The East Coast Community Ocean Forecast System (ECCOFS) is a data assimilating ocean analysis and forecast system being developed by Rutgers University, the University of California Santa Cruz, Fathom Science Inc., and the National Ocean Service (NOS) of NOAA for transition to operations at NCEP in 2028. The ECCOFS domain spans the eastern seaboard of North America and Intra-Americas Seas from the Grand Banks of Newfoundland in the north to the mouth of the Orinoco River, Venezuela, in the south. ECCOFS will complement the existing WCOFS (West Coast Operational Forecast System) to achieve complete forecast coverage of U.S. territorial seas adjacent to the 48 contiguous states and Puerto Rico. Each day ECCOFS generates an ocean analysis using the Regional Ocean Modeling System (ROMS) 4-Dimensional Variational (4D-Var) data assimilation (DA) system constrained by ocean observations. The forecast resolution is 3 km in the horizontal at 50 vertical terrain-following levels.
🌐 View data in browser · 💻 Data access (code) · 📦 Data access (NOAA, AWS Open Data)
Daily output of the NOAA NOS ECCOFS forecast model
(a ROMS configuration of the East Coast Coastal Ocean Forecast System), published as
Icechunk stores on
Source Cooperative so that months of daily 3-D fields
open in about a second as one lazy xarray datacube.
Both stores are virtual: each holds only Zarr metadata and byte-range references to the NetCDF files in
NOAA's public bucket s3://noaa-nos-eccofs-pds (us-east-1). Nothing is copied, and both are kept up to date
automatically (see How it stays up to date).
The links use the gridlook build published beside the HYCOM store, which
reads any store named in the URL. dimIndices_s_rho=49 is the surface level (the last of 50, since ROMS counts
from the bottom up) and dimIndices_time=160 is 2026-11-16; use the sliders to change both.
The data arrays are read from NOAA's bucket, which sends Access-Control-Allow-Origin: * and allows Range
requests (checked 2026-11-19), so no CORS workaround should be needed.
Required: Python ≥ 3.12, icechunk ≥ 2.2. No credentials are needed. On older Python, pip install icechunk
quietly installs icechunk 1.x, which will not work.
For the forecast-run store use the eccofs-avg-fmrc URL; its variables have dimensions
(time, step, ...) and the best-estimate series can be rebuilt from it with
Rolodex. A complete example with maps and a point time series is in
eccofs_avg_best_series_surface_temp.ipynb.
Reads are by whole chunk. One chunk is one vertical level of one 145 × 167 tile, so a surface map reads about 100 chunks (roughly 10 MB) and not the whole water column. The first surface map from a laptop took a few seconds; from a cloud machine in us-east-1 it takes about half a second.
Each ECCOFS run is a short analysis followed by a forecast, so the same day appears in several runs. The best
series takes each valid time from one run, following the modeler's recommendation for avg:
_0002 file, lead day 1) for all past times, which avoids
start-of-cycle transients, and_0003 to _0008) as the forecast, so the series always ends about a week
after the newest run.If a run is missing (a few days, such as 2026-06-15 and 2026-06-20, are absent from the bucket), the
next most recent run that covers that day is used. Choosing the files is done with Rolodex
BestEstimate(offset=1) on the forecast-run store, so the two always agree. The root attribute sources
lists the source file behind every time, and latest_run is the newest run reflected.
The last six times are a forecast, not an analysis, and each new run replaces them. Past times do not change. Old snapshots are expired after a week, so Icechunk time travel on the live stores only reaches back that far.
Spatial chunks (145 × 167, a 10 × 10 tiling) are inherited from NOAA's files and cannot be changed in a virtual store. The files are not compressed, so each vertical level inside a source chunk is a contiguous byte range, and the stores reference one level per chunk. In one measurement from a laptop, a surface map moved about 30 MB over the network where whole 50-level chunks moved about 530 MB.
Two GitHub Actions workflows run daily at 15:30 UTC:
eccofs-avg-fmrc.eccofs-avg-best up to date: it compares the source
file of each stored time with the current best series, cuts the time axis back to the first time that
changed, and appends references for the changed files only (about a week of files per run, one commit).Both sign in to Source Cooperative with GitHub's OIDC token, so the repository holds no stored credentials.
scripts/build_avg.py and scripts/build_avg_best.py do the work and can be run locally; --dry-run shows
what would change. The his/ product and the other prefixes of the bucket (qck, sta, frc, obs, riv)
are not included yet.
The references are made with VirtualiZarr, using the
subchunk-first-nontrivial-axis branch of a fork
that splits a chunk along the first non-trivial axis (here, the vertical). The scripts also build by hand the
one level of the 51-level s_w variables that the fork leaves in a trailing partial chunk. Both stores were
checked against the source files by comparing values read through the store with values read directly.
This work is released under Apache-2.0. The stores do not contain the ECCOFS data, only references to NOAA's files, so please follow NOAA's terms for the data itself and acknowledge NOAA NOS and the ECCOFS developers.
If you use the stores or scripts in published work, please cite:
Signell, R. (2026). NOAA ECCOFS as virtual Icechunk stores. ocean-icechunks/noaa-eccofs. https://github.com/rsignell/eccofs
avg files to use, from John Wilkin.