sciCUN: A deep learning model for daily sea surface current fields inference—A case study of the Gulf of Riga

The precise forecasting of sea surface currents is essential for applications including navigation, pollution management, and ecosystem monitoring. Conventional high-resolution hydrodynamic models, such as NEMO, provide detailed short-term forecasts; however, they are computationally intensive and resource-demanding. To address these challenges, we introduce sciCUN: a deep-learning model for surface current inference using CNN-U-Net. As a case study, we used sciCUN to forecast daily current fields in the Gulf of Riga. During the training process, the model learns how the atmospheric forcing of the next day affects the fields of previous sea surface currents. sciCUN was trained from 1993 to 2019 and evaluated over a 4-year (2020–2023) prediction performance test. The results of the performance evaluations showed that somewhat less accurate predictions were mostly found in coastal regions close to river mouths and along the Baltic Sea border in the Irbe Strait, where, in contrast to hydrodynamic models, the data-driven modeling process did not apply boundary conditions. Nevertheless, sciCUN showed good predictive performance throughout its four-year testing period, achieving an average Euclidean distance of 2.30 cm/s between its prediction outputs and the original data. Furthermore, sciCUN obtained an average component-wise MAE of 1.45 cm/s and an average correlation coefficient of 0.92. sciCUN further demonstrated its ability to predict dominant daily surface current patterns through additional SOM analyses, using various clustering grid sizes to classify daily surface current maps into groups ranging from two to twelve prototypes. When the cluster size was reduced to two, focusing on the most dominant and distinctive patterns, sciCUN-predicted outputs achieved 97% accuracy in matching the correct cluster. By increasing the clustering grid size to categorize daily sea surface current maps into 12 prototypes, sciCUN still achieved 87% accuracy. Notably, most mismatches occurred between clusters whose prototypes exhibited closely resembling internal patterns. These results show that sciCUN is a computationally efficient and reliable way to emulate daily sea surface current forecasts.