CycloneNet: A Forensic Physics-Guided Deep Learning Framework for Atmospheric Singularity Mapping

CycloneNet V2 is a research release for forensic (hindcast) analysis of tropical cyclones. Rather than forecasting, it retrospectively produces FuelMap-style attribution/hypothesis maps—interpretable regions whose patterns are physically plausible correlates of historical rapid intensification (RI) —using a fully auditable artifact-driven pipeline and weak physics guidance (a physics-derived fuel prior plus diagnostic consistency constraints). The project is designed around a deliberate high-recall operating point to reduce missed RI cases. External spatial validation against independent oceanic energy proxies (e. g. , TCHP/OHC) remains future work. Included artifacts (this Zenodo record): CycloneNet A Forensic Physics-Guided Deep Learning Framework for Atmospheric Singularity Mapping. pdf — methodology, experiments, and results testₘetrics. json — ROC-AUC, PR-AUC, recall, precision, F1, Brier, etc. testₚredictions. csv — per-event probabilities and FuelMap coordinate estimates trainₕistory. json — per-epoch losses and AUC bestₐucₘodel. pt — trained PyTorch checkpoint (best AUC) bestₘodel. pt — trained PyTorch checkpoint (best overall selection) Contact: estefano. senhor@gmail. comGitHub: https: //github. com/estefano-ferreira/cyclone-net