Physics-constrained spatio-temporal generative adversarial learning for daily-scale sea ice thickness prediction

With the rapid shrinkage of Arctic sea ice, predicting its spatio-temporal changes has become an important technical basis for studies of polar climate. Existing physics-driven and data-driven paradigms show respective advantages in representing sea-ice process mechanisms and learning complex spatio-temporal patterns; however, both remain challenged in balancing prediction accuracy with physics consistency. Therefore, this study proposes a daily-scale spatio-temporal sea ice thickness (SIT) prediction method driven by both data and physics, namely the Physics-Constrained Generative Adversarial Network (PCGAN). To incorporate sea-ice physics principles into the data-driven prediction process, PCGAN is designed at three levels: the generator, the discriminators, and physics-rule constraints. First, a Physics-Augmented Cross-Process Interaction (PA-CPI) module is designed in the generator to enhance the ability of the model to represent the physically coupled mechanisms among multiple geophysics variables, including sea ice, snow, wind fields, and temperature. Second, a spatio-temporal dual-discriminator architecture is proposed to jointly constrain the local spatial structures and temporal increments of SIT. Finally, the freezing-degree-day growth relationship and the daily-scale thickness change-rate relationship are transformed into bounded physics consistency scores and embedded into temporal adversarial learning to improve the physics plausibility of the generator-predicted results. Experimental results in the Arctic Ocean during the cold season show that PCGAN maintains stable performance under different random initializations. Its average RMSE is 0.1035 m, which is 12.73% lower than that of MSS-STFormer, the best-performing baseline model. Physics-related statistical metrics, such as the anomalous thinning rate, also show that the predictions of PCGAN are closer to the observed statistical characteristics than those of the other deep learning models. Multidimensional spatio-temporal analyses further confirm the stable predictive performance of PCGAN under different prediction conditions. The interpretability analysis shows that the model’s responses to multiple geophysical factors are broadly consistent with established physical understanding of sea-ice evolution. Ablation experiments and threshold-sensitivity analyses verify the effectiveness of the key model structures and the physics consistency constraints. PCGAN demonstrates the potential of incorporating physics consistency constraints into data-driven frameworks to improve SIT prediction accuracy and physical plausibility, providing methodological support for Arctic SIT prediction during the cold season.