era5wrf/atmospair

US Mid-Atlantic ERA5 / WRF Paired Atmospheric Dataset (2007–2015)

A paired ERA5 (coarse reanalysis) and WRF (high-resolution simulation) atmospheric dataset over the US mid-Atlantic for 2007–2015. The corpus contains 26,296 timesteps at 3-hourly cadence, with 68 ERA5 channels and 49 WRF channels on a shared 112 × 112 grid. ERA5 has been pre-regridded onto the WRF mass grid so both sides share identical spatial coordinates and can be consumed channel-for-channel.

⚠️ WRF-derived channels are experimental. Eleven of the 49 WRF target channels are derived in-pipeline from the native WRF state: WPD (wind power density at seven hub-height levels), and four surface-level diagnostics — MCAPE, MCIN, SPEED10, and WINDMASK10. They are released for completeness as side-products of the WRF post-processing pipeline but have not been independently validated; use at your own risk and prefer the native fields (T2, U10, V10, U, V, TK, QVAPOR, Z) for downstream evaluation.

The dataset is released as a single consolidated zarr store. Per-channel normalization parameters (*_center and *_scale) computed over the development period and per-timestep validity masks are bundled in. A 1 GB sample subset (July 2015) is shipped alongside for quick inspection.

🌟 Highlights

• 9 years of paired data — 2007-01-01 to 2015-12-31, 3-hourly UTC, 26,296 paired timesteps.
• Shared 112 × 112 grid — ERA5 already regridded onto the WRF mass grid; no additional alignment needed.
• 117 atmospheric channels total — 68 ERA5 input + 49 WRF target, including derived wind power density (WPD).
• Self-describing zarr — variable names, vertical levels, normalization statistics, and validity masks all stored alongside the array data.
• Streamable — open directly with xr.open_zarr("hf://datasets/...", consolidated=True); chunks are pulled on demand.
• Sample subset included — one full month (July 2015, 248 timesteps, ~1 GB) for fast inspection without downloading the full corpus.

📊 Repository structure

era5wrf/atmospair/
├── atmospair.zarr.tar                # full corpus tar archive (~95 GB, 26,296 timesteps; download + untar)
├── atmospair_sample.zarr/            # July 2015 only (~1 GB, 248 timesteps; streamable)
├── metadata/                         # tabular indices for the HF Dataset Viewer
│   ├── variables_era5.jsonl          # 68 rows, ERA5 channel manifest
│   ├── variables_wrf.jsonl           # 49 rows, WRF  channel manifest
│   ├── timesteps.jsonl               # 26,296 timesteps for the full corpus (no train/eval split)
│   ├── variables_era5.md             # ERA5 channel table fragment for embedding
│   └── variables_wrf.md              # WRF  channel table fragment for embedding
├── README.md                         # this card
├── LICENSE                           # CC-BY-4.0
├── ATTRIBUTION.md                    # required ERA5 / WRF citations
├── croissant.json                    # Croissant 1.1 metadata
├── example_dataset.py             # PyTorch Dataset reference
└── example_usage.ipynb               # walkthrough notebook

🚀 Quick start

Stream the sample subset from Hugging Face

The 1 GB July-2015 sample can be opened directly without downloading:

import xarray as xr

ds = xr.open_zarr(
    "hf://datasets/era5wrf/atmospair/atmospair_sample.zarr",
    consolidated=True,
)
print(ds)

Get the full corpus (download + extract)

The full 9-year corpus ships as a single tar archive. Download it, then untar it locally:

huggingface-cli download era5wrf/atmospair atmospair.zarr.tar \
    --repo-type dataset --local-dir ./atmospair_data
tar -xf ./atmospair_data/atmospair.zarr.tar -C ./atmospair_data

You'll end up with ./atmospair_data/atmospair.zarr/ — open it the same way as the sample:

import xarray as xr
ds = xr.open_zarr("./atmospair_data/atmospair.zarr", consolidated=True)

