Abstract

Permafrost degradation, accelerated by environmental variability, presents growing challenges for environmental stability and infrastructure resilience in polar and high-altitude regions. While traditional physics-based models have been foundational for characterizing permafrost behavior, recent advances in machine learning (ML) and deep learning (DL) have introduced scalable, data-driven alternatives capable of leveraging heterogeneous environmental datasets. This review synthesizes key developments in computational methods for predicting permafrost conditions, emphasizing model architectures, data integration strategies, and evaluation benchmarks. It surveys applications of classical ML algorithms, deep neural networks, and hybrid approaches, with particular attention to the distinction between physics-aware, physics-guided, and physics-informed strategies. Techniques for integrating satellite observations, in situ measurements, and reanalysis data are examined, alongside preprocessing workflows essential for model robustness and generalizability. Special attention is given to emerging paradigms such as physics-informed learning, interpretable AI, and ensemble modeling frameworks. The review identifies persistent challenges by systematically evaluating these methodologies, including data sparsity, cross-regional transferability, and computational scalability. It outlines research priorities to improve the operational utility of permafrost forecasting systems. The synthesis offers a foundation for advancing predictive modeling under changing environmental conditions.