Dynamic evolution mechanism driven lifetime prediction for global mesoscale eddy

Accurately predicting ocean eddy lifetimes is challenging due to their nonlinear energy dynamics and complex environmental interactions. This study utilizes a global ocean eddy trajectory dataset to analyse spatio-temporal dynamics, from energy accumulation to dissipation, identifying key lifetime factors: self energy, spatial distribution, and seasonal variability. We propose the first deep learning framework (ELPNet) designed to model the complete lifetimes of mesoscale eddies. It employs an adaptive time-series dimensionality reduction technique to present the dynamic evolution of eddies in a real-time and normalized manner. Furthermore, it integrates Long Short-Term Memory (LSTM) networks and Transformer models within a U-shaped encoder-decoder architecture to capture multi-scale spatio-temporal patterns, thereby enhancing the representation capability of eddies’ complete lifecycles. To address limitations of conventional metrics for nonlinear dissipation, we introduce the Eddy Lifetime Prediction Accuracy (ELPA) metric. With ELPA at a 30% tolerance, ELPNet achieves 71.2% prediction accuracy, doubling the performance of traditional methods. Error analysis highlights seafloor topography, energy dissipation rates, and eddy survival stages as primary accuracy constraints. This scalable approach extends to forecasting other oceanic phenomena like ocean fronts and internal waves, thereby enhancing intelligent ocean science forecasting systems.