Published October 10, 2026
| Version v1.1.0
Software
Open
brightbandtech/ExtremeWeatherBench: v1.1.0
Authors/Creators
- 1. @TheWeatherCompany
- 2. @Brightbandtech
- 3. @fraunhoferhhi
Description
Added
- Unified evaluation progress bar in
progress.py: one 0–100% bar with ETA in both serial and parallel mode, plus--no-progressandEWB_DISABLE_PROGRESS(#381). - Optional
output_format="xarray"(a flat Dataset) and CLI writes to netCDF or zarr, including--sparse(#399, #407). - Python 3.14 support (#386).
- Hypothesis property tests for forecast/target input variation (#388).
- CAPE accuracy note in the docs (#384).
- A docs check in the test suite: every
ewb.*reference inREADME.mdanddocs/must exist, Python snippets must parse, theirextremeweatherbenchimports must resolve, and calls must use real keyword arguments (#423). - A "Time coordinates" docs section describing which coordinates EWB uses to decide whether an input covers a case (#426).
- A warning when a forecast's
valid_timespansinit_timeandlead_timebut differs from their sum, which usually meanslead_timeis in the wrong units (integers are read as hours) (#426). - A warning when gridded data with a custom
preprocessreaches the coverage check without its time coordinates; previously every case was skipped with only an INFO-level "no data" log (#429).
Changed
- Peak metrics (
MaximumMeanAbsoluteError,MinimumMeanAbsoluteError,MaximumLowestMeanAbsoluteError) reduce each forecast initialization over its ownlead_timewindow, not across different model runs at a fixed lead time. Completeness guards count present values rather than coordinate slots (#393). - CAPE now marches the moist adiabat and uses a single epsilon constant. Numeric CAPE output changes (about 2.6% of the mean signal on the ERA5 reference profiles). Stored CAPE-derived scores will not match 1.0.2 (#384).
- Package imports are lazy via
lazy-loader. Legacy aliases such asewb.forecasts.ZarrForecastandewb.targets.ERA5no longer work; useewb.inputs.ZarrForecastandewb.inputs.ERA5(#352). - Requires Python
>=3.12,<3.15. Python 3.11 is no longer supported (#386). cartopymoved out of core dependencies into thedata-prepextra. Natural Earth land masks now use pooch and geopandas (#386).- Dependency refresh:
icechunk>=2.1.2,kerchunk>=0.2.10from PyPI (git override removed),pandas>=2.2.3,<3,numba>=0.66, andvirtualizarr>=2.7.3indata-prep(#386). MaximumMeanAbsoluteErrorhonors the instancereduce_spatial_dimsinstead of hardcoding latitude/longitude (#383).- Evaluation reuses forecast and target datasets within a process, groups operators that share a case and forecast, and computes aligned data once so each metric does not rebuild the same dask graph (#399).
- Gridded
preprocessfunctions run after the coverage check and case subset, so they only process the case's data. They can no longer create or rename the coordinates EWB uses to find a case (time coordinates,latitude/longitude, the 0–360 longitude convention); a 1.0.x preprocess that did so must move that work tovariable_mappingor the dataset passed toXarrayForecast(#399, #429). - The coverage check decides from one time source, in order:
init_time+lead_time(forecasts; a forecast's ownvalid_timeis ignored), thenvalid_time, thentime(#426). KerchunkForecastconverts CIRA'stimeaxis tolead_timewhen it opens a reference, so CIRA kerchunk data needs no preprocess (#429).DurationMeanErrorkeeps only the case dayofyear slice of climatology before stacking to valid time (#400).- Atmospheric river objects use 2D labels plus union-find instead of a
4D connected-component scan, skip unused forecast
(lead, valid_time)pairs, and stream masks one lead at a time (#401, #408). - Tropical cyclone spatial masks use a chunked haversine grid instead of a Python loop over IBTrACS points (#402).
- LSR parquet reads filter
valid_timeto the case window, with a fallback to the full file if the store rejects filters (#403). - Unique target pipelines are computed once in the parent and reused across forecast workers (#408).
