Arctic cyclones (ACs) are powerful weather systems that influence Arctic weather and pose risks to transportation and operations in the region. Forecasting them accurately is important but remains difficult. Recently, new machine learning (ML) weather models have shown promise as faster alternatives to traditional forecast systems, but their skill in the Arctic is not well known. Meanwhile, recent study had shown that ACs can intensify through different processes, such as upper-level atmospheric forcing, diabatic heating, or strong low-level temperature gradients. This study compares two advanced ML models, GenCast and NeuralGCM, with the operational ECMWF ensemble system (IFS-ENS), focusing on their ability to forecast AC track and intensity. We test whether ML models outperform traditional forecasts, and whether forecast skill depends on how the cyclone intensifies.
Analyzing 97 ACs from 2021 to 2023, we group them into three types based on dominant intensification mechanisms. Results show no clear advantage of ML models overall. While GenCast and NeuralGCM detect slightly more cyclones, IFS-ENS more accurately predicts storm intensity, especially low-level winds. Cyclones driven by upper-level forcing are best predicted, while those with strong low-level baroclinicity remain the hardest. Our results highlight that understanding the intensification mechanism is essential for improving the forecast.