# Hongyan Zhu: Tropical Cyclone simulations: Machine learning versus physical model

**URL:** <https://forum.access-hive.org.au/t/hongyan-zhu-tropical-cyclone-simulations-machine-learning-versus-physical-model/6637>\
**Category:** Atmosphere/ML Workshop 2026 Abstracts\
**Created:** [16 July 2026 07:31 UTC](https://forum.access-hive.org.au/t/hongyan-zhu-tropical-cyclone-simulations-machine-learning-versus-physical-model/6637 "2026-07-16T07:31:10Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![lachlanswhyborn](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.access-hive.org.au/lachlanswhyborn/32/1837_2.png) [@lachlanswhyborn](https://forum.access-hive.org.au/u/lachlanswhyborn)\
**Post date:** [16 July 2026 07:31 UTC](https://forum.access-hive.org.au/t/hongyan-zhu-tropical-cyclone-simulations-machine-learning-versus-physical-model/6637/1 "2026-07-16T07:31:10Z")

</div>

Hongyan Zhu: **Tropical Cyclone simulations: Machine learning versus physical models**

**Tropical Cyclone simulations: Machine learning versus physical models**

Hongyan Zhu, Debbie Hudson, Catherine de Burgh-Day, Griffith Young, Rabi Rivett, Jim Fraser

As artificial intelligence reshapes weather prediction, a key question arises: can machine learning models match or even surpass traditional physics-based models in simulating tropical cyclones?

This talk compares the performance of the ECMWF AI Forecasting System (AIFS) and its ensemble version (AIFS-ENS) with the Bureau’s physics-based models, ACCESS-G4 and ACCESS-GE4, for tropical cyclone forecasts. We assess their ability to predict cyclone track, intensity, rainfall, and structure.

Results indicate that AIFS provides more skilful track forecasts, likely due to its stronger representation of large-scale atmospheric circulation patterns. During the 2026 tropical cyclone season, AIFS-ENS achieved track errors that were approximately 40% lower than those of ACCESS-GE4 at a lead time of 108 hours. AIFS-ENS also demonstrates enhanced skill in intensity prediction relative to ACCESS-GE4 ensemble systems. Furthermore, ML models are increasingly capable of capturing the thermodynamic structure and evolution of tropical cyclones throughout their lifecycle. These findings highlight both the growing potential and remaining limitations of AI-based forecasting, and the complementary strengths of machine learning and physical modelling approaches for next-generation tropical cyclone prediction.