NOAA RRFS and REFS source data exploration

[aldenks]

Run the source data exploration guide on NOAA's Rapid Refresh Forecast System (RRFS, deterministic) and the RRFS Ensemble Forecast System (REFS), leading with the AWS Open Data buckets and noting NOMADS and other mirrors as fallbacks.

[aldenks]

NOAA RRFS and REFS source data exploration

Filled-out source data exploration for the NOAA Rapid Refresh Forecast System (RRFS) deterministic forecast and the RRFS Ensemble Forecast System (REFS). Explored on 2026-09-16/17 for issue #1069. Every value below was read from live S3 listings, .idx files, or GRIB2 files opened with rasterio; NWS/NCO documentation is cited only to interpret what the files show and for the license.

Key takeaways

• RRFS v1 and REFS v1 become operational on 2026-10-14 12Z (SCN 26-48 Updated); they replace NAM, HREF, SREF and HiresW (SCN 26-47 Updated). HRRR and RAP are not retired by these notices; HRRR is a REFS member.
• Preferred source: the AWS Open Data operational bucket s3://noaa-rrfs-ops-pds (anonymous S3 and HTTPS, .idx for every GRIB, CC0-1.0). It has carried the pre-implementation parallel feed since 2026-08-12 with the exact operational file layout, and the same bucket will carry the operational feed. This is a forward-rolling archive: as of exploration it holds ~5 weeks.
• The older prototype bucket s3://noaa-rrfs-pds is no longer a usable archive — its rrfs_a/ and rrfs_public/ real-time prefixes have been emptied. Only three ~1-month retrospective case-study windows (Jul 2023, Jan–Feb 2024, May 2024) remain, on a different (pre-v1) file layout.
• NOMADS mirrors the same layout under .../com/rrfs/{para,v1.0}/ and .../com/refs/{para,v1.0}/ with ~2 days of retention — a fallback only. No Google Cloud or Azure mirror was found.
• The RRFS/REFS conus 3 km grid is exactly the HRRR CONUS grid (1799×1059 Lambert conformal, identical WKT, origin and pixel centers to noaa/hrrr/.../templates/latest.zarr). The ak (polar stereographic, 1649×1105), hi/pr (Mercator, 2.5 km) and na (13 km rotated lat-lon) domains are each their own grid.
• RRFS cycles: hourly-lead 2dfld/prslev files exist only for 00/06/12/18Z (f000–f084, all domains) and 03/09/15/21Z (f000–f018, no na domain). The other 16 cycle hours publish only the sub-hourly subh file (34 surface fields at 15/30/45/60 min, f001–f018), which does include the top-of-hour values.
• REFS = ensemble products (refs.*/ensprod, 8 product types × 4 domains, f01–f60 hourly) computed from 14 members: the current and 6 h-old RRFS deterministic and 5-member rrfsens runs (12), plus current and 6 h-old HRRR for conus/ak (2). The individual rrfsens members (m001–m005, f000–f060) appeared in the bucket only from 2026-09-09 12Z; refs.* products from 2026-08-12 18Z.
• Core-variable gaps to plan around: RRFS has all 14 core variables (MSLET rather than PRMSL). REFS mean/sprd/prob lack DSWRF, DLWRF, surface PRES, 2 m RH/SPFH and 100 m wind; rrfsens members lack DLWRF, SPFH 2 m and 100 m wind (80/160/320 m instead).
• The REFS ak ensemble products mask (9999) every cell outside the HRRR Alaska domain for the fields where HRRR contributes (TMP, MSLET, APCP, …; 36.4% of the grid), while UGRD 10 m is unmasked — verified cell-for-cell against the HRRR AK grid footprint.

---

Source landscape

Source	What	Coverage	Role
s3://noaa-rrfs-ops-pds (us-east-1), registry	RRFS, rrfsens, REFS ensprod, firewx; v1 layout	2026-08-12 → present, rolling	Preferred
https://nomads.ncep.noaa.gov/pub/data/nccf/com/rrfs/{para,v1.0}/, .../com/refs/{para,v1.0}/	Same files and layout	~2 days (only *.20260916/ and *.20260917/ present on 2026-09-17)	Fallback / latency cross-check
s3://noaa-rrfs-pds, registry	Prototype bucket. rrfs_a/ and rrfs_public/ are empty. Remaining: retro_output_final/{summer,winter,spring}/ (Jul 2023; 2024-01-08 → Feb 2024; 2024-05-02 → May 2024), rrfsdesi/rrfsens.20260902/ (a one-cycle sample), rrfs_retro_maps/, sample_rotlat_awips/	Three ~1-month retro windows	Not a backfill source (different product set: prslev only, no resolution token in names, no na domain)
Google Cloud / Azure	none found	—	—
ftpprd.ncep.noaa.gov	Not documented in the SCN for RRFS/REFS; returned 403 to our fetches	—	Unverified

The registry text says the prototype feed stopped updating on 2026-08-12 when the NOMADS/AWS parallel began (Herbie issue #553 reports the same). SCN 26-48 says the parallel and operational feeds share the same rrfs.YYYYMMDD/CC/... and refs.YYYYMMDD/CC/ensprod/... layout; the AWS registry says the ops bucket files "will align with those described in the Service Change Notice". Whether every product/variable stays byte-identical across the 2026-10-14 cutover cannot be verified before it happens — re-check the first operational cycles.

