Predictive modeling of mechanical properties in CO₂ laser welding of hastelloy X using supervised machine learning

A supervised machine learning (ML) regression study was performed to predict the tensile strength of CO₂ laser-welded Hastelloy X joints using data from 27 experimental trials involving variations in laser power, welding speed, and focal length. Five regression algorithms — Decision Tree (DT), Random Forest (RF), XGBoost (XGB), CatBoost (CB), and AdaBoost — were evaluated, initially using an 80:20 train-test split, under which Random Forest produced the strongest apparent performance (R² = 0.9483, MAE = 9.28, MSE = 132.937). Given the limited size of the experimental dataset, 5-fold and leave-one-out cross-validation were additionally performed to obtain a more robust assessment of model generalization; under both schemes, Random Forest and CatBoost consistently outperformed the remaining models, with cross-validated R² values of 0.61–0.65, indicating more modest but reliable predictive performance relative to the initial split evaluation. Analysis of variance (ANOVA) performed on the experimental dataset independently confirmed laser power as the dominant and only statistically significant process parameter influencing tensile strength, corroborating the feature importance rankings obtained from the machine learning models. These findings demonstrate the effectiveness of ensemble-based machine learning approaches, particularly Random Forest and CatBoost, in predicting tensile strength from laser welding process parameters, while highlighting the importance of cross-validation for reliable model assessment on small experimental datasets, with implications for reducing experimental effort and improving process optimization in the fabrication of high-temperature alloy joints.