Analysis-Ready, Cloud Optimized ERA5
September 10, 2025 · View on GitHub
Recipes for reproducing Analysis-Ready & Cloud Optimized (ARCO) ERA5 datasets.
Introduction • Overview • Get Started with Colab Notebooks • Analysis Ready Data • Raw Cloud Optimized Data • Project roadmap • How to reproduce • FAQs • How to cite this work • License
Introduction
Our goal is to make a global history of the climate highly accessible in the cloud. To that end, we present a curated copy of the ERA5 corpus in Google Cloud Public Datasets.
What is ERA5?
ERA5 is the fifth generation of ECMWF's Atmospheric Reanalysis. It spans atmospheric, land, and ocean variables. ERA5 is an hourly dataset with global coverage at 30km resolution (~0.28° x 0.28°), ranging from 1979 to the present. The total ERA5 dataset is about 5 petabytes in size.
Check out ECMWF's documentation on ERA5 for more.
What is a reanalysis?
A reanalysis is the "most complete picture currently possible of past weather and climate." Reanalyses are created from assimilation of a wide range of data sources via numerical weather prediction (NWP) models.
Read ECMWF's introduction to reanalysis for more.
So far, we have ingested meteorologically valuable variables for the land and atmosphere. From this, we have produced a cloud-optimized version of ERA5, in which we have converted grib data to Zarr with no other modifications. In addition, we have created "analysis-ready" versions on regular lat-lon grids, oriented towards common research & ML workflows.
This two-pronged approach for the data serves different user needs. Some researchers need full control over the interpolation of data for their analysis. Most will want a batteries-included dataset, where standard pre-processing and chunk optimization is already applied. In general, we ensure that every step in this pipeline is open and reproducible, to provide transparency in the provenance of all data.
Overview
| Location | Type | Description |
|---|---|---|
$BUCKET/ar/ | Analysis Ready | An ML-ready, unified (surface & atmospheric) version of the data in Zarr. |
$BUCKET/co/ | Cloud Optimized | A port of gaussian-gridded ERA5 data to Zarr. |
$BUCKET/raw/ | Raw Data | All raw grib & NetCDF data. |
- The
gcp-public-data-arco-era5bucket is stored in theus-central1(Iowa) Google Cloud region. - The stable version of ERA5 is updated on a monthly cadence (on roughly the 9th of each month) with a 3 month delay. The preliminary version of ERA5, known as ERA5T is available with approximately 1 week delay (where 5-6 days delay are due to processing at ECWMF).
- The most recent data available can be found by examining the metadata associated with each Zarr store. The metadata encompasses three essential attributes:
valid_time_start,valid_time_stop(for ERA5), andlast_updated. For ERA5T, usevalid_time_stop_era5tinstead. These attributes specify the start date, stop date, and most recent time of update for the dataset's data, respectively. Please note that both start and end times are inclusive, and all times are given in UTC.
Get Started with Colab Notebooks
Explore the ERA5 dataset interactively with these notebooks:
| Notebook | Description | Open in Colab |
|---|---|---|
| Surface Reanalysis Walkthrough | Learn how to work with surface-level ERA5 data |  |
| Model Levels Walkthrough | Explore atmospheric model levels data | |
Analysis Ready Data
These datasets have been regridded to a uniform 0.25° equiangular horizontal resolution to facilitate downstream analyses, e.g., with WeatherBench2.
0.25° Pressure and Surface Level Data
This dataset contains most pressure-level fields and all surface-level field regridded to a uniform 0.25° resolution. It is a superset of the data used to train GraphCast and NeuralGCM.
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/ar/full_37-1h-0p25deg-chunk-1.zarr-v3',
chunks=None,
storage_options=dict(token='anon'),
)
ar_full_37_1h = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Times:
00/to/23 - Levels:
1/2/3/5/7/10/20/30/50/70/100/125/150/175/200/225/250/300/350/400/450/500/550/600/650/700/750/775/800/825/850/875/900/925/950/975/1000 - Grid: equiangular lat-lon
- Size: 2.05 PB
- Chunking:
{'time': 1, 'latitude': 721, 'longitude': 1440, 'level': 37} - Chunk size (per variable): 154 MB
Data summary table
0.25° Model Level Data
This dataset contains 3D fields at 0.25° resolution with ERA5's native vertical coordinates (hybrid pressure/sigma coordinates).
