BUG: All NOAA NOMADS Dependent Atmosphere Models Broken

[MateusStano]

OpenDAP support for NOMADS environmental data fetching no longer works

You can find more information here PDF

The change is big and is definitely not easy to fix. From what I understand there are two ways we can go about fixing this

• Use the Siphon library

  • We can use a tool called Siphon. It connects to a different type of server (THREDDS) that still works.
  • The Good: It is very similar to our old method. It keeps the data in a format that RocketPy already understands (netCDF4).
  • The Bad: We have to add siphon as a new requirement for RocketPy.

• Use "GRIB Filter" with basic web requests

  • We can write code to download small data files directly from NOAA using simple web links.
  • More on this here: https://nomads.ncep.noaa.gov/info.php?page=gribfilter
  • The Good: We do not need to add any new libraries
  • The Bad: The files come in a format called "GRIB2." RocketPy usually uses "NetCDF." Reading GRIB2 files seems difficult and ive read that it can be cumbersome to do it on windows

It seems that using Siphon is the way to go here, but I might be missing an easier way to fix the issue

[Gui-FernandesBR]

This is my claude's review:

Context

What happened: NOAA terminated OpenDAP access on nomads.ncep.noaa.gov effective Feb 24, 2026 (SCN25-81). All RocketPy fetchers in fetchers.py call netCDF4.Dataset(opendap_url) which no longer works.

Affected models: GFS, NAM, RAP, HIRESW, GEFS (all 5 active NOMADS-based models)

NOAA's recommended migration paths:

1. GRIB Filter: https://nomads.ncep.noaa.gov/cgi-bin/filter_gfs_0p25.pl?... (returns GRIB2 subsets)
2. HTTPS bulk download: https://nomads.ncep.noaa.gov/pub/data/nccf/com/ (full GRIB2 files)

Current RocketPy deps: numpy, scipy, matplotlib, netCDF4, requests, pytz, simplekml, dill

---

Current Architecture (what breaks)

fetchers.py:
  fetch_gfs_file_return_dataset()
    → netCDF4.Dataset("https://nomads.ncep.noaa.gov/dods/gfs_0p25/...")  ← DEAD
    → returns netCDF4.Dataset object

environment.py:
  set_atmospheric_model(type="Forecast", file="GFS")
    → calls fetcher → gets netCDF4.Dataset
    → process_forecast_reanalysis(dataset, dictionary)
      → data.variables["tmpprs"][time, :, lat, lon]  ← expects netCDF4 interface
      → bilinear_interpolation → Function objects

Key files:

• rocketpy/environment/fetchers.py (lines 95-406) — 6 broken fetchers
• rocketpy/environment/environment.py (line 1675) — process_forecast_reanalysis() consumes netCDF4.Dataset
• rocketpy/environment/weather_model_mapping.py — variable name mappings (tmpprs, ugrdprs, etc.)
• rocketpy/environment/tools.py — helpers (find_time_index, bilinear_interpolation, etc.)

---

Option Analysis

Option 1: Siphon + Unidata THREDDS

How: Use siphon library to query Unidata's THREDDS server (thredds.ucar.edu), which mirrors NOAA data and returns netCDF4 via NCSS (NetCDF Subset Service).

python

from siphon.catalog import TDSCatalog
cat = TDSCatalog('https://thredds.ucar.edu/thredds/catalog/grib/NCEP/GFS/Global_0p25deg/catalog.xml')
ncss = list(cat.datasets.values())[0].subset()
query = ncss.query()
query.lonlat_point(lon, lat)  # or lonlat_box for 2x2 grid
query.time(datetime)
query.accept('netcdf4')
query.variables('Temperature_isobaric', 'u-component_of_wind_isobaric', ...)
data = ncss.get_data(query)  # returns netCDF4.Dataset

Verified available on THREDDS:

• GFS 0.25° ✅ (grib/NCEP/GFS/Global_0p25deg)
• NAM CONUS 12km ✅ (grib/NCEP/NAM/CONUS_12km)
• RAP CONUS 13km ✅ (grib/NCEP/RAP/CONUS_13km)
• GEFS — needs verification (likely available)
• HIRESW — needs verification

New dependency: siphon (pure Python: requests, protobuf, beautifulsoup4)

Criteria | Rating -- | -- Refactor scope | Small — returns netCDF4, same interface as before New dependencies | Light — siphon is pure Python, pip-installable everywhere Windows support | Excellent — no C libraries needed Server reliability | Medium — single server (thredds.ucar.edu), run by Unidata/NSF Model coverage | Good — GFS, NAM, RAP confirmed; GEFS/HIRESW likely Data subsetting | Yes — NCSS supports point/box queries, variable/level selection Maintenance | Good — maintained by Unidata (same team behind netCDF4)

---

Recommendation

For RocketPy's use case (single-point atmospheric profiles for rocket simulation):

Primary recommendation: Option 1 (Siphon + THREDDS) as an optional dependency rocketpy[forecast]

Rationale:

1. Minimal disruption — returns netCDF4, so process_forecast_reanalysis() needs only variable name mapping updates
2. Pure Python — installs via pip install siphon on all platforms, no C library issues
3. Lightweight — doesn't bloat RocketPy with weather-specific heavy deps
4. Good enough reliability — Unidata THREDDS has been running reliably for years; used by pvlib, MetPy, and many other projects
5. Clean API — catalog-based discovery instead of URL guessing

If reliability is a top concern, consider Option 5 (Hybrid) with GRIB filter fallback, accepting the cfgrib dependency cost.

If NOAA-official matters most, go with Option 2 (GRIB Filter + cfgrib), accepting the larger refactor and C library dependency.

---

Implementation Sketch (for chosen approach)

If Siphon is chosen:

1. Add optional dependency group in pyproject.toml:

toml

   [project.optional-dependencies]
   forecast = ["siphon>=0.9"]

2. Rewrite fetcher functions in fetchers.py:
   • Replace netCDF4.Dataset(opendap_url) with Siphon NCSS queries
   • Return netCDF4.Dataset-compatible object
   • THREDDS catalog URLs for each model
3. Update weather_model_mapping.py:
   • THREDDS variable names differ (e.g., Temperature_isobaric vs tmpprs)
   • Add THREDDS-specific mappings
4. Minimal changes to environment.py:
   • Pass lat/lon to fetchers (currently not passed; fetchers return full dataset)
   • OR keep returning full dataset and let process_forecast_reanalysis handle extraction
5. Update tests

If GRIB Filter is chosen:

1. Add optional dependency: cfgrib, eccodes, xarray
2. Rewrite fetchers to download GRIB2 via HTTP + parse with cfgrib
3. Rewrite data extraction in process_forecast_reanalysis (or create adapter)
4. New variable name mappings for GRIB2 conventions
5. Handle GRIB2-specific data structures (flat messages vs 4D arrays)
6. Update tests

---

Open Questions for Decision

1. Which option do you want to proceed with?
2. For Siphon: should we verify GEFS and HIRESW availability on THREDDS before committing?
3. Should there be a graceful error message when the optional dep is not installed?
4. Should we deprecate/remove any models in this fix, or maintain all 5?