XAI‐GPI: An Interpretable and Adaptive Machine Learning Genesis Index for Tropical Cyclones

Abstract Tropical cyclones are extreme weather events that threaten the lives of millions of people worldwide and cause widespread damage. Accurately detecting the genesis of tropical cyclones (TCG), particularly its interannual variability across ocean basins, remains a major challenge. This work aims to improve annual TCG estimation while simultaneously identifying key formation drivers and providing explanatory insights into their influence on cyclone occurrences and interannual variability. The approach combines spatial clustering for dimensionality reduction, ensemble‐based feature selection, and interpretability via SHapley Additive exPlanations (SHAP). Based on the results of the framework, we propose a Machine Learning‐based alternative to conventional Genesis Potential Indices (GPIs) for the estimation of basin‐wide aggregated numbers of TCG: the XAI‐GPI. Unlike fixed‐form GPIs such as the Emanuel and Nolan GPI, our index identifies and explains key predictors of TCG variability directly from the data. Applied across six tropical basins, the XAI‐GPI outperforms conventional GPIs in capturing observed year‐to‐year variability while offering physical interpretability. SHAP analysis reveals how drivers such as vertical shear, relative humidity, and ENSO‐related indices modulate interannual fluctuations across the basins, while also identifying less intuitive predictors that act as proxies for large‐scale variability. This explanatory capacity provides a transparent and scalable method for advancing cyclone genesis estimation in a changing climate, positioning XAI‐GPI as a next‐generation, explainable GPI for use in both projections and attribution contexts.