Cross-process transfer learning guided by materials knowledge for data-efficient prediction of phase transformation temperatures in twin-wire directed energy deposition-arc NiTi-X alloys

Twin-wire directed energy deposition-arc (TW-DED-arc) presents a transformative route for fabricating large-scale NiTi-based shape memory alloys (SMAs). However, accurate prediction of phase-transformation temperatures in TW-DED-arc NiTi-X (Cu/Nb/Hf) alloys remains challenging because of the scarcity of process-specific data. Therefore, this study proposes a materials-knowledge-guided transfer-learning framework for the accurate prediction of phase-transformation temperatures, including A f and M s . A hybrid dataset integrating literature data and experimental data was constructed. By incorporating materials-knowledge-guided feature engineering and an instance-weighted transfer learning strategy, the framework effectively optimizes the predictive performance of traditional machine learning models on the nine-sample TW-DED-arc target-domain evaluation set. Comparative analysis reveals that the optimized Extra Trees (ET) regressor outperforms eleven other algorithms, achieving R 2 values of 0.958 for A f and 0.986 for M s . Experimental verification on NiTiCu, NiTiNb, and NiTiHf ternary alloy systems confirmed the model's versatility. Furthermore, SHapley Additive exPlanations (SHAP) analysis identified the principal feature contributions to the ET model predictions across the three NiTi-X alloy systems. This work establishes a data-efficient framework for predicting the phase-transformation temperatures of TW-DED-arc-fabricated SMAs by leveraging historical materials data for an emerging manufacturing process.