Huge Ensembles Heatwave Forecasting over North Africa: Case Study of 3 Low-Likelihood, High-Impact Events.

Kwesi Twentwewa Quagraine, William D. Collins, Ankur Mahesh, Boris Bonev, Karthik Kashinath, Thorsten Kurth, Michael S Pritchard, Shashank Subramanian

Abstract: As the scientific community increasingly explores artificial intelligence and machine learning (AI/ML) weather models, it is essential to assess the usefulness of large ensembles of such models over data-sparse regions. This study evaluates the capability of the NVIDIA Spherical Fourier Neural Operator (SFNO) huge ensemble dataset to forecast heatwaves over North Africa, using ERA5 reanalysis as a benchmark. We focus on three extreme events observed in July 2023 in Algiers and Biskra (Algeria) and Tunis (Tunisia), addressing the question: “How well can AI/ML weather models forecast low-likelihood, high-impact heatwave events?’ Our findings show that the SFNO huge ensemble demonstrates skill in capturing the tails of the temperature distribution, an essential feature for forecasting extremes. Forecast performance could however be modulated by mesoscale to synoptic scale complexities such as land–ocean contrasts. Nevertheless, the ensemble is able to predict the selected heatwave events at lead times ranging from three to seven days. This work provides one of the first systematic evaluations of AI/ML ensemble forecasts of extreme heat in North Africa and highlights their potential for supporting decision-making in vulnerable, data-scarce regions.

External IDs:doi:10.22541/au.176599423.36287665/v1