License (both buckets): CC0-1.0 per the NODD registry entries; NOAA requests attribution for unaltered data and prohibits implying NOAA endorsement. No LICENSE/README object exists inside either bucket (only an S3 Explorer index.html at the root of noaa-rrfs-pds). Nothing non-open found.

Documentation state: the NCO product pages (rrfs, refs) list file types but their per-file parameter inventories are still "coming soon" — every variable list below comes from the .idx files.

---

Dataset: NOAA RRFS (deterministic forecast)

Source Information

• Summary of data organization: One GRIB2 file per init time, lead hour, domain and product family, containing every variable of that family: 2dfld (surface/2-D fields, 315 messages at f001 on conus), prslev (pressure-level fields, 675 messages), 2dfld.*.subh (sub-hourly subset of 2dfld, 34 fields × 4 time steps), 2dfld_awipsubset (14-field AWIPS subset on the na grid at 3-hourly leads). Domains are separate files on separate grids: conus/ak (3 km), hi/pr (2.5 km), na (13 km), and a relocatable 1.5 km fire-weather nest firewx_lcc under its own firewx.YYYYMMDD/ prefix.

• File format: GRIB2 with a NOAA-style .idx for every file.

• Temporal coverage: rrfs.20260812/21 (first cycle with 2dfld/prslev; the 20–23Z prefixes of 2026-08-12 were uploaded on 2026-08-13 and 20Z holds only hi/pr subh files) to present. Every date from 2026-08-13 has all 24 cycle prefixes.

• Temporal frequency (verified for all 24 hours of all 37 dates from the census):

  • 00/06/12/18Z: 2dfld and prslev for all domains, hourly f000–f084 (85 files each), plus subh f001–f018 and (from 2026-08-27/28) 2dfld_awipsubset at f000, f003…f060 (3-hourly), f066…f084 (6-hourly) = 25 files.
  • 03/09/15/21Z: 2dfld and prslev for conus/ak/hi/pr only (no na 13 km, no awipsubset, no BUFR), hourly f000–f018, plus subh f001–f018.
  • All other hours (01,02,04,05,…,23Z): only 2dfld.{3km,2p5km}.subh for conus/ak/hi/pr, f001–f018. No 2dfld, prslev, na or awipsubset files exist for these cycles. (SCN 26-48's "18 h for other hourly cycles" is therefore delivered only through the sub-hourly file, and only for the subh variable set.)
  • firewx: 00/06/12/18Z only, hourly f000–f036, 2dfld and prslev on the firewx_lcc grid.
  • Leads stay hourly to f084 — checked f061/f070/f084 .idx files; no 3-hourly coarsening past f060.

• Latency (S3 LastModified via boto3, UTC):

  Cycle	f000	f001	f018	f060	f084	2dfld cycle span
  2026-09-10 00Z	+1:48:45	+1:54:22	+2:18:18	+3:02:40	+3:28:23	1:39:38
  2026-09-11 12Z	+1:51:06	+1:54:33	+2:21:20	+3:01:48	+3:32:02	1:40:56
  2026-09-12 05Z (subh only)	—	+1:20:44	+1:46:26	—	—	0:27:13

  prslev lands within 1–3 minutes of the matching 2dfld file at every lead checked.

• Access notes: anonymous S3 (--no-sign-request) and plain HTTPS. Sizes at f001 conus: 2dfld ≈ 353 MB, prslev ≈ 605 MB (about 10% / 3% smaller than the 2026-08-12 files, 391 / 627 MB); ak 2dfld ≈ 309 MB, prslev ≈ 545 MB; na 13 km 2dfld ≈ 152 MB; hi/pr 2dfld 10 / 27 MB; subh conus ≈ 143 MB. A full 00Z cycle is ~1,900 objects. .idx byte ranges make single-variable reads cheap (the 2 m temperature message is 1.7 MB).

• License: CC0-1.0 (see Source landscape).

• Browse root: https://noaa-rrfs-ops-pds.s3.amazonaws.com/index.html

• URL format:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.{YYYYMMDD}/{HH}/rrfs.t{HH}z.{2dfld|prslev}.{3km|2p5km|13km}.f{FFF}.{conus|ak|hi|pr|na}.grib2[.idx]
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.{YYYYMMDD}/{HH}/rrfs.t{HH}z.2dfld.{3km|2p5km}.subh.f{FFF}.{conus|ak|hi|pr}.grib2[.idx]
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.{YYYYMMDD}/{HH}/rrfs.t{HH}z.2dfld_awipsubset.3km.f{FFF}.na.grib2[.idx]
https://noaa-rrfs-ops-pds.s3.amazonaws.com/firewx.{YYYYMMDD}/{HH}/rrfs.t{HH}z.{2dfld|prslev}.1p5km.f{FFF}.firewx_lcc.grib2[.idx]

3km pairs with conus/ak, 2p5km with hi/pr, 13km with na. Same paths under https://nomads.ncep.noaa.gov/pub/data/nccf/com/rrfs/v1.0/.

