Amidst the accelerating activation of polar cryosphere tipping points due to global warming， significant challenges must be overcome to understand their state and changes， including sparse observations， insufficient physical knowledge， and limitations of traditional model simulations. Artificial Intelligence （AI） provides a powerful tool for efficiently extracting information from vast polar datasets and bridging cognitive gaps. This paper summarizes notable progress by Chinese researchers in AI applications for the polar cryosphere： Sea ice forecasting: purely data-driven deep learning models (e.g., SICNet, SIPNet) have been developed, significantly improving weekly, monthly, and seasonal-scale forecasts of Arctic/Antarctic sea ice concentration. Some models incorporate physical constraints and outperform traditional dynamical and statistical models. Various methods (e.g., improved U-Net, EW-Net, SAC-Net, and PMDRnet) have been proposed for sea ice type identification, lead extraction, sea ice thickness relationship modeling, and enhancing the spatial resolution of passive microwave imagery. Ice sheet surface hydrology: Applied Random Forest (RF) and BP neural networks were applied to estimate surface melt of the Greenland Ice Sheet and identify supraglacial lakes. An improved U-Net model was used to automatically extract surface water bodies of the Antarctic ice sheet/ice shelf with high accuracy, thereby overcoming the limitations of traditional NDWI methods. Subglacial systems: A novel method based on Variational Autoencoders (VAE) and unsupervised clustering was used to automatically detect subglacial lakes from ice-penetrating radar data, thereby improving efficiency and accuracy. Crevass identification: Improved U-Net and its variants (e.g., ResUNet) were applied to automatically extract surface crevasse distributions on Antarctic ice shelves from SAR and optical imagery. Ice stratigraphy and topography: Deep learning (e.g., EisNet and ST-SOLOv2) was employed to automatically extract internal isochronous layers and bedrock interfaces from radargrams, aiming to solve this long-standing manual bottleneck. Other applications include: Mass balance reconstruction of covered ice sheets (fusing multi-source data with SVM/BPNN), radiation balance dataset construction (RF), near-surface air temperature inversion (RF/DNN), ice shelf basal channel identification (improved U-Net), intelligent classification of glacial seismic events (autoencoders and Gaussian mixture models), GPS data interpolation, tropospheric delay modeling, and identification of geological structures. Although Chinese polar cryosphere AI research began relatively late, it has developed rapidly and yielded fruitful results, demonstrating significant potential in data-driven modeling, automated feature extraction, and multisource information fusion. Current challenges include model interpretability, insufficient integration of physical mechanisms, scarcity of high-quality labeled data, and limited generalization ability in complex regions. Future efforts should focus on developing physically constrained AI models, advancing multimodal learning, enhancing model robustness and interpretability, and strengthening international collaboration and data-sharing to more accurately characterize polar cryosphere changes and support global climate response and risk assessment.