Improving long-term forecast stability of OceanCastNet: Efficient fine-tuning via a temporal change loss function

This paper introduces an efficient fine-tuning method to improve the long-term stability and accuracy of deep learning wave forecast models, particularly under data-limited conditions. By incorporating a Temporal Change Loss function into the OceanCastNet (OCN) model, we guide the training process to learn robust physical dynamics rather than merely fitting static states. A primary advantage of this approach is data efficiency. Even with the training dataset halved, the fine-tuned model not only restores performance but exceeds that of the baseline model (trained on the complete dataset) for medium- and long-range forecasts. Furthermore, the method preserves the capability to capture extreme events, as demonstrated by a case study of Super Typhoon Goni. It also effectively suppresses error accumulation during 100-day continuous forecast simulations. These results suggest that supervising the temporal evolution of physical variables is a concise and effective strategy. This approach addresses the critical challenges of data dependency and long-term drift in data-driven geoscience modeling.