• Example URLs:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/00/rrfs.t00z.2dfld.3km.f000.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/00/rrfs.t00z.2dfld.3km.f001.conus.grib2.idx
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/00/rrfs.t00z.prslev.3km.f084.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/03/rrfs.t03z.2dfld.3km.f018.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/05/rrfs.t05z.2dfld.3km.subh.f001.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/00/rrfs.t00z.2dfld.13km.f001.na.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260915/00/rrfs.t00z.2dfld.2p5km.f001.hi.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/firewx.20260915/00/rrfs.t00z.2dfld.1p5km.f001.firewx_lcc.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfs.20260812/21/rrfs.t21z.2dfld.3km.f001.conus.grib2   (earliest)

GRIB Index

• Index files available: Yes, for every GRIB2 object.
• Index style: NOAA (colon-separated).
• Example lines:

1:0:d=2026091500:REFC:entire atmosphere (considered as a single layer):1 hour fcst:
74:76682577:d=2026091500:TMP:2 m above ground:1 hour fcst:
97:101256626:d=2026091500:APCP:surface:0-1 hour acc fcst:

Message number in the .idx equals the rasterio band index.

Coordinate Reference System

One grid per domain; all on a 6,371,229 m sphere. Unchanged between the 2026-08-12 and 2026-09-15 files.

Domain	Projection	PROJ	Size (x × y)	Step
conus 3 km	Lambert conformal conic	+proj=lcc +lat_0=38.5 +lon_0=-97.5 +lat_1=38.5 +lat_2=38.5 +x_0=0 +y_0=0 +R=6371229 +units=m	1799 × 1059	3000 m
ak 3 km	Polar stereographic	+proj=stere +lat_0=90 +lat_ts=60 +lon_0=210 +x_0=0 +y_0=0 +R=6371229 +units=m	1649 × 1105	2976 m
hi 2.5 km	Mercator	+proj=merc +lat_ts=20 +lon_0=0 +x_0=0 +y_0=0 +R=6371229 +units=m	321 × 225	2500 m
pr 2.5 km	Mercator (same CRS as hi)	as above	544 × 310	2500 m
na 13 km	Rotated lat-lon (GRIB pole rotation)	+proj=ob_tran +o_proj=longlat +o_lon_p=0 +o_lat_p=35 +lon_0=-113 +R=6371229 (southern pole lat −35, lon −113)	1127 × 683	0.1083°
firewx_lcc	Lambert conformal conic, relocatable	+proj=lcc +lat_0=25 +lon_0=-95 +lat_1=25 +lat_2=25 +R=6371229 +units=m on 2026-09-15	522 × 390	1270 m

The conus WKT, transform (Affine(3000, 0, -2699020.142521929, 0, -3000, 1588193.847443335)) and pixel-center x/y arrays are identical to the repo's HRRR CONUS template. The firewx_lcc centre moves with the daily fire-weather target and its nominal "1.5 km" is 1270 m in the file.

Dimensions & Dimension Coordinates (conus)

Dimension	Min	Max	Step	Notes
init_time	2026-08-12T21	present	6 h (f084 cycles) / 3 h (f018 cycles) / 1 h (subh only)	see Temporal frequency
lead_time	0 h	84 h (00/06/12/18Z) or 18 h (03/09/15/21Z)	1 h	subh: 15 min steps 0:15–18:00
y	−1,587,306.15 m	1,586,693.85 m	−3000 m	1059, pixel centres, = HRRR
x	−2,697,520.14 m	2,696,479.86 m	3000 m	1799, pixel centres, = HRRR
pressure_level (prslev)	2 hPa	1000 hPa	irregular	45 levels: 2, 5, 7, 10, 20, 30, 50, 70, 100, then every 25 hPa 125–1000

Other domains (pixel centres): na grid_latitude −36.93…36.93 (step −0.1083°), grid_longitude −61.00…60.95; ak y −4,810,103…−1,524,599 m, x −2,619,396…2,285,052 m; hi y 1,920,585…2,480,585 m, x −16,878,201…−16,078,201 m; pr y 1,585,610…2,358,110 m, x −7,889,207…−6,531,707 m.

Data Variables (conus 2dfld f001; band tags from rasterio)

GDAL normalises temperatures to °C in GRIB_UNIT/values (GRIB_NORMALIZE_UNITS); the GRIB encodes K, as with HRRR in this repo.

