Logo image
Atlantic tropical easterly wave and cyclogenesis forecasting in the physical and new AI forecast systems at ECMWF
Journal article   Peer reviewed

Atlantic tropical easterly wave and cyclogenesis forecasting in the physical and new AI forecast systems at ECMWF

Sharanya J. Majumdar, Linus Magnusson, Quinton A. Lawton, Rebecca Emerton, Simon T. K. Lang, Michael Maier-Gerber and David S. Richardson
Weather and forecasting
2026-04-17

Abstract

ECMWF forecasts of African easterly waves and tropical cyclogenesis in the Atlantic basin are investigated during 2020-2024, with a focus on Integrated Forecasting System (IFS) upgrades and the new Artificial Intelligence Forecasting System (AIFS). Ensemble-based probabilistic forecasts, valid at the time a tropical storm was named, exhibited high variability, with sensitivity to the maximum wind speed threshold. In many cases, the probabilities rose sharply when the lead time was reduced from 72 h to 48 h. The average probabilities have generally increased in each year, including when the IFS ensemble grid spacing was reduced from 18 km to 9 km in 2023. Across 18 developing tropical cyclones in 2024, the probabilities based on the AI-based ensemble system (“AIFS-CRPS”) often exceeded IFS probabilities for 84-120 h lead times, especially for stronger waves. In contrast, the AIFS-CRPS probabilities were lower for 36-48 h lead times, especially for the weakest waves. The average AIFS-CRPS ensemble mean position forecast error was often lower than that of the AIFS-Single, IFS deterministic, and IFS ensemble mean forecasts. The wave locations in the control (deterministic) IFS forecasts exhibited a greater southward bias than the single AIFS forecasts (“AIFS-Single”), whereas both models exhibited a slow bias in longitude. Mean absolute errors in AIFS-Single were mostly smaller for 3-5 day forecasts of 850-500 hPa wave-relative environmental vorticity and specific humidity, with lower biases in these fields. Overall, the AIFS-CRPS ensemble and AIFS-Single forecasts serve as a useful complement to the IFS.

Metrics

1 Record Views

Details

Logo image