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/ar/model-level-1h-0p25deg.zarr-v1',
chunks=None,
storage_options=dict(token='anon'),
)
ar_native_vertical_grid_data = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
It can combined with surface-level variables from the 0.25° pressure- and surface-level dataset:
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/ar/full_37-1h-0p25deg-chunk-1.zarr-v3',
chunks=None,
storage_options=dict(token='anon'),
)
ar_full_37_1h = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
ar_model_level_and_surface_data = xarray.merge([
ar_native_vertical_grid_data, ar_full_37_1h.drop_dims('level')
])
- Times:
00/to/23 - Levels:
1/to/137 - Grid: equiangular lat-lon
- Size: 5.88 PB
- Chunking:
{'time': 1, 'hybrid': 18, 'latitude': 721, 'longitude': 1440} - Chunk size (per variable): 74.8 MB
Data summary table
Raw Cloud Optimized Data
These datasets contain the raw data used to produce the Analysis Ready data. Whenever possible, parameters are represented by their native grid resolution See this ECMWF documentation for more.
Please view our walkthrough notebook for demos of these cloud-optimized datasets:
Model Level Wind
This dataset contains model-level wind fields on ERA5's native grid, as spherical harmonic coefficients.
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/co/model-level-wind.zarr-v2',
chunks=None,
storage_options=dict(token='anon'),
)
model_level_wind = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Levels:
1/to/137 - Times:
00/to/23 - Grid:
T639spherical harmonic coefficients (docs) - Size: 664 TB
- Chunking:
{'time': 1, 'hybrid': 1, 'values': 410240} - Chunk size (per variable): 1.64 MB
Data summary table
| name | short name | units | docs | config |
|---|---|---|---|---|
| vorticity (relative) | vo | s^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=138 | era5_ml_dve.cfg |
| divergence | d | s^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=155 | era5_ml_dve.cfg |
| temperature | t | K | https://apps.ecmwf.int/codes/grib/param-db?id=130 | era5_pl_hourly.cfg |
| vertical velocity | w | Pa s^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=135 | era5_ml_tw.cfg |
Model Level Moisture
This dataset contains model-level moisture fields on ERA5's native reduced Gaussian grid.
import xarray
ds = xr.open_zarr(
'gs://gcp-public-data-arco-era5/co/model-level-moisture.zarr-v2/',
chunks=None,
storage_options=dict(token='anon'),
)
model_level_moisture = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Levels:
1/to/137 - Times:
00/to/23 - Grid:
N320, a Reduced Gaussian Grid (docs) - Size: 1.54 PB
- Chunking:
{'time': 1, 'hybrid': 1, 'values': 542080} - Chunk size (per variable): 2.17 MB
Data summary table
| name | short name | units | docs | config |
|---|---|---|---|---|
| specific humidity | q | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=133 | era5_ml_o3q.cfg |
| ozone mass mixing ratio | o3 | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=203 | era5_ml_o3q.cfg |
| specific cloud liquid water content | clwc | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=246 | era5_ml_o3q.cfg |
| specific cloud ice water content | ciwc | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=247 | era5_ml_o3q.cfg |
| fraction of cloud cover | cc | (0 - 1) | https://apps.ecmwf.int/codes/grib/param-db?id=248 | era5_ml_o3q.cfg |
| specific rain water content | crwc | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=75 | era5_ml_qrqs.cfg |
| specific snow water content | cswc | kg kg^-1 | https://apps.ecmwf.int/codes/grib/param-db?id=76 | era5_ml_qrqs.cfg |
Single Level Surface
This dataset contains single-level renanalysis fields on ERA5's native grid, as spherical harmonic coefficients.
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/co/single-level-surface.zarr-v2/',
chunks=None,
storage_options=dict(token='anon'),
)
single_level_surface = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Times:
00/to/23 - Grid:
TL639spherical harmonic coefficients (docs) - Size: 2.42 TB
- Chunking:
{'time': 1, 'values': 410240} - Chunk size (per variable): 1.64 MB
Data summary table
| name | short name | units | docs | config |
|---|---|---|---|---|
| logarithm of surface pressure | lnsp | Numeric | https://apps.ecmwf.int/codes/grib/param-db?id=152 | era5_ml_lnsp.cfg |
| surface geopotential | zs | m^2 s^-2 | https://apps.ecmwf.int/codes/grib/param-db?id=162051 | era5_ml_zs.cfg |
Single Level Reanalysis
This dataset contains single-level renanalysis fields on ERA5's native reduced Gaussian grid.