Variable name	Level	Units	Available from	Notes
temperature_2m	2 m	K (rasterio shows [C])	2026-08-12	TMP:2 m above ground, band 74, PDTN 0
wind_u_10m / wind_v_10m	10 m	m/s	2026-08-12	bands 88/89; also WIND:10 m speed, MAXUW/MAXVW hourly max
wind_u_100m / wind_v_100m	100 m	m/s	2026-08-12	bands 219/220; also 30, 50, 80, 160, 320 m AGL and 305…4572 m AMSL
precipitation_surface	surface	kg/m²	2026-08-12	APCP: f001 has 0-1 hour acc; from f002 two messages per file, the 1-h bucket (N-1)-N hour acc and the since-init total 0-N hour acc, hourly through f084. Rasterio names the band APCP01/APCP02…
downward_short_wave_radiation_flux_surface	surface	W/m²	2026-08-12	DSWRF: two messages, (N-1)-N hour ave (PDTN 8) and instantaneous N hour fcst (PDTN 0)
downward_long_wave_radiation_flux_surface	surface	W/m²	2026-08-12	DLWRF: same two-message pattern
pressure_surface	surface	Pa	2026-08-12	PRES:surface, band 31
pressure_reduced_to_mean_sea_level	MSL	Pa	2026-08-12	MSLET only ("MSLP (Eta model reduction)"); no PRMSL message
total_cloud_cover_atmosphere	atmosphere	%	2026-08-12	TCDC:entire atmosphere instantaneous plus a 0-1 hour ave; also boundary-layer/low/mid/high
relative_humidity_2m	2 m	%	2026-08-12	band 78
specific_humidity_2m	2 m	kg/kg	2026-08-12	band 75
dew_point_temperature_2m	2 m	K ([C] in rasterio)	2026-08-12	band 76

Also in 2dfld (288 distinct var:level pairs): REFC/REFD/RETOP/VIL, CAPE/CIN (surface, 0–90/180/255 hPa AGL, 0–3 km), HLCY/UPHL/MXUPHL, GUST, VIS, CEIL, cloud base/top heights, HPBL, TMAX/TMIN/MAXRH/MINRH, ASNOW/FROZR/FRZR/WEASD/SNOD/SNOWC, PRATE, CRAIN/CSNOW/CICEP/CFRZR/CPOFP, PWAT, LTNG, ULWRF/USWRF (surface and TOA), SHTFL/LHTFL/GFLUX, soil moisture/temperature at 9 layers, smoke/dust (MASSDEN, COLMD, AOTK), synthetic satellite brightness temperatures, and 15 NCEP-local parameters GDAL cannot name. prslev (20 variables): HGT, TMP, DPT, RH, SPFH, UGRD, VGRD, ABSV, DZDT, CLMR, ICMR, RWMR, SNMR, GRLE on all 45 levels; TMP/RH/SPFH/UGRD/VGRD also on six 30 hPa-thick boundary-layer layers (30-0 … 180-150 mb above ground); PRES only on those six layers; POT, PLI, PWAT only on 30-0 mb above ground; MCONV on 30-0 mb above ground, 850 and 950 hPa; STRM on 250 and 500 hPa.

subh files (34 fields × 15/30/45/60 min): APCP, ASNOW, FROZR, FRZR, TSNOWP (since-init 0-N min acc totals only, no buckets), PRATE, TMP/DPT/SPFH 2 m, UGRD/VGRD 10 m and 80 m, GUST, PRES surface, MSLET, DSWRF/DLWRF/USWRF/ULWRF (instantaneous only), VBDSF/VDDSF, REFC/REFD/RETOP/VIL, VIS, cloud base/ceiling/top HGT, CRAIN/CSNOW/CICEP/CFRZR/CPOFP, UPHL. No RH 2 m, no TCDC, no 100 m wind.

Temporal availability changes:

• 2dfld_awipsubset.3km.*.na first appears 2026-08-27 (12Z and 18Z), and from 2026-08-28 on all four 00/06/12/18Z cycles (daily object count 11,480 → 11,580 → 11,680).
• No variable additions or removals detected between the 2026-08-12/13 and 2026-09-15/16 2dfld/prslev message lists (same message counts and core set).

Completeness (census of every rrfs.* date/hour for conus 2dfld/prslev): all 37 dates follow the cycle rule above. The only gaps in the whole archive are single leads in four recent cycles: 2026-09-14 15Z prslev f000, f002; 2026-09-14 18Z 2dfld f000, f005 and prslev f042, f062; 2026-09-15 00Z 2dfld f003 and prslev f007, f059, f075; 2026-09-15 06Z 2dfld f049 and prslev f002, f012. Other domains were spot-checked complete (85/85) for 2026-09-15 00Z; the 05Z subh sample had one missing pr lead. Treat dropped single files as routine.

