Physics-informed and hybrid machine learning in friction stir welding: toward autonomous and certifiable control

Physics-informed and hybrid machine learning approaches are emerging as critical enablers of robust and autonomous friction stir welding (FSW) systems. While conventional finite element models provide physical interpretability but suffer from high computational cost, and purely data-driven machine learning models offer predictive speed but lack generalization and explainability, neither approach alone is sufficient for safety–critical and industrial-scale welding applications. This review critically examines the evolution from empirical and physics-based modeling toward hybrid and physics-informed artificial intelligence frameworks that integrate thermomechanical governing equations with data-driven learning architectures. The paper synthesizes recent advances in supervised and unsupervised learning for weld quality prediction, defect detection, tool wear monitoring, and microstructure estimation, and highlights their limitations in transferability, data dependency, and certification readiness. Particular emphasis is placed on physics-informed neural networks (PINNs), hybrid finite element-machine learning models, reduced-order surrogate modeling, and multi-fidelity data fusion strategies that embed conservation laws, constitutive behavior, and process constraints directly into learning algorithms. These approaches demonstrate improved robustness, reduced training data requirements, enhanced extrapolation capability, and greater interpretability compared to purely black-box models. The review further examines explainable artificial intelligence (XAI) techniques for linking process parameters to microstructural evolution and mechanical performance and discusses the integration of hybrid models into real-time control architectures, reinforcement learning frameworks, and digital twin systems. Key challenges including data standardization, model generalization across alloys and joint geometries, computational efficiency, cybersecurity, and industrial validation are critically analyzed. By unifying thermomechanical physics with adaptive learning, physics-informed modeling provides a pathway toward trustworthy, scalable, and self-optimizing FSW systems. The review concludes by outlining a roadmap for autonomous intelligent welding platforms capable of supporting next-generation aerospace and automotive manufacturing within Industry 4.0 ecosystems.