CycloneNet: A Deep Learning Framework for Atmospheric Singularity Mapping via Spatio-Temporal Attention

Abstract CycloneNet is a software engineering framework with publicly available source code, released under the CC BY-NC 4.0 license. It represents a novel, hybrid approach for the forensic analysis of tropical cyclones. Its core innovation is a physics-guided deep learning system designed not to predict, but to diagnostically pinpoint—with sub-pixel spatial accuracy—the precise geographic locations of thermodynamic energy sources that fueled historical storm intensification. This record contains the following primary scientific artifacts: Technical Paper (PDF): Detailed methodology, architectural design, and forensic results. Scientific Dataset (CSV): The complete audit trail of 18 hurricanes (1989-2024), including "Target Lock" coordinates and confidence weights. Execution Log (TXT): The official validation reports and pipeline logs providing the raw output of the system's performance metrics (AUC 0.97, Recall 0.92) for independent verification. The framework functions as a forensic tool that applies physics-guided artificial intelligence to historical data, learning to identify and geolocate critical thermodynamic hotspots with a mean spatial error of approximately 26 km.