(EXPERIMENTAL) NOAA GraphCast Global Forecast System (GFS) (EXPERIMENTAL)

The GraphCast Global Forecast System (GraphCastGFS) is an experimental system set up by the National Centers for Environmental Prediction (NCEP) to produce medium range global forecasts. The horizontal resolution is a 0.25 degree latitude-longitude grid (about 28 km). The model runs 4 times a day at 00Z, 06Z, 12Z and 18Z cycles. Major atmospheric and surface fields including temperature, wind components, geopotential height, specific humidity, and vertical velocity, are available. The products are 6 hourly forecasts up to 10 days. The data format is GRIB2. <br/> <br/> The GraphCastGFS system is an experimental weather forecast model built upon the pre-trained Google DeepMind’s GraphCast Machine Learning Weather Prediction (MLWP) model. The GraphCast model is implemented as a message-passing graph neural network (GNN) architecture with “encoder-processor-decoder” configuration. It uses an icosahedron grid with multiscale edges and has around 37 million parameters. This model is pre-trained with ECMWF’s ERA5 reanalysis data. The GraphCastGFSl takes two model states as initial conditions (current and 6-hr previous states) from NCEP 0.25 degree GDAS analysis data and runs GraphCast (37 levels) and GraphCast_operational (13 levels) with a pre-trained model provided by GraphCast. Unit conversion to the GDAS data is conducted to match the input data required by GraphCast and to generate forecast products consistent with GFS from GraphCastGFS’ native forecast data. <br/> <br/> <strong>The GraphCastGFS version 2 made the following changes from the GraphcastCastGFS version 1.</strong> <br/><ol> <li>The 37 vertical levels model is removed due to the storage restriction and limited accuracy.<br/></li> <li>The 13 levels graphcast ML model was fine-tuned with NCEP’s GDAS data as inputs and ECMWF ERA5 data as ground truth from 20210323 to 20220901, validated from 20220901 to 20230101. Evaluation is done with forecasts from 20230101-20240101. The new weights created from the training are used to create global forecasts. It is important to note that the GraphCastGFS v1 model weights obtained from Google’s DeepMInd were provided based on 12 timesteps training with ERA5 data, while the GraphCastGFS v2 model weights resulted from training with 14 timesteps with GDAS and ERA5 data that significantly increased the accuracy of the forecasts compared with GraphCastGFS V1.<br/> <br/> The input data generated from the GDAS data as GraphCast input is provided under input/ directory. An example of file names is shown below <br/> <br/> source-gdas_date-2024022000_res-0.25_levels-13_steps-2.nc <br/> <br/> The files are under forecasts_13_levels/. There are 40 files under each directory covering a 10 day forecast. An example of file name is listed below <br/> <br/> graphcastgfs.t00z.pgrb2.0p25.f006 <br/> <br/></li> </ol> <em>The GraphCastGFS version 2.1 change log:</em> <br/> <br/><ol> <li>Starting from 06 cycle on 20240710, the forecast length is increased from 10 days to 16 days.<br/> <br/> Please note that this NOAA GraphCastGFS Model was produced using a code package released by Google DeepMind. For information on Google DeepMind, please visit their github page listed in the documentation and license sections of this page.</li> </ol>

GraphCast 全球预报系统（GraphCastGFS）是由国家环境预测中心（NCEP）设立的一项实验性系统，旨在生成中期全球预报。该系统的水平分辨率采用0.25度经纬度网格（约28公里）。模型每日运行四次，分别在00Z、06Z、12Z和18Z时次。主要大气和地表场包括温度、风速分量、位势高度、比湿和垂直速度等数据均可用。预报产品为每6小时一次，预报时长可达10天。数据格式为GRIB2。

GraphCastGFS系统是一个基于预训练的Google DeepMind GraphCast 机器学习天气预报（MLWP）模型的实验性天气预报模型。GraphCast模型采用信息传递图神经网络（GNN）架构，配置为“编码器-处理器-解码器”模式。它使用二十面体网格和多层次边，参数量约为3700万。此模型使用ECMWF的ERA5再分析数据进行预训练。GraphCastGFSl采用两个模型状态作为初始条件（当前状态和6小时前状态），从NCEP的0.25度GDAS分析数据中获取，并使用GraphCast（37层）和GraphCast操作（13层）模型进行运行，该预训练模型由GraphCast提供。对GDAS数据进行单位转换，以匹配GraphCast所需的输入数据，并生成与GraphCastGFS原生预报数据一致的预报产品。

GraphCastGFS版本2相较于GraphcastCastGFS版本1做出了以下变更。

<ul>
<li>由于存储限制和精度限制，取消了37个垂直层级模型。</li>
<li>使用NCEP的GDAS数据作为输入和ECMWF ERA5数据作为基准，从20210323至20220901期间对13层级的Graphcast ML模型进行了微调，验证期为20220901至20230101。评估基于20230101至20240101的预报。从训练中创建的新权重用于生成全球预报。值得注意的是，GraphCastGFS v1模型权重基于与ERA5数据训练的12个时间步长，而GraphCastGFS v2模型权重则是通过使用GDAS和ERA5数据训练的14个时间步长生成的，与GraphCastGFS V1相比，显著提高了预报精度。</li>
</ul>

从GDAS数据生成的作为GraphCast输入的数据提供在input/目录下。以下为文件名示例

source-gdas_date-2024022000_res-0.25_levels-13_steps-2.nc

文件位于forecasts_13_levels/目录下。每个目录下有40个文件，覆盖10天的预报。以下为文件名示例

graphcastgfs.t00z.pgrb2.0p25.f006

GraphCastGFS版本2.1变更日志：

<ul>
<li>从20240710的06时次开始，预报时长从10天增加到16天。</li>
<li>请注意，此NOAA GraphCastGFS模型使用的是Google DeepMind发布的代码包。有关Google DeepMind的信息，请访问本页面的文档和许可部分列出的github页面。</li>
</ul>