Physics-Guided Machine Learning Framework for Defect Prediction and Mechanical Property Modeling in WAAM of Nitinol Shape Memory Alloy

Wire Arc Additive Manufacturing (WAAM) has emerged as a promising technique for fabricating large-scale metallic components; however, process instability and defect formation remain critical challenges, particularly for thermally sensitive materials such as Nitinol shape memory alloys (SMAs). In this study, a physics-guided machine learning framework is developed to predict defect occurrence and mechanical performance in WAAM-fabricated Nitinol. Ten experimental cases were produced under varying current, voltage, and torch travel speed conditions, and the corresponding heat input was calculated using thermodynamic energy balance principles. Defect observations, hardness values (before and after heat treatment), and ultimate tensile strength (UTS) were experimentally evaluated. A hybrid modeling strategy integrating process parameters with physics-derived heat input was implemented using classification and regression approaches. Leave-One-Out Cross Validation (LOOCV) was employed to ensure robust evaluation given the limited dataset. Results indicate that inclusion of heat input as a derived physical feature significantly improves prediction consistency and interpretability. An optimal process stability window was identified around intermediate heat input levels, corresponding to defect-free builds and stable mechanical response. The proposed physics-guided learning framework demonstrates the potential of integrating metallurgical understanding with data-driven modeling for intelligent process control in additive manufacturing of shape memory alloys.