In recent years, the application of artificial intelligence (AI) technology in the field of biology has made rapid progress, with the most notable achievements emerging in protein structure prediction and design—recognized by the awarding of the Nobel Prize in Chemistry in 2024. It is foreseeable that the accurate prediction of various physicochemical properties of proteins will become a key direction for future development in the protein prediction field. Among these properties, protein thermodynamic stability holds great significance for advancing our understanding of life mechanisms, facilitating drug development, enabling disease diagnosis and treatment, and supporting the biotechnology industry in areas such as enzyme production, biosensor development, and protein drug formulation.The accurate prediction of protein thermodynamic stability using AI technologies can greatly enhance both research capabilities and industrial efficiency. This article reviews the development of protein thermodynamic stability prediction technologies, covering the evolution from biological experimental methods and traditional energy function-based approaches to modern machine learning techniques. The focus is on machine learning-based prediction models, particularly recent breakthroughs involving deep neural networks, graph neural networks, and attention mechanisms. Core challenges in mutation stability prediction—such as dataset quality-quantity imbalance, model overfitting, and the incorporation of protein dynamics—are discussed in depth. This paper aims to provide researchers with a comprehensive reference framework to advance the development of mutation-based protein thermodynamic stability prediction technologies.