Sea Ice Concentration Prediction for Arctic Shipping: A Transformer-Based Approach with Latent Uncertainty Modeling

Arctic shipping routes offer substantial and economic benefits, yet their operational reliability is often hindered by the unpredictable and dynamic nature of Arctic environments. To address this, we develop the CryoFormer, which is a Transformer-based model for predicting sea ice concentration with enhanced accuracy and uncertainty awareness. CryoFormer integrates a deterministic stage-leveraging dartboard spatial self-attention and causal temporal self-attention to efficiently capture local and sequential dependencies-and a stochastic stage that introduces latent variables via variational inference to model environmental and sensor uncertainties. Compared to LSTM and STGNN, CryoFormer achieves up to 24.8% lower MAE and 6.4% lower RMSE on the NOAA NSIDC dataset, demonstrating superior spatiotemporal modeling and uncertainty quantification. These improvements are essential for safer Arctic navigation. Future work will focus on incorporating multivariate climate inputs and enabling real-time deployment in operational systems.