Data driven prognosis of clinch joints in multi-material design

<jats:p>Abstract. Clinching is a conventional mechanical joining process used in Multi-Material Design in the automotive sector. To receive the desired geometrical characteristics in clinch joints, correct process design is required. To reduce the cost of finding fitting process parameters, numerical simulation of the joining process can be used to predict the geometrical characteristics, such as interlock, instead of an experimental approach. These numerical simulation models consume computational resources and time. In this paper machine learning is used to find correlations between features of the joining process and geometrical characteristics in the joint. This serves the purpose of predicting the joint’s target values more resource-efficiently. Modelling with machine learning requires a structured dataset with sufficient parameter variation. To create this data base the following procedure was used. For joining partners, a HC340LA steel alloy with 2 mm material thickness was used punch-sided and an EN AW 5182 aluminum alloy with 1.5 mm thickness was used die-sided. For this combination a suitable tool combination and punch distance was experimentally identified. A finite element model was created to reproduce the joining process. For the modelling of the material of both joining partners flow curves determined by Vallaster et al. were used [1]. The punch and die were recreated digitally by opto-electronic measurements and transformed into a mesh suitable for numerical simulation. The model was validated by comparing process values like the maximum force applied by the punch and geometrical values in the joints cross section. Additionally, a process window for suitable punch distances was experimentally determined. Afterwards a variation of 70 different process designs was conducted with variants inside and outside the process window. The results were used for training, testing and validating various machine learning models. All models competed against each other to find the must suited model to predict every geometric value. To ensure good model performances and prevent the model from overfitting, a tenfold cross validation was used for validating the models. Analysis of the results gives the following key findings: i) Good predictability is reached for the interlock and sheet thickness of the joint. ii) Prediction neck thickness showed low error values, but also low correlation. iii)The prediction of those key values for evaluating clinch joint characteristics by machine learning models positively impacts needed resources in comparison to numerical models.</jats:p>