Atmospheric Forcing Datasets

As of 2026-10-04. Which atmospheric datasets can force a coastal ocean model through this service or through NumericalEarth, how they compare, and which to choose.

An ocean model’s surface fluxes need, at every hour of the run: 10 m wind (eastward and northward), 2 m air temperature and specific humidity, surface pressure, precipitation, and downwelling shortwave and longwave radiation. The bulk formulae in NumericalEarth turn these into momentum, heat and freshwater fluxes; any dataset missing one of them cannot drive the model on its own.

Summary

Dataset

Type

Grid

Output step

Coverage

Latency

Path to a model

HRRR (NOAA GSL, NCEP)

Hourly-cycling forecast, 3 km convection-allowing, radar assimilation

3 km, Lambert conformal over CONUS

1 h (15 min for some fields)

CONUS; archive on AWS from 2014-07-30

About an hour after each cycle

forcingkit /api/v1/atmosphere (schema hrrr-atm-v1); the default

ERA5 (ECMWF, Copernicus C3S)

Global reanalysis (4D-Var)

0.25 degrees (about 31 km)

1 h

1940 to present

ERA5T about 5 days; final ERA5 2 to 3 months

NumericalEarth ERA5PrescribedAtmosphere

RRFS v1 (NOAA GSL, NCEP)

Hourly-cycling forecast, FV3 limited-area, 3 km

3 km, North America

1 h

Operational from 2026-10-06 (12 UTC cycle)

Hourly cycles to 18 h; to 84 h every 6 h

Not integrated; replaces NAM, SREF, HREF and HiresW, while HRRR stays in operations

HRDPS (ECCC MSC)

Forecast, GEM limited-area, 2.5 km

0.0225 degrees (about 2.5 km), rotated latitude-longitude over Canada and the northern US

1 h

Runs on the MSC Datamart for the last 30 days; no longer public archive found

00, 06, 12 and 18 UTC runs to 48 h, complete about 3 h after the run

forcingkit fetchers.hrdps.fetch_points (points only)

NAM CONUS Nest (NOAA NCEP)

Forecast, 3 km

3 km, CONUS

1 h

Retired 2026-10-06 in favour of RRFS

Four cycles a day, to 60 h

Not integrated; do not adopt

GFS (NOAA NCEP)

Global forecast, FV3

0.25 degrees

1 h to 120 h, then 3 h

Global; NODD archive on AWS

Four cycles a day, to 384 h

Not integrated; the natural extension past HRRR’s 48 h in forecast mode

ECMWF IFS open data

Global forecast

0.25 degrees

3 h, then 6 h

Global; real time

Four cycles a day, to 15 days (00 and 12 UTC)

Not integrated; CC-BY-4.0 since 2025-10-01

JRA55-do (JMA, MRI)

Reanalysis adjusted for ocean-sea-ice models

About 0.5 degrees

3 h

1958-01-01 to 2024-02-01, final version 1.6.0

Discontinued (JRA-55 ended January 2024; successor JRA-3Q)

NumericalEarth JRA55PrescribedAtmosphere (its catalogue ends 2019-12-31)

ECCO v4 (ECCO Consortium, NASA JPL)

Ocean state estimate’s adjusted forcing

About 1 degree

Monthly

Multi-decadal

Not real time

NumericalEarth ECCOPrescribedAtmosphere; climate scale only

Datasets in use

HRRR (default)

Provenance. NOAA’s High-Resolution Rapid Refresh, version 4, developed by the NOAA Global Systems Laboratory and run by NCEP: a 3 km convection-allowing model that assimilates radar every 15 minutes and starts a new forecast every hour (Dowell et al., 2022; James et al., 2022). Distributed through the NOAA Open Data Dissemination programme in the noaa-hrrr-bdp-pds bucket on AWS (us-east-1), anonymous, from 2014-07-30 to the present. NOAA open data: free to use; NOAA asks for attribution and that modified products not be presented as NOAA’s.

