UTRI-Net: Universal Rapid Intensification Forecasting via Multi-Scale Temporal Features

We present UTRI-Net, a machine learning framework for predicting tropical cyclone rapid intensification (RI) that achieves robust cross-basin generalization. By eliminating geographic coordinates and incorporating multi-scale temporal features (6–12 hour windows), our approach captures universal physical processes instead of location-specific patterns. Under rigorous cross-basin and out-of-time testing, UTRI-Net demonstrates strong discriminative skill: cross-basin validation yields AUC = 0.9149 (Atlantic → Western Pacific) and AUC = 0.9347 (Western Pacific → Atlantic), with an average cross-basin AUC = 0.9248. A strict temporal split (1995–2015 → 2016–2024) achieves AUC = 0.9392 and MCC = 0.3308. When trained on Atlantic+Western Pacific and evaluated on an independent Eastern Pacific set, the model attains AUC = 0.9629 with a confusion matrix of [[19126, 1190], [64, 246]] at a 0.5 threshold. Additionally, inter-hemispheric validation on the Indian Ocean (zero-shot) yields AUC = 0.8163, demonstrating true universality. These results substantially outperform a climatological baseline (AUC = 0.5000) and indicate that RI predictability benefits more from dynamic temporal evolution than from static geographic climatology.