[GISWLH]

Summary

Two forecast scales, one interface. Leads ≤ 15 days are weather (fields at lead hours); leads > 15 days are S2S (weekly means, week k = days 7(k-1)..7k after init).

ds  = wp.forecast("graphcast", "z500", "7d", init="2020-10-03")
cmp = wp.compare(["ifs-ext", "gefs", "cfsv2", "persistence", "climatology"], "t2m", "6w", init="2020-10-01..2020-10-29")
cmp.table(); cmp.ranking(); cmp.plot("scores.png"); cmp.plot_maps("week3.png", "t2m", 3)

Simple interface

• wp.forecast(model, variable, lead, init=...). Lead strings: 48h, 7d, 15d, 6w, 45d, week3-4, w3, 24,48,72. Variable aliases work (precip, mslp, T2M).
• wp.compare(...) returns a Comparison with .table(), .ranking(), .summary(), .plot() and .plot_maps().
  • init="START..END" averages over several inits: RMSE is pooled as the square root of the mean MSE, ACC is the plain mean.
• Old keyword calls still work: lead_days=, lead_hours=, preset=, and init as the second positional argument.
• New CLI: weatherpre forecast MODEL [VAR] [LEAD], weatherpre compare MODELS [VAR] [LEAD], weatherpre models --scale s2s [--all].
  • Breaking: the old compare --models ... command is now compare-wb2.

Model catalog and new models

• weatherpre/catalog.py is now the single source of truth. It drives backend routing, weatherpre models, and the generated README tables (scripts/gen_model_table.py).
• New hosted models (no GPU needed):
  • NOAA GEFSv12 ensemble mean: 0.5°, to 35 days, 2020-09-23 onwards.
  • NOAA CFSv2: 9 months, member 1.
  • ECMWF extended range ensemble mean: WeatherBench 2, 2016–2022.
  • S2S baselines: climatology and persistence.
• Earth2Studio route rewritten:
  • Catalog entries: AIFS v2, AIFS-ENS v2, Aurora 1.5, FourCastNet 3, Atlas, U-CAST, WeatherNext 2, GenCast-mini, SFNO, FengWu, FuXi-S2S, DLESyM, ACE2.
  • Any "e2s:<ClassName>" also works.
  • Class names were checked against earth2studio 0.19.

S2S verification

• Weekly means are scored against ERA5 weekly means, with RMSE and ACC (ACC uses the 1990–2017 weekly climatology), via WeatherBench-X.
• Models on other grids are box-smoothed to the WeatherBench 2 1.5° grid before scoring.
• S2S anomaly maps are added (wp.anomaly, maps.plot_s2s), plus a combined RMSE/ACC skill plot.

Project / README

• The README is now bilingual (README.md in English, README_zh.md in Chinese), with a hero figure and real accuracy tables.
• The long 2020-10-03 weather examples moved to docs/EXAMPLES.md.
• Added: tests, ci.yml (ruff + pytest on 3.11 and 3.12; live-data tests are opt-in), CONTRIBUTING, CITATION.cff, CHANGELOG and issue templates.
• pyproject now has the extras [gpu], [hf] and [dev], and the version is unified at 0.3.0.

Fixes

• The interpreter crashed at exit ("double free") when eccodes was loaded before PROJ. Fixed by importing pyproj before cfgrib.
• Scoring failed when there was only one lead or one week (WeatherBench-X results were over-squeezed). Fixed.
• When the latest cycle is scored against the IFS analysis, models on other grids (e.g. GEFS) are now interpolated to the analysis grid. Any remaining scoring error is reported as skipped instead of crashing.
• GEFS S3 requests now retry on 503 "Slow Down".

Results (real runs, in the README)

• S2S, October 2020 (all 9 Monday/Thursday inits), t2m ACC:
  • ECMWF extended range: 0.84 in week 1, 0.41 in week 3.
  • GEFS: 0.77 in week 1, 0.42 in week 3.
  • From week 3 on, ensemble-mean RMSE is only slightly below climatology.
• Weather, 2020-10-03, 7 days, mean rank: GraphCast 1.3, Pangu 1.8, IFS HRES 3.2, GEFS 3.8.

Test plan

• pytest -q: 52 passed (offline)
• WEATHERPRE_NETWORK=1 pytest -m network: 5 passed. Covers WeatherBench 2, GEFS, ifs-ext vs climatology, single lead/week scoring, and Earth2Studio Persistence on CPU.
• ruff check weatherpre tests scripts
• python -m compileall weatherpre scripts hf (the check workflow from main)
• CLI models, forecast, compare run end to end. Each README example was run once. Exceptions: the aurora / e2s:FCN3 lines need a GPU or HF token, and latest AIFS was reached only through the AWS mirror (data.ecmwf.int is blocked here), where S3 throttled it.
• Large Earth2Studio models (FCN3, AIFS v2, …) on a real GPU. This was not possible here: no GPU.
• First GitHub Actions run of ci.yml. It installs weatherbenchX from git and may need tuning.

🤖 Generated with Claude Code

https://claude.ai/code/session_012XaUZ3NVhMrZPpC27Jw2Js

---

Generated by Claude Code