Leveraging Machine Learning to Predict Welding Quality in Ultrasonically Assisted TIG Welded Inconel 625 Joints

The study employs machine learning models to predict hot crack length and microhardness in ultrasonically assisted tungsten inert gas (UA-TIG) welded Inconel 625 joints. Welding parameters such as current, gas flow rate, ultrasonic vibration, and filler metal are varied. Fifty-two welding experiments are conducted using a central composite design matrix with four variables: two continuous variables with three levels and two categorical variables with two levels. Based on the design matrix, the experiments yield data analyzed via machine learning models to predict outcomes and assess differences between actual and predicted values. The predictive models evaluated include Random Forest Regression (RFR), Artificial Neural Networks (ANN), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Multiple Linear Regression (MLR). Comprehensive statistical performance measures and visual techniques were used to determine the most effective predictive model. Among these, the RFR model demonstrates superior performance, achieving high accuracy with coefficients of determination (R2) of 0.9866 (training) and 0.9772 (testing) for crack length and 0.9906 (training) and 0.9612 (testing) for microhardness. These results highlight the robustness of the RFR model in reliably predicting welding outcomes, even with a limited dataset. This study underscores the value of integrating machine learning techniques with welding experiments, providing a framework for improving weld quality and mitigating hot cracks in nickel-based superalloy joints.