How this service builds it (fetchers/hrrr_atmosphere.py). Hour t is the one-hour forecast (wrfsfcf01) of the cycle that started at t - 1 h. Every record is therefore a short forecast from the latest analysis, and cycles are chained, so a run of any length is covered without reaching into the long-range end of any single forecast. Eight GRIB2 messages are read per hour by byte range from the .idx sidecar:

Message

Store variable

Treatment

UGRD, VGRD at 10 m

u10, v10 (m/s)

Rotated from the Lambert grid axes to east and north (HRRR winds are grid-relative), angle sin(38.5 deg) * (lon + 97.5 deg), checked against pyproj

TMP at 2 m

t2m (K)

SPFH at 2 m

q2m (kg/kg)

Used directly; no dewpoint conversion

PRES at the surface

sp (Pa)

Surface, not sea-level, pressure

APCP 0-1 h accumulation

prate (kg m-2 s-1)

Mean of the accumulations ending at t and t + 1 h, divided by 3600 s: a rate centred on t

DSWRF, DLWRF at the surface

dswrf, dlwrf (W m-2)

Instantaneous at t; hourly sampling resolves the diurnal cycle but aliases short cloud transients. Hourly means exist only for shortwave, in the sub-hourly files

Fields are interpolated (barycentric, weights built once per request) to a regular 0.03 degree longitude-latitude grid covering the bbox plus 0.25 degrees, which is the grid type NumericalEarth’s atmosphere regridder accepts. Records run from one hour before the run start to one hour after its end; a missing message or hour raises, and the store is published only when complete. Building one hour takes about 6.5 s on a warm connection, so a 168 h window takes about 19 minutes.

Strengths. Resolves the land-sea contrast, sea breezes and frontal timing at the scale of coastal domains from about 15 to 85 km across. Hourly radiation and precipitation. No credentials.

Limits. CONUS only. A forecast product, not a reanalysis: it carries forecast-model biases and is not homogeneous across HRRR versions (v4 since December 2020). Radiation is instantaneous.

HRRR forecasts at points

forcingkit.fetchers.hrrr_points.fetch_points(cycles, leads, points, fields) returns HRRR forecasts at caller-supplied points for any cycles and lead hours, one row per (issue time, valid time, lead, point, field, value). Keeping both times makes the record bitemporal: a forecast can be compared with what was later observed, by how far ahead it was made. Fields are 10 m wind (rotated to east and north), gust, 2 m temperature and dew point, sea-level pressure and visibility, in the units HRRR publishes. Each point takes an inverse-distance-squared mean of up to four HRRR water cells within 6 km, so a coastal point never takes a land cell’s temperature or wind.

Leads a cycle does not carry are skipped (HRRR v4 runs to 48 h at 00, 06, 12 and 18 UTC and to 18 h otherwise); a missing field in a published file raises. Three leads of seven fields at four points took 28 s on a warm connection, so one cycle’s leads 1 to 18 take about three minutes.

from forcingkit.fetchers.hrrr_points import fetch_points

df = fetch_points(["2026-10-14T08:00"], [1, 6, 18], {"buoy": (-73.58, 40.96)})

HRDPS forecasts at points

forcingkit.fetchers.hrdps.fetch_points(cycles, leads, points, fields) returns forecasts from Environment and Climate Change Canada’s High Resolution Deterministic Prediction System (HRDPS) in the same bitemporal rows as HRRR at points: 10 m wind rotated to east and north (u10, v10) and the instantaneous 10 m gust (gust), in m/s. HRDPS is a second, independent short-range model at a similar resolution (2.5 km), run by a different centre with different physics and data assimilation, so its errors at a station can be measured alongside HRRR’s.

Source. The MSC Datamart (https://dd.weather.gc.ca) publishes each 00, 06, 12 and 18 UTC run to 48 h as one GRIB2 file per variable and lead (about 2 MB each, with no byte-range index), so a run’s three fields at all 49 leads are about 300 MB. The Datamart keeps 30 days; no longer public archive of the GRIB2 files was found, so a record of HRDPS errors has to be accumulated from the day collection starts. newest_cycle() returns the newest run whose 48 h files are all posted (about 3 h after the run). base_url=hrdps.HPFX_URL reads the same files from MSC’s HPFX server, which has more bandwidth but no 24/7 support.

