Evaluating the Predictions of Horn of Africa 2020-2022 Drought by GraphCast

AI weather emulators, such as GraphCast, offer promising potential for improving weather forecasting. While GraphCast outperforms state-of-the-art numerical forecasting models in global-mean variables at the surface and lower troposphere, its skill in predicting regional extreme events remains unclear. This study evaluates GraphCast's performance in forecasting the 2020-2022 Horn of Africa (HoA) drought, the longest and most severe drought in this region on record. We generate 40-day deterministic and ensemble forecasts using the operational GraphCast model, initialized in October-November-December (OND) and March-April-May (MAM) seasons during 2020-2022. Comparing GraphCast's deterministic and ensemble forecasts to ECMWF IFS forecasts shows that GraphCast predicts the amount of precipitation and the drought condition in most regions over the HoA with higher accuracy than IFS HRES forecasts, though biases increase at longer lead times. We evaluate GraphCast’s predictions of key physical variables identified as important factors to HoA precipitations in prior studies and their correlations with the precipitations to assess GraphCast's ability to capture dynamics and teleconnections. Preliminary results indicate that GraphCast can reproduce lower-tropospheric relationships but struggles to predict upper-tropospheric connections due to significant biases in the variables at those levels. Our study highlights GraphCast's potential for predicting regional weather extremes like droughts and identifies potential issues and limitations in subseasonal predictions with the GraphCast model.