Sample Files Examined

• Early archive: 2026-08-12 21Z, rrfs.20260812/21/rrfs.t21z.{2dfld,prslev}.3km.f001.conus.grib2
• Recent data: 2026-09-15 00Z conus 2dfld f000/f001 and prslev f001; .idx at f001, f002, f003, f006, f012, f018, f019, f024, f061, f070, f084; subh f001/f002/f018 (00Z and 05Z); awipsubset f000/f001
• Other domains: 2026-09-15 00Z ak, hi, pr, na 2dfld f001; firewx.20260915/00 2dfld f001; 2026-09-15 12Z ak 2dfld f001 (mask check)
• Prototype retro: s3://noaa-rrfs-pds/retro_output_final/spring/rrfs.20240502/12/rrfs.t12z.prslev.f001.hi.grib2 — same hi grid as today; only prslev files, no resolution token in the name, no na domain

Notable Observations

• Core surface variables have no declared nodata and no sentinels at f001 on conus; APCP f001/f002 min/max/mean = 0/87.4/0.14 and 0/137.7/0.29 kg/m² with no masked cells. Other bands use per-band nodata, mostly 9999 (occasionally 9.999e20): e.g. CEIL masks 72.6% of cells (no ceiling), a "not applicable" NaN. The na and firewx_lcc files report a dataset-level nodata=9999.
• Accumulated and averaged fields need to be selected by the .idx time-range string, not by element name: APCP, ASNOW, FROZR carry both a 1-h bucket and a since-init total from f002; DSWRF/DLWRF/TCDC carry both an hourly average and an instantaneous value. WEASD is a state (N hour fcst), not an accumulation.
• PRMSL is absent everywhere (RRFS, REFS, members); MSLET is the MSL pressure.
• firewx_lcc relocates daily; it is not a fixed grid.

---

Dataset: NOAA REFS (RRFS Ensemble Forecast System)

REFS is a product-generation system, not a separate model. Per SCN 26-48 and the prob fcst 0/14 tag in the files, each product is derived from 14 members: current + 6 h-old RRFS deterministic, current + 6 h-old rrfsens m001–m005, and (for conus/ak only) current + 6 h-old HRRR. Two prefixes matter: refs.* (the derived products) and rrfsens.* (the five raw members).

Source Information — refs.* ensemble products

• Summary of data organization: One GRIB2 file per init time, product type, lead hour and domain, holding every field (and every probability threshold) of that product: mean (weighted ensemble mean, 70 messages at f01), sprd (spread, 76), prob (threshold/neighborhood probabilities, 144 at f01 → 199 at f36), pmmn (probability-matched mean, 7), lpmm (localized PMM, APCP only), avrg (APCP), eas (ensemble-agreement-scale APCP/ASNOW probabilities), ffri (conus only: probability of exceeding flash-flood guidance, PPFFG, plus APCP recurrence-interval probabilities). Message counts grow with lead as more accumulation windows become available.
• File format: GRIB2 with .idx.
• Temporal coverage: refs.20260812/18 (a bootstrap cycle with 9 files) then complete from refs.20260813/00 to present.
• Temporal frequency: 00/06/12/18Z; leads f01–f60 hourly, all domains, all products (60 files per product/domain). NB: SCN 26-47 says REFS NOAAPORT/SBN products are 3-hourly f03–f48 then 6-hourly to f60 — that cadence applies to the WMO feed, not to these files.
• Latency (boto3 LastModified, UTC): f01 at +1:59:59 to +2:10:23; f36 at +2:32 to +2:57; f60 at +3:11:52 to +3:37:22 (cycles 2026-09-01 00Z, 2026-09-08 12Z, 2026-09-15 18Z). mean and prob for the same lead land within seconds of each other. Products lag member m001's f001 by 24–39 minutes.
• Access notes: anonymous S3 / HTTPS. Sizes on conus: mean 64–98 MB, sprd 64–97 MB, prob 31–72 MB, pmmn 4–10 MB, eas 0.5–13 MB, lpmm/avrg/ffri ≤ 7 MB; hi and pr files are 1–10% of that. A day is 13,920 objects.
• License: CC0-1.0.
• Browse root: https://noaa-rrfs-ops-pds.s3.amazonaws.com/index.html
• URL format:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.{YYYYMMDD}/{HH}/ensprod/refs.t{HH}z.{mean|sprd|prob|pmmn|lpmm|avrg|eas}.f{FF}.{conus|ak|hi|pr}.grib2[.idx]
https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.{YYYYMMDD}/{HH}/ensprod/refs.t{HH}z.ffri.f{FF}.conus.grib2[.idx]

(two-digit lead). Same paths under https://nomads.ncep.noaa.gov/pub/data/nccf/com/refs/v1.0/.