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/co/single-level-reanalysis.zarr-v2',
chunks=None,
storage_options=dict(token='anon'),
)
single_level_reanalysis = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Times:
00/to/23 - Grid:
N320, a Reduced Gaussian Grid (docs) - Size: 60.9 TB
- Chunking:
{'time': 1, 'values': 542080} - Chunk size (per variable): 2.17 MB
Data summary table
Single Level Forecast
This dataset contains single-level forecast fields on ERA5's native reduced Gaussian grid.
import xarray
ds = xarray.open_zarr(
'gs://gcp-public-data-arco-era5/co/single-level-forecast.zarr-v2/',
chunks=None,
storage_options=dict(token='anon'),
)
single_level_forecasts = ds.sel(time=slice(ds.attrs['valid_time_start'], ds.attrs['valid_time_stop']))
- Times:
06:00/18:00 - Steps:
0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18 - Grid:
N320, a Reduced Gaussian Grid (docs) - Size: 53.2 TB
- Chunking:
{'time': 1, 'step': 1, 'values': 542080} - Chunk size (per variable): 2.17 MB
Data summary table
Project roadmap
Updated on 2024-06-25
- Phase 0: Ingest raw ERA5
- Phase 1: Cloud-Optimize to Zarr, without data modifications
- Use Pangeo-Forge to convert the data from grib to Zarr.
- Create example notebooks for common workflows, including regridding and variable derivation.
- Phase 2: Produce an Analysis-Ready corpus
- Update GCP CPDs documentation.
- Create walkthrough notebooks.
- Phase 3: Automatic dataset updates, data is back-fillable.
- WIP Phase 4: Mirror ERA5 data in Google BigQuery.
- Phase 5: Derive a high-resolution version of ERA5
- Regrid datasets to lat/long grids.
- Convert model levels to pressure levels (at high resolution).
- Compute derived variables.
- Expand on example notebooks.
How to reproduce
All phases of this dataset can be reproduced with scripts found here. To run them, please clone the repo and install the project.
git clone https://github.com/google-research/arco-era5.git
Or, via SSH:
git clone git@github.com:google-research/arco-era5.git
Then, install with pip:
cd arco-era5
pip install -e .
Acquire & preprocess raw data
Please consult the instructions described in raw/.
Cloud-Optimization
All our tools make use of Apache Beam, and thus are portable to any cloud (or Runner). We use GCP's Dataflow to produce this dataset. If you would like to reproduce the project this way, we recommend the following:
- Ensure you have access to a GCP project with GCS read & write access, as well as full Dataflow permissions (see these "Before you begin" instructions).
- Export the following variables:
export PROJECT=<your-gcp-project> export REGION=us-central1 export BUCKET=<your-beam-runner-bucket>
From here, we provide examples of how to run the recipes at the top of each script.
pydoc src/single-levels-to-zarr.py
pydoc src/ar-to-zarr.py
You can also discover available command line options by invoking the script with -h/--help:
python src/model-levels-to-zarr.py --help
Automating dataset Updates in zarr and BigQuery
This feature is works in 4 parts.
- Acquiring raw data from CDS, facilitated by
weather-dltool. - Splitting raw data using
weather-sp. - Ingest this splitted data into a zarr file.
How to Run.
-
Set up a Cloud project with sufficient permissions to use cloud storage (such as GCS) and a Beam runner (such as Dataflow).
Note: Other cloud systems should work too, such as S3 and Elastic Map Reduce. However, these are untested. If you experience an error here, please let us know by filing an issue.
-
Acquire one or more licenses from Copernicus.
-
Add the all
Copernicuslicenses into the secret-manager with value likes this: {"api_url": "URL", "api_key": "KEY"}NOTE: for every API_KEY there must be unique secret-key.
-
Update all of these variable in docker-file.
