The paper focuses on optimizing Fourier neural operators (FNOs) for ocean dynamics modeling through a multiobjective hyperparameter search approach. FNOs are a data-driven model capable of simulating complex ocean behaviors.
The key highlights and insights are:
Careful selection of model hyperparameters is crucial for the performance of deep learning models in ocean modeling, but manual tuning is infeasible due to the vast search space.
The authors leverage DeepHyper's advanced search algorithms for multiobjective optimization to efficiently explore hyperparameters associated with data preprocessing, FNO architecture, and training strategies.
In addition to the commonly used mean squared error (MSE) loss, the authors propose adopting the negative anomaly correlation coefficient (ACC) as an additional loss term to improve model performance and investigate the potential trade-off between the two.
The experimental results show that the optimal set of hyperparameters enhanced model performance in single timestepping forecasting and greatly exceeded the baseline configuration in the autoregressive rollout for long-horizon forecasting up to 30 days.
The authors demonstrate that there is no trade-off between minimizing MSE and maximizing ACC, and the simple addition of negative ACC can benefit ocean modeling using FNOs.
The proposed approach offers a scalable solution with improved precision for ocean dynamics forecasting using FNOs.
Egy másik nyelvre
a forrásanyagból
arxiv.org
Mélyebb kérdések