Machine learning modeling of a deep drawing process for predicting resulting component properties after springback

Abstract. Batch and process fluctuations in the fabrication of sheet metal components can lead to variations in product properties, potentially impacting subsequent manufacturing stages and increasing rejection rates. Identifying and analyzing these deviations is essential to maintain consistent product quality. To counteract such variations, a metamodel for a process chain, including deep drawing, clamping, and clinching, is proposed. This work presents a machine learning (ML) modeling approach for a deep drawing process with an s-rail tool to predict key properties such as effective plastic strain, maximum principal strain, von Mises stress, and residual sheet thickness after springback. The goal was to present a feasible approach for the ML modeling of the investigated deep drawing process and assess prediction qualities across different component sections. It was demonstrated that, with a moderate number of simulations, the predictive capabilities of the models for the forecasted parameters, as measured by the normalized root mean square error, averaged no more than just below 12 percent. However, certain regions, such as critical ends of the flange areas and transitions or straight sections near curved areas, had worse prediction quality compared to other sections. These regions were highlighted for future research, which should focus on incorporating additional features into the ML training process to improve predictions in these identified sections. In a broader scope, this ML modeling for the deep drawing process could be combined with ML modeling of subsequent processes, such as clamping and clinching, to develop a holistic metamodel for the entire process chain.