PROJECTREGIONBUCKETMANIFEST_LOCATIONAPI_KEY_*- In case of multiple API keys, API_KEY must follow this format:
API_KEY_*. here * can be numeric value i.e. 1, 2. - API_KEY_* value is the resource name of secret-manager key and it's value looks like this ::
projects/PROJECT_NAME/secrets/SECRET_KEY_NAME/versions/1
- In case of multiple API keys, API_KEY must follow this format:
WEATHER_TOOLS_SDK_CONTAINER_IMAGE- Is made using this dockerfile and is stored in a docker registry.
ARCO_ERA5_SDK_CONTAINER_IMAGE- Is made using this dockerfile and is stored in a registry.
-
Create docker image.
export PROJECT_ID=<your-project-here>
export REPO=<repo> eg:arco-era5-raw-to-zarr-to-bq
gcloud builds submit . --tag "gcr.io/$PROJECT_ID/$REPO:latest"
- Run script to create cloud run jobs. create_job
python deployment/create_job.py
- There will be 5 different cloud run jobs.
arco-era5-zarr-ingestion- For zarr data ingestion.arco-era5t-daily-executor- Triggers daily to process era5t-daily data.arco-era5t-monthly-executor- Triggers monthly to process era5t-monthly data.arco-era5-sanity- Sanity job to validate the data era5 vs era5t and replace in case of difference.arco-era5-executor- Triggers every month to run a sanity job for every zarr available.
- Set up cloud schedulers to trigger above jobs on specified frequencies.
arco-era5t-daily-executor- Schedule a daily trigger forera5t-dailydata.arco-era5t-monthly-executor- Schedule a monthly trigger forera5t-monthlydata.arco-era5-executor- Schedule a monthly trigger to do era5 vs era5t sanity forcurrent_month - 3month.
Making the dataset "High Resolution" & beyond...
This phase of the project is under active development! If you would like to lend a hand in any way, please check out our contributing guide.
FAQs
How did you pick these variables?
This dataset originated in Loon, Alphabet’s project to deliver internet service using stratospheric balloons, and is now curated by Google Research & Google Cloud Platform. Loon’s Planning, Simulation and Control team needed accurate data on how the stratospheric winds have behaved in the past to evaluate the effectiveness of different balloon steering algorithms over a range of weather. This led us to download the model-level data. But Loon also needed information about the atmospheric radiation to model balloon gas temperatures, so we downloaded that. And then we downloaded the most commonly used meteorological variables to support different product planning needs (RF propagation models, etc)...
Eventually, we found ourselves with a comprehensive history of weather for the world.
Where are the U/V components of wind? Where is geopotential height? Why isn’t X variable in this dataset?
We intentionally did not include many variables that can be derived from other variables. For example, U/V components of wind can be computed from divergence and vorticity; geopotential is a vertical integral of temperature.
In the second phase of our roadmap (towards "Analysis Ready" data), we aim to compute all of these variables ourselves. If you’d like to make use of these parameters sooner, please check out our example notebooks where we demo common calculations. If you notice non-derived missing data, such as surface variables, please let us know of your needs by filing an issue, and we will be happy to incorporate them into our roadmap.
Do you have plans to get all of ERA5?
We aim to support hosting data that serves general meteorological use cases, rather than aim for total completeness. Wave variables are missing from this corpus, and are a priority on our roadmap. If there is a variable or dataset that you think should be included here, please file a Github issue.
For a complete ERA5 mirror, we recommend consulting with the Pangeo Forge project (especially staged-recipes#92).
Why are there two model-level datasets and not one?
It definitely is possible for all model level data to be represented in one grid, and thus one dataset. However, we opted to preserve the native representation for variables in ECMWF's models. A handful of core model variables (wind, temperature and surface pressure) are represented as spectral harmonic coefficients , while everything else is stored on a Gaussian grid. This avoids introducing numerical error by interpolating these variables to physical space. For a more in depth review of this topic, please consult these references:
- Fundamentals of Numerical Weather Prediction by Jean Coiffier
- Atmospheric modeling, data assimilation, and predictability by Eugenia Kalnay
Please note: in a future releases, we intend to create a dataset version where all model levels are in one grid and Zarr.
Why doesn’t this project make use of Pangeo Forge Cloud?
We are big fans of the Pangeo Forge project, and of Pangeo in general. While this project does make use of their Recipes, we have a few reasons to not use their cloud. First, we would prefer to use internal rather than community resources for computations of this scale. In addition, there are several technical reasons why Pangeo Forge as it is today would not be able to handle this case (0, 1, 2, 3). To work around this, we opted to combine familiar-to-us infrastructure with Pangeo-Forge's core and to use the right tool for the right job.
