Development of an approach for predicting mechanical properties of adhesive bonds using machine learning

In joint design, the determination of adhesive properties can be energy intensive, time consuming and costly as they require multiple steps and long curing times for the adhesives. In recent years, the use of machine learning (ML) algorithms to map complex relationshipshas become established. The increasing availability of data as well as advances in computingpower and data driven algorithms have arrived as a solution to databased joint design. The aim of this thesis is to develop a machine learning based approach for predicting characteristic values and stressstrain curves of adhesive bonding using existing data. In this work,existing data from peerreviewed publication and reports of publiclyfunded projects were collected and structured in a standardized form. Data from epoxy (EP), polyurethane (PU)and silane-modified polymer (SMP) adhesives under tensile and shear load are considered. Stressstrain curves for which the original measurement data are not available are digitized. All data is analyzed with Python and the curves are shortened, smoothed and interpolatedto a uniform set of data points. After further data preprocessing, the resulting data set is employed to develop two approaches to predict mechanical parameters. For an adhesive test, the first approach predicts the strain at σmax and failure stress, while the second approach predicts the entire stressstrain curve. The features used in both approaches are the experimental parameters such as data sheet information, joint geometry and test conditions.