An integrated machine learning and optimization approach for enhanced strength prediction in riveted joints

In this study, experimental tests, finite element analysis, and machine learning techniques were integrated to predict the maximum shear stress of riveted joints. First, ten tensile tests were conducted, yielding an average stress of 76.0383 MPa on the plates, and the regression analysis of the stress–strain data demonstrated a 95-99% level of agreement. A finite element analysis performed in Ansys under conditions similar to the experiments, produced a result of 78.875 MPa, which falls within an error margin consistent with the literature. Subsequently, a dataset of 20 samples containing various rivet diameters, plate thicknesses, hole coordinates and tensile loads was generated and used, to train a Regression Decision Tree model in MATLAB. For a new design case, the model predicted 7.480 MPa instead of 7.291 MPa, corresponding to an error of approximately 2.50%. When the dataset was expanded to 50 samples, this deviation decreased to 0.82%, indicating a significant improvement in accuracy. Overall, the results demonstrate that the machine learning model rapidly improves, as additional data become available and can provide reliable, fast predictions, including the effects of bending moments, offering a promising approach that may reduce the need for extensive experimental and numerical analyses.