Use the bundled PyTorch dataloader

from huggingface_hub import hf_hub_download
hf_hub_download("era5wrf/atmospair",
                "example_dataset.py", repo_type="dataset")

from example_dataset import PairedDownscalingDataset

# Full corpus (after download + extract):
ZARR = "./atmospair_data/atmospair.zarr"
# or sample (streamable, no local download):
# ZARR = "hf://datasets/era5wrf/atmospair/atmospair_sample.zarr"

train = PairedDownscalingDataset(ZARR, years=range(2007, 2014))   # normalize=True default
sample = train[0]
print(sample["era5"].shape)   # (68, 112, 112)
print(sample["wrf"].shape)    # (49, 112, 112)

📐 Data layout

The zarr store is organised into three xarray-level groupings: Dimensions (the named axes), Coordinates (label arrays attached to those axes), and Data variables (the actual data, including auxiliary arrays for normalization and validity).

Dimensions

Dimension	Size	Description
time	26,296	UTC timestamps at 3-hourly cadence
era5_channel	68	ERA5 input channel axis
wrf_channel	49	WRF target channel axis
south_north	112	Mass-grid y axis
west_east	112	Mass-grid x axis
south_north_stag	112	Staggered y axis (Arakawa-C V faces)
west_east_stag	112	Staggered x axis (Arakawa-C U faces)

Coordinates

Coordinate	Dimensions	Dtype	Chunks	Description
XLAT	(south_north, west_east)	float32	(112, 112)	Mass-grid latitude (degrees north)
XLONG	(south_north, west_east)	float32	(112, 112)	Mass-grid longitude (degrees east)
XLAT_U	(south_north, west_east_stag)	float32	(112, 112)	x-staggered latitude
XLONG_U	(south_north, west_east_stag)	float32	(112, 112)	x-staggered longitude
XLAT_V	(south_north_stag, west_east)	float32	(112, 112)	y-staggered latitude
XLONG_V	(south_north_stag, west_east)	float32	(112, 112)	y-staggered longitude
time	(time)	datetime64[ns]	(1,)	UTC timestamps
era5_channel	(era5_channel)	int64	(68,)	ERA5 channel index 0…67
era5_variable	(era5_channel)	<U26	(68,)	ERA5 variable name per channel
era5_pressure	(era5_channel)	float64	(68,)	ERA5 vertical-level identifier per channel
wrf_channel	(wrf_channel)	int64	(49,)	WRF channel index 0…48
wrf_variable	(wrf_channel)	<U26	(49,)	WRF variable name per channel
wrf_pressure	(wrf_channel)	float64	(49,)	WRF vertical-level identifier per channel

Data variables

Variable	Dimensions	Dtype	Chunks	Role
era5	(time, era5_channel, south_north, west_east)	float32	(1, 68, 112, 112)	Coarse ERA5 input fields
wrf	(time, wrf_channel, south_north, west_east)	float32	(1, 15, 56, 56)	High-resolution WRF target fields
era5_center	(era5_channel)	float32	(68,)	Per-channel normalization center
era5_scale	(era5_channel)	float32	(68,)	Per-channel normalization scale
era5_valid	(time, era5_channel)	bool	(1, 68)	Per-(time, channel) validity mask
wrf_center	(wrf_channel)	float32	(15,)	Per-channel normalization center
wrf_scale	(wrf_channel)	float32	(15,)	Per-channel normalization scale
wrf_valid	(time)	bool	(1,)	Per-timestep validity mask

Note on chunks: the Chunks column shows the on-disk chunk size. The wrf array is chunked into 15-channel × 56 × 56 blocks per timestep, while the era5 array stores one full timestep (68 channels, 112 × 112) per chunk. Chunk sizes affect read performance, not the number of values stored.

*_center and *_scale hold the per-channel normalization parameters computed over the development period (2007–2013). Use them to normalize before training and to invert the transform on predictions before computing metrics in physical units.

era5_valid and wrf_valid are boolean masks. A timestep is usable only if both sides are valid; the reference dataloader applies this filter automatically.

