Machine learning predictive approaches for hot crack mitigation in modified TIG welded AA7075 joints

Abstract

ABSTRACT The potential of sophisticated machine learning methodological approaches such as Random Forest Regression (RFR), Artificial Neural Network (ANN), and Multi-Linear Regression (MLR), for the prediction of hot cracking sensitivity and microhardness of Ultrasonic Assisted Tungsten Inert Gas (UA-TIG) welded AA7075 joints is examined in this study. Three regression models are developed on the basis of the aforementioned approaches. Welding current (A) and gas flow rate (L/min) are continuous variables, while ultrasonic vibration and filler material are categorical variables. Hot cracking sensitivity and hardness are taken as responses. The trials are carried out using Central Composite Design (CCD) based Design of Experiments. Among the 52 experiments, 41 are assigned for training models, while the rest are utilized to test models. Various visual and performance metrics are used to evaluate predictive models. The results highlight that the RFR model (R2 testing = 0.964 for hot cracking sensitivity & R2 testing = 0. 976 for microhardness) performs better than other predictive models. Further, a sensitivity analysis is performed to identify the significant impact of each process parameter on the aforementioned responses. Ultrasonic vibration is found to be the most influential factor affecting hot cracking sensitivity and microhardness.