GitHub

This is the official repository for the Pangu-Weather papers.

Accurate medium-range global weather forecasting with 3D neural networks, Nature, Volume 619, Pages 533–538, 2023.

Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast, arXiv preprint: 2211.02556, 2022.

by Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu and Qi Tian

Note: the arXiv version offers more technical details, and the Nature paper contains some new figures.

Resources including pseudocode, pre-trained models, and inference code are released here.

The slides used in a series of recent talks are attached here. Baidu Netdisk, extraction code: zjj4

中文版PPT请参见链接: 百度网盘, 提取码: 7nzb

The downloaded files shall be organized as the following hierarchy:

If you use a CPU environment, please run:

If you use a GPU environment, please first confirm that the cuda version is 11.6 and the cudnn version is the 8.2.4 for Linux and 8.5.0.96 for Windows (please see this page for details). Then, please run:

Please download the four pre-trained models (~1.1GB each) from Google drive or Baidu netdisk:

The 1-hour model (pangu_weather_1.onnx): Google drive/Baidu netdisk

The 3-hour model (pangu_weather_3.onnx): Google drive/Baidu netdisk

The 6-hour model (pangu_weather_6.onnx): Google drive/Baidu netdisk

The 24-hour model (pangu_weather_24.onnx): Google drive/Baidu netdisk

These models are stored using the ONNX format, and thus can be used via different languages such as Python, C++, C#, Java, etc.

Please prepare the input data using numpy. There are two files that shall be put under the input_data folder, namely, input_surface.npy and input_upper.npy.

input_surface.npy stores the input surface variables. It is a numpy array shaped (4,721,1440) where the first dimension represents the 4 surface variables (MSLP, U10, V10, T2M in the exact order).

input_upper.npy stores the upper-air variables. It is a numpy array shaped (5,13,721,1440) where the first dimension represents the 5 surface variables (Z, Q, T, U and V in the exact order), and the second dimension represents the 13 pressure levels (1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa in the exact order).

In both cases, the dimensions of 721 and 1440 represent the size along the latitude and longitude, where the numerical range is [90,-90] degree and [0,359.75] degree, respectively, and the spacing is 0.25 degrees. For each 721x1440 slice, the data format is exactly the same as the .nc file download from the ERA5 official website.

Note that the numpy arrays should be in single precision (.astype(np.float32)), not in double precision.

We support ERA5 initial fields and ECMWF initial fields (e.g., the initial fields of the HRES forecast), where the latter often leads to a slight accuracy drop (mainly for T2M because the two fields are quite different in temperature). A .nc file of ERA5 can be transformed into a .npy file using the netCDF4 package, and a .grib file of the ECMWF initial fields can be transformed into a .npy file using the pygrib package. Note that Z represents geopotential, not geopotential height, so a factor of 9.80665 should be multiplied if the raw data contains the geopotential height.

We temporarily do not support other kinds of initial fields due to the possibly dramatic differences in the fields when Z