Machine Learning based bidirectional modelling of rotary friction welding

The objective of this article is to develop bidirectional predictive models for estimation of mechanical properties namely tensile strength and metal loss and input process parameters such as heating pressure, heating time, upsetting pressure, and upsetting time of rotary friction welding of stainless steel (AISI 304) using machine-learning (ML) techniques namely support vector machines (SVM) based on input-output training data. Machine learning (ML) techniques are increasingly employed to establish quantitative relationships between input parameters and critical process responses in welding operations. However, despite their growing adoption, there remains limited clarity on the development of an integrated ML framework that concurrently accounts for both dataset characteristics and model selection to ensure accurate and reliable prediction outcomes. In this study, the datasets underwent systematic preprocessing to eliminate noise, enhance data quality, and extract relevant features representing the inherent physical behavior of the process. Model performance was subsequently evaluated using an independent test dataset, and predictive accuracy was quantified through the mean absolute percentage deviation (MAPD). The developed models have been tested using a separate set of test data and compared in terms of mean absolute mean deviations (MAPD) in predictions for both the outputs and inputs. The prediction accuracies of the developed ML-based models for both forward and inverse cases were found satisfactory.