📋 ERA5 channels (68)

Click to expand the full ERA5 channel table

dataset	channel_index	variable	units	pressure_level	center	scale
era5	0	t	K	0	287.1305	9.5996
era5	1	t	K	1	283.5162	9.5766
era5	2	t	K	2	280.1069	8.7027
era5	3	t	K	3	273.1340	7.1026
era5	4	t	K	4	266.6559	6.6218
era5	5	t	K	5	258.2419	6.5334
era5	6	t	K	6	247.1094	6.4059
era5	7	t	K	7	232.7205	5.5636
era5	8	t	K	8	224.6013	4.7817
era5	9	t	K	9	217.8080	4.8177
era5	10	t	K	10	213.4181	5.1442
era5	11	t	K	11	210.0171	4.2864
era5	12	t	K	12	213.0829	3.1485
era5	13	u	m/s	0	1.1196	3.8207
era5	14	u	m/s	1	3.4831	6.2723
era5	15	u	m/s	2	5.9743	6.6670
era5	16	u	m/s	3	10.5453	8.2880
era5	17	u	m/s	4	13.5919	10.0540
era5	18	u	m/s	5	16.9016	12.2022
era5	19	u	m/s	6	20.7918	14.8525
era5	20	u	m/s	7	25.6838	17.9557
era5	21	u	m/s	8	28.4261	19.1515
era5	22	u	m/s	9	29.7888	18.5770
era5	23	u	m/s	10	25.9950	15.2068
era5	24	u	m/s	11	17.4025	12.2504
era5	25	u	m/s	12	4.2265	9.3743
era5	26	v	m/s	0	0.1347	4.4303
era5	27	v	m/s	1	-0.1084	6.7701
era5	28	v	m/s	2	-0.3832	7.0498
era5	29	v	m/s	3	-0.1781	7.9483
era5	30	v	m/s	4	0.2044	9.2985
era5	31	v	m/s	5	0.5442	11.1328
era5	32	v	m/s	6	1.0272	13.6729
era5	33	v	m/s	7	1.5899	16.8197
era5	34	v	m/s	8	1.4533	18.0082
era5	35	v	m/s	9	0.7237	16.9319
era5	36	v	m/s	10	0.3115	12.1493
era5	37	v	m/s	11	0.0597	7.5972
era5	38	v	m/s	12	0.0511	3.8560
era5	39	q	kg/kg	0	0.0075	0.0046
era5	40	q	kg/kg	1	0.0065	0.0040
era5	41	q	kg/kg	2	0.0054	0.0036
era5	42	q	kg/kg	3	0.0029	0.0023
era5	43	q	kg/kg	4	0.0018	0.0016
era5	44	q	kg/kg	5	0.0010	0.0009
era5	45	q	kg/kg	6	0.0005	0.0005
era5	46	q	kg/kg	7	0.0002	0.0002
era5	47	q	kg/kg	8	0.0001	0.0001
era5	48	q	kg/kg	9	0.0000	0.0000
era5	49	q	kg/kg	10	0.0000	0.0000
era5	50	q	kg/kg	11	0.0000	0.0000
era5	51	q	kg/kg	12	0.0000	0.0000
era5	52	z_pl	m^2/s^2	0	1416.0063	572.3361
era5	53	z_pl	m^2/s^2	1	7822.6069	561.1865
era5	54	z_pl	m^2/s^2	2	14686.3877	644.2152
era5	55	z_pl	m^2/s^2	3	30147.3047	971.0875
era5	56	z_pl	m^2/s^2	4	42108.3828	1238.1178
era5	57	z_pl	m^2/s^2	5	55859.3164	1552.5444
era5	58	z_pl	m^2/s^2	6	72054.3203	1936.0114
era5	59	z_pl	m^2/s^2	7	91865.4609	2381.0647
era5	60	z_pl	m^2/s^2	8	103827.8906	2577.6716
era5	61	z_pl	m^2/s^2	9	117973.2109	2651.0669
era5	62	z_pl	m^2/s^2	10	135762.6094	2463.3850
era5	63	z_pl	m^2/s^2	11	160372.3750	2079.9705
era5	64	z_pl	m^2/s^2	12	202335.7344	2018.9246
era5	65	t2m	K	—	286.9804	9.6489
era5	66	u10	m/s	—	0.9429	2.9314
era5	67	v10	m/s	—	0.0205	3.3605

