Predictive Tropical Cyclogenesis Using Deep Learning: A Novel Approach to Early Warning Systems

Author

• Nikolay Todorov, Chapman UniversityFollow

Date of Award

Fall 12-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computational and Data Sciences

First Advisor

Cyril Rakovski

Second Advisor

Arin Ghazarian

Third Advisor

Jochen Albrecht

Abstract

Tropical cyclogenesis (TCG) is among the least predictable stages of the tropical cyclone life cycle, yet the timing and location of genesis are exactly what early warning systems (EWS) depend on. We present a spatio-temporal deep learning framework that forecasts genesis as a normalized probability field and delivers that skill through a real-time warning pipeline. The model couples a two-layer ConvLSTM encoder with learned temporal-attention pooling over nine 6-hourly steps (0–48 h lead) and a convolutional refinement head that outputs per-cell genesis probability. It is trained on 1980–2025 ERA5 fields and Emanuel potential intensity at 0.25° and evaluated by leave-one-year-out cross-validation against IBTrACS genesis locations across two contrasting basins, the North Atlantic/Caribbean/Gulf of Mexico (267 storms) and the Western Pacific. A novel objective drives spatial sharpness: positive cases minimize SoftMax-KL divergence against σ-widened Gaussian targets over ocean cells, while non-genesis snapshots incur a log-sum-exp peakedness penalty. In the North Atlantic the model localizes genesis to a median of 115 km at genesis time, with 69% of storms within 200 km and 81% within 500 km, and skill decays smoothly toward a basin-shared predictability floor by 48 h. The compact Atlantic domain achieves sharp near-genesis localization while the larger basin shows the same qualitative decay, indicating the model extracts real near-genesis environmental structure rather than domain artifacts. Crucially, the trained model drives a live EWS web application that converts probabilistic output into genesis-probability alerts in real time, demonstrating an end-to-end path from probabilistic prediction to decision-ready warning, with particular value for under-observed, high-vulnerability basins.

Creative Commons License

This work is licensed under a Creative Commons Attribution-Share Alike 4.0 License.

Recommended Citation

N. Todorov, "Predictive tropical cyclogenesis using deep learning: A novel approach to early warning systems," Ph.D. dissertation, Chapman University, Orange, CA, 2026. https://doi.org/10.36837/chapman.000774