Abstract: Friction stir additive manufacturing (FSAM), as an emerging solid-state additive manufacturing technology, has demonstrated significant application potential in difficult-to-process fields such as high-temperature alloys, composite materials, and integrated fabrication of ultra-large components. The process involves complex thermo-mechanical-material multi-field coupling effects, and the relationship between process parameters and forming quality exhibits strong nonlinear characteristics, posing major challenges for process optimization and control. In recent years, deep learning has provided new research avenues for the intelligent development of the FSAM process chain due to its powerful capabilities in complex data mining and nonlinear modeling. By constructing end-to-end intelligent model systems, it enables the automatic extraction and integrated analysis of high-dimensional features from multi-source heterogeneous process data, overcoming the inherent limitations of traditional methods in modeling nonlinear relationships, dynamic process adaptability, and multi-objective collaborative optimization. Unprecedented technical pathways for enhancing process stability and precisely controlling microstructural evolution have been thereby offered. A systematic review of research progress in deep learning applications for friction stir additive manufacturing is presented. Fundamental principles and microstructural evolution mechanisms of the FSAM process are first elucidated, followed by systematic classification and organization of the multi-source data involved. Subsequently, application outcomes of deep learning are focused on across several key domains, including process monitoring and defect detection, parameter prediction and performance analysis, prediction of microstructural evolution, and process optimization strategies. Comprehensive analysis of existing research indicates that although remarkable success has been achieved by deep learning in enhancing the controllability and prediction accuracy of the FSAM process, core challenges are still faced, such as scarcity of multi-source data, insufficient model interpretability, and limited generalization ability across different process scenarios. Thus, extensive research opportunities and development potential remain for the deep integration and broader application of deep learning in the FSAM field.