ERA5 variables: t (temperature), u / v (zonal / meridional wind), q (specific humidity), z_pl (geopotential), each at 13 vertical levels, plus surface fields t2m, u10, v10. The pressure_level column holds the raw vertical-level identifier from the zarr; refer to the accompanying manuscript for the physical mapping.

📋 WRF channels (49)

Click to expand the full WRF channel table

dataset	channel_index	variable	units	pressure_level	center	scale
wrf	0	T2	K	—	284.9958	10.8910
wrf	1	U10	m/s	—	0.6777	3.2215
wrf	2	V10	m/s	—	-0.0494	3.4986
wrf	3	U	m/s	0	0.8094	3.7761
wrf	4	U	m/s	1	1.1772	4.8100
wrf	5	U	m/s	2	1.4701	5.5486
wrf	6	U	m/s	3	1.7235	6.0609
wrf	7	U	m/s	4	1.9603	6.4170
wrf	8	U	m/s	5	2.2024	6.6678
wrf	9	U	m/s	6	2.4834	6.8449
wrf	10	V	m/s	0	0.0058	4.0690
wrf	11	V	m/s	1	0.1636	5.3363
wrf	12	V	m/s	2	0.2736	6.2563
wrf	13	V	m/s	3	0.3512	6.8525
wrf	14	V	m/s	4	0.4148	7.2018
wrf	15	V	m/s	5	0.4832	7.3822
wrf	16	V	m/s	6	0.5741	7.4496
wrf	17	TK	K	0	285.2459	10.7315
wrf	18	TK	K	1	285.3293	10.5748
wrf	19	TK	K	2	285.1434	10.5474
wrf	20	TK	K	3	284.8106	10.5561
wrf	21	TK	K	4	284.3615	10.5634
wrf	22	TK	K	5	283.8329	10.5513
wrf	23	TK	K	6	283.2098	10.5048
wrf	24	QVAPOR	kg/kg	0	0.0077	0.0051
wrf	25	QVAPOR	kg/kg	1	0.0075	0.0050
wrf	26	QVAPOR	kg/kg	2	0.0073	0.0049
wrf	27	QVAPOR	kg/kg	3	0.0072	0.0048
wrf	28	QVAPOR	kg/kg	4	0.0070	0.0047
wrf	29	QVAPOR	kg/kg	5	0.0069	0.0046
wrf	30	QVAPOR	kg/kg	6	0.0067	0.0045
wrf	31	Z	m	0	1567.9675	1836.4984
wrf	32	Z	m	1	2104.7300	1833.1614
wrf	33	Z	m	2	2739.8284	1829.5360
wrf	34	Z	m	3	3472.8694	1825.8800
wrf	35	Z	m	4	4303.5020	1822.4232
wrf	36	Z	m	5	5231.9858	1819.4709
wrf	37	Z	m	6	6306.7202	1817.2397
wrf	38	MCAPE	J/kg	—	424.7575	574.7105
wrf	39	MCIN	J/kg	—	42.9617	59.6902
wrf	40	SPEED10	m/s	—	3.8272	2.9039
wrf	41	WINDMASK10	—	—	0.0000	1.0000
wrf	42	WPD	W/m^2	0	172.0187	362.2594
wrf	43	WPD	W/m^2	1	324.6923	488.1293
wrf	44	WPD	W/m^2	2	505.4665	657.0610
wrf	45	WPD	W/m^2	3	670.8135	875.4647
wrf	46	WPD	W/m^2	4	800.6617	1095.5264
wrf	47	WPD	W/m^2	5	895.9377	1293.6201
wrf	48	WPD	W/m^2	6	964.7072	1466.2845

WRF variables: native fields — surface T2, U10, V10 and upper-air U, V, TK, QVAPOR, Z at 7 levels each (38 channels); plus 11 experimental WRF-derived diagnostics — WPD at 7 hub-height levels, and the surface diagnostics MCAPE, MCIN, SPEED10, WINDMASK10 — see the disclaimer at the top of this card.

