  Earth2Studio (any px model) · HF ZeroGPU Space (Aurora)
   ▼
 one schema: (lead | week, latitude, longitude), z500 / t850 / t2m / msl / tp
   ▼
 wp.compare(): regrid to the WB2 1.5° grid (or 0.25°) ─► WeatherBench-X RMSE / ACC ─► table · ranking · figures

Truth: ERA5 6-hourly (weather, 2018–2021), ERA5 weekly means + 1990–2017 weekly climatology (S2S, to 2023-01), IFS analysis as a proxy for recent weather cycles. Future valid times cannot be scored; for those, compare reports the inter-model spread.

⚠️ Honest limits

• Hosted archives differ: WeatherBench 2 forecasts are 1.5° and mostly 2018–2022; ECMWF Open Data keeps only ~4 days; GEFSv12 starts 2020-09-23 and reaches 35 days (00Z); CFSv2 (member 1) is not bias-corrected, so its S2S anomalies carry model climate drift (e.g. over high terrain).
• Scores of models with other native grids are computed after box-smoothing to the WB2 1.5° grid; this is fair between models but not identical to WB2's conservative regridding.
• GPU routes are tested end-to-end with Earth2Studio's Persistence model on CPU; the large models need a CUDA GPU and their Earth2Studio extra, and several weights are non-commercial.
• Aurora in 2020 uses ERA5 initial conditions inside its training period (a pipeline check, not out-of-sample skill).
• canglong serves only the latest canglong.earth run (no archive), so it cannot be scored against ERA5 yet. It is not drift-corrected: for the run starting 2026-10-08, the area-mean 2 m temperature poleward of 60°N is +5 K above the ERA5 1990–2017 weekly climatology in week 1 and +14 K in week 6, while 60°S–60°N stays within ±1 K. Treat its high-latitude wp.anomaly() as model drift.

🛣️ Roadmap

• Probabilistic scores (CRPS, spread–skill) for ensembles (GEFS members, AIFS-ENS, GenCast, FCN3)
• Bias-corrected S2S anomalies from reforecasts (ECMWF / CFSv2 hindcasts)
• Point / city time series (wp.forecast(..., at=(35.7, 139.7))) and regional crops
• Published GPU runs (Earth2Studio) for a shared leaderboard
• Hosted docs site

Ideas and new models are welcome: open an issue or see CONTRIBUTING.md.

📚 Cite & licence

Code: Apache-2.0 (CITATION.cff). Data and weights keep their own terms: ECMWF Open Data CC-BY-4.0 (credit ECMWF), NOAA public domain, ERA5 / WeatherBench 2 per Copernicus and WB2 terms; non-commercial weights are never committed. Please also cite the models and WeatherBench 2 you use.

Built on WeatherBench 2 / WeatherBench-X, Earth2Studio, ecmwf-opendata, microsoft-aurora, NOAA Open Data Dissemination and cartopy-robinson-lat-clip. Related: GISWLH/WeatherAI (model ports).

Editable artwork (Excalidraw): header · logo