Machine Learning-Driven Fatigue Life Prediction in Notched Friction Welded UNS S31603 Stainless Steel

Welded joints are used in many engineering applications, including aerospace, automotive, and structural engineering. However, notches in the welded components reduce their fatigue strength and lead to early failure through stress concentration. It is, therefore, crucial to comprehend and estimate the fatigue behaviour of friction-welded joints with a notch to ensure the structural integrity and component durability. This study aims to examine the notch fatigue behaviour of friction welded joints of UNS S31603 stainless steel and to build up regression models based on Machine Learning, such as Gaussian Process Regression (GPR) and Support Vector Regression (SVR) for accurate fatigue life prediction. Fatigue tests were performed on friction-welded UNS S31603 stainless steel specimens under strain-controlled conditions with fully reversed loading (R=-1) at a constant strain amplitude of 0.3%. The effect of notch width, depth and central angle on fatigue life was investigated. Experimental data was used to train the ML models, and hyperparameter tuning was done using grid search and 4-fold cross-validation. Model accuracy was measured by R², RMSE and MAPE values. The experimental results indicated that the fatigue life of notched welded joints is drastically reduced compared to the base metal, and deeper notches reduce the fatigue strength by up to 84%. It was observed that GPR outperformed SVR in the prediction of fatigue life and had an R² value of 0.9932 with RMSE of 8.61 for the parent material and an R² value of 0.9933 with RMSE of 15.76 for the welded joints. The results highlight the need to optimise notch design and welding conditions to enhance fatigue resistance in engineering applications. Major Findings: The current research investigated the detrimental effect of notch parameters on the fatigue life of both parent material and friction welded joints of UNS S31603. The work also found the GPR model to be the best fit regression model compared to the SVR.

View article page

Download PDF