Why use this dataset? What uses are there for the data?
ERA5 can be used in many applications. It can be used to train ML models that predict the impact of weather on different phenomena. ERA5 data could also be used to train and evaluate ML models that forecast the weather. The data could be used to compute climatologies, or the average weather for a region over a given period of time. ERA5 data can be used to visualize and study historical weather events, such as Hurricane Sandy.
Where should I be cautious? What are the limitations of the dataset?
 Mumbai, India |  San Francisco, USA |
 Tokyo, Japan |  Singapore |
 ERA5 Topography |  GMTED2010 Topography |
It is important to remember that a reanalysis is an estimate of what the weather was, it is not guaranteed to be an error-free estimate. There are several areas where the novice reanalysis user should be careful.
First, the user should be careful using reanalysis data at locations near coastlines. The first figure shows the fraction of land (1 for land, 0 for ocean) of ERA5 grid points at different coastal locations. This is important because the land-surface model used in ERA5 tries to blend in the influence of water with the influence of land based on this fraction. The most visible effect of this blending is that as the fraction of land decreases, the daily variation in temperature will also decrease. Looking at the first figure, there are sharp changes in the fraction of land between neighboring grid cells so there could be differences in daily temperature range that might not be reflected in actual weather observations.
The user should also be careful when using reanalysis data in areas with large variations in topography. The second figure is a plot of ERA5 topography around Mount Everest compared with GMTED2010 topography. The ERA5 topography is completely missing the high peaks of the Everest region and missing most of the structure of the mountain valleys. Topography strongly influences temperature and precipitation rate, so it is possible that ERA5’s temperature is too warm and ERA5’s precipitation patterns could be wrong as well.
ERA5’s precipitation variables aren’t directly constrained by any observations, so we strongly encourage the user to check ERA5 against observed precipitation (for example, Wu et al., 2022). A study comparing reanalyses (not including ERA5) against gridded precipitation observations showed striking differences between reanalyses and observation Lisa V Alexander et al 2020 Environ. Res. Lett. 15 055002.
Can I use the data for {research,commercial} purposes?
Yes, you can use our ERA5 data according to the terms of the Copernicus license.
Researchers, see the next section for how to cite this work.
Commercial users, please be sure to provide acknowledgement to the Copernicus Climate Change Service according to the Copernicus Licence terms.
How to cite this work
Please cite our presentation at the 22nd Conference on Artificial Intelligence for Environmental Science describing ARCO-ERA5.
Carver, Robert W, and Merose, Alex. (2023):
ARCO-ERA5: An Analysis-Ready Cloud-Optimized Reanalysis Dataset.
22nd Conf. on AI for Env. Science, Denver, CO, Amer. Meteo. Soc, 4A.1,
https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/415842
In addition, please cite the ERA5 dataset accordingly:
Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A.,
Muñoz‐Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D.,
Simmons, A., Soci, C., Abdalla, S., Abellan, X., Balsamo, G.,
Bechtold, P., Biavati, G., Bidlot, J., Bonavita, M., De Chiara, G.,
Dahlgren, P., Dee, D., Diamantakis, M., Dragani, R., Flemming, J.,
Forbes, R., Fuentes, M., Geer, A., Haimberger, L., Healy, S.,
Hogan, R.J., Hólm, E., Janisková, M., Keeley, S., Laloyaux, P.,
Lopez, P., Lupu, C., Radnoti, G., de Rosnay, P., Rozum, I., Vamborg, F.,
Villaume, S., Thépaut, J-N. (2017): Complete ERA5: Fifth generation of
ECMWF atmospheric reanalyses of the global climate. Copernicus Climate
Change Service (C3S) Data Store (CDS). (Accessed on DD-MM-YYYY)
Hersbach et al, (2017) was downloaded from the Copernicus Climate Change
Service (C3S) Climate Data Store. We thank C3S for allowing us to
redistribute the data.
The results contain modified Copernicus Climate Change Service
information 2022. Neither the European Commission nor ECMWF is
responsible for any use that may be made of the Copernicus information
or data it contains.
License
This is not an official Google product.
Copyright 2022 Google LLC
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.