README.md

---
library_name: pytorch
tags:
  - weather-forecasting
  - earth-system
  - ensemble-forecasting
  - temporal-downscaling
  - climate
  - zarr
  - netcdf
---

# Coupled Earth-System Forecast Sample and Temporal Downscaling Model

本仓库提供一个全球海陆气冰耦合预报样例，以及配套的时间降尺度模型权重。样例包含
2020-01-04 起报的初始场和 48 个集合成员、40 个预报时次的离线结果。

This repository provides a global atmosphere-land-ocean-sea-ice coupled forecast
sample and a temporal downscaling model checkpoint. The sample is initialized on
2020-01-04 and contains 48 ensemble members with 40 forecast steps.

## Repository Structure

```text
.
├── model/
│   └── 8000_G.pth
└── data/
    ├── inp_data/
    │   └── era5_oras_20200103_20200104.zarr/
    └── oup_data/
        └── 20200104/
            ├── 0.nc
            ├── 1.nc
            ├── ...
            └── 47.nc
```

`oup_data` follows the existing artifact name and means output data.

## Artifact Relationship

The files represent different stages of the forecasting workflow:

```text
Two-day normalized initial state
  -> FuXi coupled forecast model
  -> 48-member daily coupled forecasts
  -> temporal downscaling model
  -> 6-hourly coupled forecasts
```

- `inp_data` is a normalized, model-ready initial state for the FuXi coupled
  forecast model.
- `oup_data` contains the stored daily FuXi coupled forecast results.
- `model/8000_G.pth` is the downstream temporal downscaling checkpoint. It is
  not the FuXi coupled forecast checkpoint.


## Temporal Downscaling Model

`model/8000_G.pth` is a PyTorch `OrderedDict` state dictionary with 544 tensors.
It uses a SwinIR-based temporal downscaling architecture.

| Item | Value |
| --- | --- |
| Checkpoint size | 129,715,342 bytes |
| Input channels | 358 (`175 x 2 + 8`) |
| Output channels | 700 (`175 x 4`) |
| Earth-system variables | 175 packed variables |
| Output times | 00, 06, 12, and 18 UTC |
| Embedding dimension | 180 |
| Block depths | `[6, 6, 6, 6, 6, 6]` |
| Attention heads | `[6, 6, 6, 6, 6, 6]` |
| Window size | 8 |
| MLP ratio | 2 |
| Residual connection | `1conv` |

The model consumes two packed daily fields plus eight static/time features and
predicts four 6-hourly residual fields. Architecture code and preprocessing
logic are required before loading the state dictionary into a model instance.

## Input Data

`data/inp_data/era5_oras_20200103_20200104.zarr` contains a normalized
atmosphere-land-ocean-sea-ice initial-state sample.

| Item | Value |
| --- | --- |
| Dates | 2020-01-03 and 2020-01-04 |
| Array shape | `(2, 211, 721, 1440)` |
| Dimensions | `time`, `channel`, `lat`, `lon` |
| Data type | `float16` |
| Spatial grid | Global 0.25 degrees |
| Latitude | 90 to -90 degrees |
| Longitude | 0 to 359.75 degrees |
| Compression | Blosc LZ4, level 5 |
| Stored size | Approximately 479 MB |

The channels cover pressure-level geopotential, temperature, wind and humidity;
single-level atmospheric and land variables; ocean salinity, temperature and
currents at depth; and sea-ice/ocean surface variables such as sea-ice
thickness, sea-surface height, sea-ice concentration, and mixed-layer
temperature.

Values are normalized model inputs rather than physical-unit observations.
Channel order must be preserved.

## Forecast Output

`data/oup_data/20200104` contains one NetCDF file for each ensemble member.

| Item | Value |
| --- | --- |
| Initialization time | 2020-01-04 |
| Ensemble members | 48 (`0.nc` to `47.nc`) |
| Forecast steps | 40 daily steps |
| Shape per member | `(1, 40, 211, 721, 1440)` |
| Dimensions | `time`, `step`, `channel`, `lat`, `lon` |
| Data type | `float32` |
| Spatial grid | Global 0.25 degrees |
| Approximate size | 35.05 GB per member; 1.68 TB in total |

Because the complete output is large, download only the required ensemble
members whenever possible. Repositories hosting these files should use Hugging
Face Xet storage.

## Loading the Files

Install the basic readers:

```bash
pip install torch xarray zarr netcdf4 huggingface_hub
```

Load the input sample:

```python
import xarray as xr

initial_state = xr.open_zarr(
    "data/inp_data/era5_oras_20200103_20200104.zarr"
)
print(initial_state)
```

Load one forecast member:

```python
import xarray as xr

member = xr.open_dataset("data/oup_data/20200104/0.nc")
print(member)
```

Inspect the checkpoint:

```python
import torch

state_dict = torch.load("model/8000_G.pth", map_location="cpu")
print(f"Number of tensors: {len(state_dict)}")
```

## Download from Hugging Face

Replace `YOUR_ORG/YOUR_REPO` with the published repository ID.

Download the checkpoint and input example:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YOUR_ORG/YOUR_REPO",
    allow_patterns=["model/*", "data/inp_data/*"],
    local_dir="coupled_forecast_sample",
)
```

Download one forecast member:

```python
from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="YOUR_ORG/YOUR_REPO",
    filename="data/oup_data/20200104/0.nc",
    local_dir="coupled_forecast_sample",
)
```

If this is published as a Hugging Face dataset repository, add
`repo_type="dataset"` to the download calls.

## Limitations

- This release contains one initialization date and is an example rather than a
  climatologically representative benchmark.
- The stored arrays are normalized/model-ready values and cannot be converted
  reliably to physical units without the matching normalization metadata.
- The temporal downscaling checkpoint is a state dictionary only; it requires
  the matching model definition and packing/preprocessing code.
- The FuXi checkpoint, complete inference pipeline, normalization constants,
  and static fields are not included in the current artifact set.
- The daily ensemble outputs are intermediate FuXi results, not direct outputs
  of `8000_G.pth`.

## Citation

Please replace this placeholder with the project publication before release:

```bibtex
@misc{coupled_earth_system_forecast,
  title  = {Coupled Earth-System Forecast Sample and Temporal Downscaling Model},
  author = {Project Team},
  year   = {2026}
}
```

## License

No license is declared in this model/data card. A license covering the model
weights, derived data, and upstream dependencies must be selected and added
before public release.