Machine learning and finite element analysis: An integrated approach for fatigue lifetime prediction of adhesively bonded joints

Abstract

Since fatigue investigations are expensive and time consuming, models capable of predicting lifetime by leveraging existing experimental data are desirable. Here, this task is accomplished by combining machine learning (ML) and finite element analysis (FEA). The dataset contains 365 points comprising four adhesives with four different joint types. The model is fed with four input parameters: stress ratio and stress amplitude (functions of the applied load), and stress concentration factor and multiaxiality, which are obtained from FEA. An extremely randomized trees (ERT) algorithm, capable of dealing with small and noisy datasets, is used to design the model. After calibration, the model's performance was assessed on unseen data and compared with a linear regression model. The ERT predictions yield a significantly smaller error factor (ER) of 2.13 than that of the linear model (ER = 5.89). A relevance analysis shows that at least one FEA-based parameter must be fed into the model.