Abstract: Mesoscale eddies are ubiquitous ocean dynamical structures, but the credibility of AI-based eddy forecasts cannot be judged by a single error metric alone. This paper proposes a multi-dimensional evaluation framework consisting of physical consistency checking, spatiotemporal evolution matching, and core parameter accuracy quantification, with quantitative metrics, pass thresholds, and a decision procedure. The framework is applied to an AI-based mesoscale eddy forecast model for the northwestern Pacific Ocean region east of the South China Sea, using daily sea level anomaly data from 2000 to 2018. The model ingests 14 days of observations and produces 14-day forecasts, and it is evaluated on ten independent 28-day test sequences. The mean 14-day sea surface height anomaly RMSE is 3.29 cm; all 55 effective truth eddies are successfully matched. The framework balances numerical accuracy and physical plausibility, supporting the development, validation, and operational use of AI-based mesoscale eddy forecasting systems.

Key words: Mesoscale eddy, AI prediction, Evaluation method, Physical consistency, Spatiotemporal evolution