Machine learning in friction stir processing: Advances in process optimisation and property prediction

Friction stir welding (FSW) and friction stir processing (FSP) use frictional heat and intense plastic deformation from a rotating tool to join or modify materials without melting them. Joint strength and microstructure depend on many interacting variables, including rotational and traverse speeds, axial force, tool design, tilt angle, number of passes and, for surface composites, reinforcement content. Classical design of experiments methods are therefore increasingly being complemented, or replaced, by machine learning (ML). This review surveys ML applications in FSW and FSP, covering process modelling, parameter optimisation, prediction of mechanical and microstructural properties, defect identification and physics-informed and interpretable methods. Artificial neural networks, tree-based ensembles and hybrid models predict tensile strength, hardness, wear rate and friction coefficient with high accuracy and one study on high-entropy-alloy-reinforced composites reports coefficients of determination above 0.96. Deep-learning models for defect classification exceed 90% accuracy on benchmark radiographic datasets. The main obstacles are scarce and inconsistent experimental data, weak generalisation across alloys and tool types, limited interpretability and little industrial deployment. Shared benchmark datasets, physics-informed and explainable models and systems capable of real-time control are identified as priorities for future work.