• Example URLs:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.20260915/12/ensprod/refs.t12z.mean.f01.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.20260915/12/ensprod/refs.t12z.mean.f01.conus.grib2.idx
https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.20260915/12/ensprod/refs.t12z.prob.f36.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.20260915/12/ensprod/refs.t12z.mean.f01.ak.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/refs.20260813/00/ensprod/refs.t00z.mean.f01.conus.grib2   (earliest complete cycle)

Source Information — rrfsens.* ensemble members

• Summary of data organization: One GRIB2 file per init time, member (m001–m005), family (2dfldnomads: 56 surface fields; prslevnomads: 8 fields × 14 pressure levels = 112 messages), domain and lead.
• File format: GRIB2 with .idx.
• Temporal coverage: rrfsens.20260909/12 to present (2026-09-09 has only 12Z and 18Z). About four weeks shorter than refs.*.
• Temporal frequency: 00/06/12/18Z; leads f000–f060 hourly (three-digit), all four domains, all five members present in every cycle.
• Latency: m001 conus 2dfldnomads f000/f001 at +1:21 to +1:45, f060 at +3:07:42 to +3:32:13 (cycles 2026-09-10 00Z, 2026-09-12 12Z, 2026-09-15 18Z). f000 sometimes lands a minute after f001.
• Access notes: sizes on conus: 2dfldnomads 43–62 MB, prslevnomads 156–188 MB per member and lead; 19,520 objects per day.
• URL format:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfsens.{YYYYMMDD}/{HH}/m{NNN}/rrfs.t{HH}z.m{NNN}.{2dfldnomads|prslevnomads}.{3km|2p5km}.f{FFF}.{conus|ak|hi|pr}.grib2[.idx]

• Example URLs:

https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfsens.20260915/12/m001/rrfs.t12z.m001.2dfldnomads.3km.f001.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfsens.20260915/12/m001/rrfs.t12z.m001.prslevnomads.3km.f001.conus.grib2
https://noaa-rrfs-ops-pds.s3.amazonaws.com/rrfsens.20260909/12/m003/rrfs.t12z.m003.2dfldnomads.2p5km.f000.hi.grib2   (earliest)

GRIB Index

• Index files available: Yes, for every object under both prefixes.
• Index style: NOAA (colon-separated) with extra trailing fields for the ensemble tag.
• Example lines:

17:18434201:d=2026091512:UGRD:10 m above ground:1 hour fcst:wt ens mean
4:2421970:d=2026091512:CAPE:90-0 mb above ground:1 hour fcst:ens spread
1:0:d=2026091512:REFC:entire atmosphere (considered as a single layer):1 hour fcst:prob >10:prob fcst 0/14:Neighborhood Probability
5:4294017:d=2026091512:APCP:surface:0-1 hour acc fcst:wt ens mean:process=193      (pmmn; lpmm uses process=200, eas process=197)
1:0:d=2026091512:REFC:entire atmosphere (considered as a single layer):1 hour fcst:ENS=+1     (rrfsens m001)

Coordinate Reference System

conus, ak, hi, pr grids are identical to the RRFS deterministic grids of the same name (verified on mean, prob and member files; conus = HRRR CONUS). No na or firewx domain in REFS.

Dimensions & Dimension Coordinates

Dimension	Min	Max	Step	Notes
init_time	2026-08-13T00 (refs) / 2026-09-09T12 (rrfsens)	present	6 h
lead_time	1 h (refs) / 0 h (rrfsens)	60 h	1 h
y, x (conus)	as RRFS		3000 m	1059 × 1799
ensemble_member (rrfsens)	m001	m005	1	the HRRR and deterministic RRFS members of REFS are not under rrfsens
pressure_level (prslevnomads)	250 hPa	1000 hPa	irregular	250, 300, 400, 500, 600, 700, 750, 800, 850, 900, 925, 950, 975, 1000

Data Variables

Variable name	Level	mean/sprd	prob	rrfsens 2dfldnomads	Notes
temperature_2m	2 m	yes (PDTN 2)	<273.15	yes (PDTN 1)	K in GRIB; rasterio shows [C]
wind_u_10m / wind_v_10m	10 m	yes	WIND speed thresholds	yes	plus WIND 10 m hourly max in members
wind_u_100m / wind_v_100m	100 m	no	no	no — 80, 160, 320 m only	mean has WIND:80 m speed only
precipitation_surface	surface	yes	thresholds e.g. >0.254, >12.7	yes	see accumulation note
downward_short_wave_radiation_flux_surface	surface	no	no	yes: 0-1 hour ave (PDTN 11) and instantaneous
downward_long_wave_radiation_flux_surface	surface	no	no	no
pressure_surface	surface	no (HGT:surface only)	no	yes
pressure_reduced_to_mean_sea_level	MSL	MSLET	no	MSLET	no PRMSL
total_cloud_cover_atmosphere	atmosphere	yes (+ low/mid/high)	no	yes (+ BL/low/mid/high, 0-1 hour ave)
relative_humidity_2m	2 m	no (RH 700 hPa only)	no	yes (+ MAXRH/MINRH)
specific_humidity_2m	2 m	no	no	no (prslevnomads only)
dew_point_temperature_2m	2 m	yes	thresholds >283.15…	yes