📑 Suggested temporal split

Split	Years	Use
Training	2007–2013	Model training
Evaluation	2014–2015	Held-out evaluation

The metadata/timesteps.jsonl file carries a split column reflecting this convention, so it can be filtered directly without recomputing.

💻 Usage example

import torch
from torch.utils.data import DataLoader
from example_dataset import PairedDownscalingDataset

ZARR = ("hf://datasets/era5wrf/atmospair/"
        "atmospair.zarr")

train = PairedDownscalingDataset(ZARR, years=range(2007, 2014))
loader = DataLoader(train, batch_size=8, shuffle=True, num_workers=4)

for batch in loader:
    era5 = batch["era5"]                # (B, 68, 112, 112), normalized
    wrf  = batch["wrf"]                 # (B, 49, 112, 112), normalized
    # ... model.forward(era5), loss against wrf, etc.

# Inverting the normalization on a model output:
pred_physical = train.denormalize_wrf(model_output)   # (B, 49, 112, 112)

🗺 Visualization

import xarray as xr
import matplotlib.pyplot as plt

ds = xr.open_zarr(
    "hf://datasets/era5wrf/atmospair/"
    "atmospair_sample.zarr",
    consolidated=True,
)

t = ds.sizes["time"] // 2
fig, axes = plt.subplots(1, 2, figsize=(10, 4))
axes[0].imshow(ds.era5.isel(time=t, era5_channel=0).values, origin="lower")
axes[0].set_title(f"ERA5 {ds.era5_variable.values[0]}")
axes[1].imshow(ds.wrf.isel(time=t, wrf_channel=0).values, origin="lower")
axes[1].set_title(f"WRF {ds.wrf_variable.values[0]}")
plt.tight_layout(); plt.show()

⚠️ Limitations

The full set of Responsible AI fields (limitations, biases, social impact, provenance) is in croissant.json. In brief:

• Single region. US mid-Atlantic only; transfer to other regions is not demonstrated.
• WRF-derived target. The high-resolution side is model output, not observation; it carries the biases of the WRF physics suite and boundary conditions used to produce it.
• ERA5 limitations apply on the input side, particularly for precipitation, convective regimes, and very-near-surface fields.
• Short window. 9 years captures interannual variability but is insufficient for multi-decadal climate-trend claims.
• 3-hourly cadence misses sub-hourly extremes such as convective gusts.
• No personal or sensitive information. Physical atmospheric variables only.

🚫 Out-of-scope uses

• Real-time operational weather forecasting (this is a static historical corpus).
• Climate trend attribution at multi-decadal scales.
• Treating WRF fields as ground truth.
• Energy-siting or infrastructure decisions without on-site measurement validation.

📝 Citation

@misc{atmospair_2026,
  title        = {A Downscaling Dataset for Multi-level Atmospheric Fields},
  author       = {Anonymous Author(s)},
  year         = {2026},
  howpublished = {Hugging Face dataset, era5wrf/atmospair},
  note         = {NeurIPS 2026 Evaluation and Dataset Track (under review)}
}

See ATTRIBUTION.md for the required ERA5 (Copernicus / ECMWF) and WRF citations.

🔧 Maintenance

• Maintainer: Anonymous Author(s) (review period)
• Contact: anonymized for review
• Versioning: this release is frozen for reproducibility. Corrections are issued as a new tagged version; old versions remain downloadable.