---
license: mit
language:
- en
- zh
metrics:
- accuracy
tags:
- climate
---

# FourCastNet: a global data-driven high-resolution weather model

This is a global data-driven high-resolution weather model implemented, trained and open sourced by [High-Flyer AI](https://www.high-flyer.cn/en/). It is the first AI weather model, which can compare with the ECMWF Integrated Forecasting System (IFS).

See also: [Github repository](https://github.com/HFAiLab/FourCastNet) and [High-flyer AI's blog](https://www.high-flyer.cn/blog/fourcastnet/)

Typhoon track prediction:

![](https://github.com/HFAiLab/FourCastNet/blob/master/img/wind_small.gif)

Water vapour prediction:

![](https://github.com/HFAiLab/FourCastNet/blob/master/img/precipitation_small.gif)

For more cases about FourCastNet prediction, please have a look at [HF-Earth](https://www.high-flyer.cn/hf-earth/), a daily updated demo released by [High-Flyer AI](https://www.high-flyer.cn/en/).

## Inference

You can load the weights `backbone.pt` and `precipitation.pt` to generate weather predictions, as shown in the following pseudocode. The complete code is released at `./infer2img.py`.

```python
import xarray as xr
import cartopy.crs as ccrs
from afnonet import AFNONet   # download the code from https://github.com/HFAiLab/FourCastNet/blob/master/model/afnonet.py

backbone_model = AFNONet(img_size=[720, 1440], in_chans=20, out_chans=20, norm_layer=partial(nn.LayerNorm, eps=1e-6))
backbone_model.load('./backbone.pt')
precip_model = AFNONet(img_size=[720, 1440], in_chans=20, out_chans=1, norm_layer=partial(nn.LayerNorm, eps=1e-6))
precip_model.load('./precipitation.pt')

input_x = get_data('2023-01-01 00:00:00')

pred_x = backbone_model(input_x)    # input Xt, output Xt+1
pred_p = precip_model(pred_x)       # input Xt+1, output Pt+1

plot_data = xr.Dataset([pred_x, pred_p])
ax = plt.axes(projection=ccrs.PlateCarree())
plot_data.plot(ax=ax, transform=ccrs.PlateCarree(), add_colorbar=False, add_labels=False, rasterized=True)
ax.coastlines(resolution='110m')
plt.savefig('img.png')
```

FourCastNet can predict 7 surface variables, plus 5 atmospheric variables at each of 3 or 4 pressure levels, for 21 variables total. The details of these variables follow the [paper](https://arxiv.org/abs/2202.11214).


## Description of Files

`backbone.pt` 
+ the weights of backbone model, 191MB, which is trained on 20 atmospheric variables from `1979-01` to `2022-12`.

`precipitation.pt`
+ the weights of precipitation model, 187MB, which is trained on the variable `total_precipitation` from `1979-01` to `2022-12`.

`infer2img.py`
+ Case code: load the above two weights to generate images of global weather prediction.