Also in mean/sprd: CAPE/CIN (surface, 0–90/180 hPa), HLCY, HGT/TMP/UGRD/VGRD/DPT on 250–925 hPa, PWAT, VIS, cloud base/ceiling HGT, LCDC/MCDC/HCDC, APCP/WEASD/ASNOW windows, CRAIN/CFRZR/CICEP/CSNOW, VWSH, and one unnamed NCEP-local parameter (parmcat=6 parm=202). prob adds REFC/REFD/RETOP/MAXREF, MXUPHL, MAXUVV, LTNG, CAT/ICI/FLGHT aviation fields, VIS, ceiling, wind speed, JFWPRB. Members (2dfldnomads) add REFC/RETOP, GUST, MAXUVV/MAXDVV, MXUPHL/MNUPHL, LTNG, ASNOW/FRZR, CPOFP, CAPE/CIN surface, PWAT, AOTK, CEIL, HLCY, VUCSH/VVCSH. Member prslevnomads: HGT, TMP, RH, DPT, SPFH, UGRD, VGRD, ABSV at 14 levels.

Accumulations: REFS products carry rolling windows ending at the lead, not since-init totals: APCP at f01 0-1 hour acc; f02 1-2; from f06 onward always the 1-h, 3-h and 6-h windows (5-6, 3-6, 0-6; … 59-60, 57-60, 54-60). WEASD/ASNOW follow the same pattern and ASNOW adds a 24-h window from f24. pmmn, lpmm, avrg, eas, ffri and prob use the same window families. Member files (2dfldnomads) carry only the 1-h bucket (N-1)-N hour acc for APCP/ASNOW/FRZR and 0-1 hour ave for DSWRF.

Temporal availability changes: none between refs.20260813/00 and refs.20260915/12 (identical grid, band count, order and metadata for mean f01). rrfsens exists only from 2026-09-09 12Z.

Completeness: all 140 cycles from 2026-08-13 00Z to 2026-09-16 18Z have the full 8-product × 4-domain × 60-lead set except refs.20260914/18 sprd conus f53 (absent). rrfsens: all (member, cycle) pairs complete except m002 2026-09-15 06Z f006 and m005 2026-09-15 00Z f046.

Sample Files Examined

• Early archive: refs.20260813/00/ensprod/refs.t00z.mean.f01.conus.grib2; refs.20260812/18 listing (bootstrap cycle: eas f51/f57–f60, mean f58, pmmn f58–f60, conus only)
• Recent data: refs.20260915/12 conus mean f01/f36, prob f36, ak/hi/pr mean f01; .idx for every product at f01 and f36; rrfsens.20260915/12/m001 conus 2dfldnomads f001 and prslevnomads f001; refs.20260915/00 mean/prob idx
• Mask check: refs.20260915/12 ak mean f01 TMP 2 m / UGRD 10 m / MSLET / APCP against rrfs.20260915/12 ak 2dfld f001 and hrrr.20260915/alaska/hrrr.t12z.wrfsfcf01.ak.grib2

Notable Observations

• ak products are masked outside the HRRR Alaska domain. refs.*.mean.f01.ak declares nodata=9999; TMP 2 m, MSLET and APCP are masked on 36.4% of the grid while UGRD 10 m is 0% masked. Reprojecting every REFS-ak pixel centre into the HRRR Alaska CRS (+proj=stere +lat_ts=60 +lon_0=225, 1299×919) and testing it against the HRRR bounds reproduces the mask exactly (masked ⇔ outside, 100% of cells). The deterministic RRFS ak file is unmasked for the same fields. So REFS ak ensemble statistics for HRRR-contributed fields cover only the HRRR-AK footprint; conus, hi and pr products showed no nodata. Whether the same applies to sprd/prob and to every field was not exhaustively checked.
• The prob fcst 0/14 tag confirms the 14-member composition; there is no per-member product for the HRRR or deterministic members, so a full member-level dataset would combine rrfsens (5), rrfs (1) and HRRR from noaa-hrrr-bdp-pds (1), each twice (current and 6 h-old cycle).
• PDTN by file: mean/sprd 2 (12 for accumulations), prob 5 (9 for accumulations), members 1 (11 for accumulations) with TYPE=4(Perturbed_forecast). For PDTN 8/11/12 messages rasterio reports GRIB_FORECAST_SECONDS=0; the lead is GRIB_VALID_TIME − GRIB_REF_TIME.
• aws s3 ls prints LastModified in the machine's local zone with no offset; all latencies above were taken from boto3/head-object, which return UTC.