Grid and winds. A rotated latitude-longitude grid of 2540 x 1290 cells at 0.0225 degrees; Long Island Sound is about 350 cells inside its southern edge. Winds are relative to the rotated grid and are rotated to east and north by the angle between each cell’s grid axis and east (about 32 degrees near Long Island Sound). For the 2026-11-09 12 UTC run, lead 1, the rotated winds’ direction differed from MSC’s own 10 m wind direction field by a median of 0.07 degrees (99th percentile 0.44 degrees) over every cell with wind above 2 m/s. Points take the same inverse-distance-squared mean of up to four water cells within 6 km as HRRR, using HRDPS’s land-sea mask.

A full run (49 leads, three fields) took about 50 s at 40 points, with a peak of 490 MB: each lead keeps only the cells its points use, and at most two leads are read at once. Times that are not runs and leads past 48 h are skipped, as is a lead with any field not yet posted.

from forcingkit.fetchers import hrdps

run = hrdps.newest_cycle()
df = hrdps.fetch_points([run], range(49), {"buoy": (-73.73, 40.88)})

License. ECCC’s Data Server End-use Licence allows use and redistribution, including commercial use, with the attribution “Data Source: Environment and Climate Change Canada” and, where possible, a link to the license. It does not allow any suggestion of official status or endorsement.

ERA5 (fallback)

Provenance. ECMWF’s fifth-generation global reanalysis for the Copernicus Climate Change Service (Hersbach et al., 2020): 0.25 degree grid (about 31 km), hourly, 1940 to the present. ERA5T, the initial release, appears about five days behind real time and is overwritten by the final ERA5 two to three months later. Requires a free Copernicus Climate Data Store account (credentials in ~/.cdsapirc); Copernicus licence, attribution required.

How it is used. A model reads it through NumericalEarth’s ERA5PrescribedAtmosphere and ERA5PrescribedRadiation, for example over the bbox padded by 0.5 degrees, with linear time indexing and one hour of padding past the end. Accumulated fields (precipitation, radiation) are hour-ending means that NumericalEarth places at the centre of their hour. NumericalEarth 0.8.1’s catalogue stops at 2025-12-31; later dates need the catalogue extended until upstream rolls it.

Strengths. Homogeneous, global, long, assimilates far more observations than any forecast; the standard against which forcing biases are judged.

Limits. At 31 km a 15 km coastal domain spans one or two ERA5 cells, so the forcing is nearly uniform and smears the coast: in the first hours of 2026-04-02 over such a domain off New Jersey, the ERA5 box (which includes New Jersey land) was about 3 K warmer at 2 m and had about half HRRR’s wind speed. Latency rules out anything closer than five days to the present.

Choosing

Need

Dataset

Hindcast of a CONUS coastal domain from 2014-07-30 on

HRRR (default)

Hindcast before 2014-07-30, outside CONUS, or a reanalysis-grade comparison run

ERA5

Forecast to 48 h

HRRR forecast mode (planned): one cycle’s f01 to f48 from the 00, 06, 12 or 18 UTC cycle

Forecast beyond 48 h

HRRR to 48 h, then GFS or ECMWF IFS open data (not integrated); RRFS to 84 h once it is established in operations

Multi-decade or climate-scale forcing

ERA5, or JRA55-do (Tsujino et al., 2018) through NumericalEarth (to 2019 in its catalogue)

References

Citations

  • Dowell, D. C., C. R. Alexander, E. P. James, et al. (2022). The High-Resolution Rapid Refresh (HRRR): An hourly updating convection-allowing forecast model. Part I: Motivation and system description. Weather and Forecasting, 37(8), 1371-1395. doi:10.1175/WAF-D-21-0151.1

  • James, E. P., C. R. Alexander, D. C. Dowell, et al. (2022). The High-Resolution Rapid Refresh (HRRR): An hourly updating convection-allowing forecast model. Part II: Forecast performance. Weather and Forecasting, 37(8), 1397-1417. doi:10.1175/WAF-D-21-0130.1

  • Hersbach, H., B. Bell, P. Berrisford, et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999-2049. doi:10.1002/qj.3803

  • Tsujino, H., S. Urakawa, H. Nakano, et al. (2018). JRA-55 based surface dataset for driving ocean-sea-ice models (JRA55-do). Ocean Modelling, 130, 79-139. doi:10.1016/j.ocemod.2018.07.002