Leveraging Transfer Learning on GraphCast to Improve GFS Forecasts

The Global Forecast System (GFS) developed by the National Centers for Environmental Prediction (NCEP) plays a pivotal role in generating weather forecasts and facilitating various downstream applications. The accuracy and efficiency of GFS forecasts are paramount as they directly influence the reliability and effectiveness of downstream products and services. To enhance the performance of GFS forecasts, we utilize a two-step approach applied to GraphCast, a cutting-edge neural network model designed for weather forecast applications. We first train GraphCast model using a 31-level ERA5 dataset from 1979 to 2023 and name the trained model GraphCastERA31Levs. This step aims to ensure the spatial resolution of the training data of GraphCastERA31Levs is consistent with GFS reanalysis and also optimize the performance of GraphCastERA31Levs by including recent atmospheric conditions captured by ERA5. Next, we employ transfer learning on GraphCastERA31Levs with 3-year GFS reanalysis name the final model GraphCastGFS. The transfer learning is necessary as the GFS reanalysis that include all the prognostic atmospheric variables is only available since March 22, 2021. The transfer learning can help improve the prediction accuracy of GraphCastGFS by leveraging the knowledge learned from 45-year ERA5 reanalysis. Upon completing the transfer learning process, we evaluate the efficacy of GraphCastGFS using two primary metrics. The first is the model accuracy, measured by the Root Mean Squared Error (RMSE), a widely recognized metric for forecast error. Furthermore, we assess the impact of GraphCastGFS on downstream applications, with a particular focus on its effect on the Hurricane Analysis Forecast System (HAFS). By comparing HAFS forecasts that use initial and boundary conditions from GraphCastGFS against those based on traditional GFS forecasts, we aim to determine the degree to which GraphCastGFS enhances forecast accuracy and reliability.