- The tropical cyclone geopotential-thickness preprocess functions share
one implementation and read only EWB variable names:
geopotential(m²/s², divided by g) or, newly,geopotential_height(m). Anything else raises aKeyErrorpointing atvariable_mapping(#420). - Ruff and the pre-commit hook are pinned to 0.16.2, with matching lint fixes (#404).
- Physical constants live in a single
constants.pymodule;calcand_capeimport them from there instead of keeping their own copies (#411). - Tabular targets (GHCN, LSR, ...) run
preprocessafter variable mapping, so preprocess functions see EWB variable names. IBTrACS still preprocesses on its raw source columns. A preprocess function that reads source column names now raisesKeyError(#413). - Docs and README examples use the namespaced API (
ewb.inputs.*,ewb.evaluate.ExtremeWeatherBench,ewb.cases.load_cases, ...) instead of removed top-level names (#423).
Deprecated
defaults.preprocess_cira_kerchunk_forecast_dataset(no preprocess is needed) and thepreprocess_cira_kerchunk_{tc,ar,severe}_forecast_datasetfunctions (use the matchingpreprocess_cira_icechunk_*function, which handles both stores). They raise aFutureWarningand will be removed in a future release (#429).
Removed
- Python 3.11 support (#386).
- Unused core pins (
eccodes,frozenlist,pyogrio) andcartopyfrom the default install (#386). cases.load_individual_cases(usecases.load_individual_cases_from_dict),cases.load_ewb_events_yaml_into_case_list(usecases.load_casesorcases.load_ewb_cases),cases.read_incoming_yaml, andutils.read_event_yaml(usecases.load_individual_cases_from_yamlto load cases from your own YAML file) (#422).
Fixed
KerchunkForecastno longer clobbers anonymous S3 access, so public CIRA/NODD kerchunk references open without credentials (#382).open_kerchunk_referenceno longer mutates the caller'sstorage_optionsdict (#382).- Documented example config uses
case_list(the name the CLI expects) (#382). check_for_spatial_datano longer crashes on antimeridian-crossing-180/180grids, and now requires a match on both latitude and longitude (#387).- Peak metrics and
DurationMeanErrorreturn NaN on degenerate time axes instead of raising (#387). - Integer
lead_timevalues are treated as hours, not nanoseconds, inconvert_init_time_to_valid_time(#387). maybe_densify_dataarraycopies instead of mutating the parent dataset (#383).- Landfall, timestep-completeness, and temporal-resolution helpers no longer materialize whole arrays to answer emptiness/shape questions (#383).
outputsis exported from the package loader, and empty-fallback time dims no longer poison Dataset dtypes (#407).- Point observations (GHCN, LSR) use a
locationdimension instead of a unique-lat × unique-lon mesh. Alignment samples the forecast at those stations, which avoids out-of-memory on large freeze cases and makes spatial-mean metrics average stations only (#410). Region.maskkeeps only the longitudes inside antimeridian- and prime-meridian-crossing regions instead of the full longitude axis (#416).Region.maskreturns an empty subset instead of raisingValueErrorwhen a wrapping region has no longitudes in the dataset (#424).- Forecasts whose
valid_timeis a non-indexed coordinate (all CIRA icechunk models: FourCastNetv2, GraphCast, Pangu, ...) are no longer rejected as having no data for the case. With xarray 2026.7, slicing a non-indexed coordinate returns nothing instead of raising, so these forecasts were silently skipped and, for example,applied_freeze.pyproduced zero rows (#391, #426). The data check now uses each forecast'sinit_time+lead_time, so forecasts initialized before a case whose lead times reach into it are also kept. - Kerchunk examples in the docs point at references that exist
(
gs://extremeweatherbench/<MODEL>.parq) (#429). CravenBrooksSignificantSevereoutput carries its own attributes instead of CAPE's (#363, #421).cases.load_cases()no longer logs a bogus deprecation warning (#422).- The CAPE accuracy note points at
constants.pyforEPSILONandMOIST_ASCENT_SUBSTEPS(#423).
Files
brightbandtech/ExtremeWeatherBench-v1.1.0.zip
Additional details
Related works
- Is supplement to
- Software: https://github.com/brightbandtech/ExtremeWeatherBench/tree/v1.1.0 (URL)
Software
- Repository URL
- https://github.com/brightbandtech/ExtremeWeatherBench