---

Recommendations for integration

1. Start with RRFS conus on the AWS ops bucket: the grid, .idx style and read path match HRRR, so the HRRR forecast integrations are a near-direct template. Decide whether the dataset is the 00/06/12/18Z f084 product, the 3-hourly f018 product, or both; the hourly subh-only cycles are a separate, narrower product.
2. Re-verify the first operational cycles after 2026-10-14 (file set, message order, sizes) before relying on the parallel-period archive as representative; SCN 26-48 warns decoders to tolerate content-order and scaling changes.
3. For REFS, the mean/sprd/prob products are the HREF-like deliverable; the ak HRRR-footprint mask must be documented as "not applicable" NaN. A member-level ensemble dataset is possible only from 2026-09-09.
4. No historical backfill exists beyond the rolling bucket; begin archiving now if a long record matters.

Sources

• https://registry.opendata.aws/noaa-rrfs-ops/ and https://github.com/awslabs/open-data-registry/blob/main/datasets/noaa-rrfs-ops.yaml
• https://registry.opendata.aws/noaa-rrfs/ and https://github.com/awslabs/open-data-registry/blob/main/datasets/noaa-rrfs.yaml
• https://www.weather.gov/media/notification/pdf_2026/scn26-048_Updated_RRFS_and_REFS_Implementation_aad.pdf
• https://www.weather.gov/media/notification/pdf_2026/SCN26-47_Updated_Retire_NAM_SREF_HREF_HiresW_NAM_MOS.aab.pdf
• https://www.nco.ncep.noaa.gov/pmb/products/rrfs/ and https://www.nco.ncep.noaa.gov/pmb/products/refs/
• https://nomads.ncep.noaa.gov/pub/data/nccf/com/rrfs/ and https://nomads.ncep.noaa.gov/pub/data/nccf/com/refs/
• RRFS and REFS Feeds Change blaylockbk/Herbie#553 and Update RRFS/REFS to operational URL(s) blaylockbk/Herbie#555
• s3://noaa-rrfs-ops-pds, s3://noaa-rrfs-pds, s3://noaa-hrrr-bdp-pds listings and files read 2026-09-16/17

[aldenks]

Proposed dataset family

Based on the exploration above. All five start with the conus domain (y/x on the HRRR CONUS grid); other domains would be separate datasets since each is its own grid.

Dataset id	Init times	Lead times	Source files	Dims
noaa-rrfs-forecast-84-hour	00/06/12/18Z	0–84 h hourly	rrfs.* 2dfld + prslev	(init_time, lead_time, y, x)
noaa-rrfs-forecast-18-hour	every 3 h	0–18 h hourly	rrfs.* 2dfld + prslev	(init_time, lead_time, y, x)
noaa-rrfs-forecast-sub-hourly	every hour	0:15–18:00 in 15 min steps (no lead 0)	rrfs.* 2dfld.*.subh	(init_time, lead_time, y, x)
noaa-rrfs-ens-forecast	00/06/12/18Z	0–60 h hourly	rrfs.* (member 0) + rrfsens.* m001–m005	(init_time, ensemble_member, lead_time, y, x)
noaa-refs-forecast	00/06/12/18Z	1–60 h hourly	refs.*/ensprod	(init_time, statistic, lead_time, y, x) + probability variables

Notes:

• Only the deterministic ids carry a horizon token, because only they come in two cadences; rrfs-ens, refs and sub-hourly each exist in exactly one configuration.
• sub-hourly is the only hourly-init product. Its 34 fields are a subset of 2dfld (no 2 m RH, TCDC, 100 m wind), its accumulations are since-init only, and its top-of-hour messages are bit-identical to the 2dfld messages for the same cycle and lead (checked TMP 2 m, UGRD 10 m, APCP, DSWRF on 2026-09-15 00Z f001).
• rrfs-ens member 0 is the deterministic RRFS run, truncated to 60 h, mirroring GEFS's member 0 = control. SCN 26-48 treats the deterministic run as a REFS member but does not call it a control; every 2dfldnomads and prslevnomads field exists with the same level and time range in the deterministic 2dfld/prslev files (checked f006), so the six members share one variable set. These 6 are not the full 14-member REFS ensemble (which adds the 6 h-old cycle and, for conus/ak, HRRR).
• refs: mean and sprd share the state-variable set and fit a statistic dimension; prob/eas/ffri are per-(field, threshold, window) probabilities and pmmn/lpmm/avrg are precipitation-only, so those become individual variables. Precipitation in every REFS product is rolling 1/3/6 h windows ending at the lead.

Open question — how refs carries the statistic. Either a statistic dimension (mean/sprd share the state-variable set, so one array per field, .sel(statistic=...)) or no dimension, with the statistic embedded in the variable name (temperature_2m_mean, temperature_2m_spread) and expressed per variable with CF cell_methods over the realization axis — CF itself has no statistic coordinate. The table above assumes the dimension. If we keep it, the second question is labeling: the files encode mean as GRIB2 derived-forecast type 1, a weighted 14-member mean (weights not in the file), and sprd as type 4, "spread of all members", unit [spread] — not type 2, standard deviation. So mean/spread is what the files assert, while CF's cell_methods vocabulary offers mean/standard_deviation and no "spread".