Integrated machine-learning modelling for mechanical property prediction – A case study on laser-welded TC4 titanium alloy

Laser welding of TC4 titanium alloy is extensively employed in high-end manufacturing sectors such as aerospace, owing to high specific strength and low density. However, the complex interactions among multiple welding process parameters present a significant challenge for a coordinated control, thus hindering further improvements in joint performance. This study employs machine learning to predict the tensile strength and elongation of laser welded joints of TC4 titanium alloy, aiming to optimize the welding process and improve joint performance. We develop predictive models correlating laser welding parameters with mechanical properties using a multilayer perceptron (MLP), support vector regression (SVR), and four ensemble algorithms—XGBoost, CatBoost, LightGBM, and random forest (RF). XGBoost model achieves the highest accuracy in predicting tensile strength, (R² = 0.90 training; 0.84 test). For elongation prediction, the CatBoost model is better than other models, (R² = 0.90 training; 0.85 test). SHAP analysis demonstrates that heat input has the most significant influence on the tensile strength prediction model, whereas tensile strength is the most critical input variable in the elongation prediction. Incorporating tensile strength as an input variable in the elongation prediction model substantially improves generalization, raising test-set R² from 0.38 to 0.85 (a 123.68% relative improvement), and simultaneously reducing hyperparameter-tuning complexity. Error-propagation analysis reveals that tensile strength prediction errors have only a minor effect on elongation prediction, supporting a phased collaborative modeling strategy between these targets. The findings provide a practical pathway for intelligent optimization of TC4 laser welding processes and precise control of joint performance, thereby advancing lightweight manufacturing of high-end equipments.