Increased Reliability of Draw-In Prediction in a Single Stage Deep-Drawing Operation via Transfer Learning

In sheet metal forming operations, finite element simulations and experimental works are used to evaluate predictions on different parameter settings. During the manufacturing process, there often exists a discrepancy of expected outcomes due to varying material properties. With the aim to save simulation and experimental resources, this paper provides a reliable transfer learning model suiting the deep-drawing case, where a model is pre-trained on simulation data, neurons are frozen in different layers and is then fine-tuned on real data. This model is evaluated in its behavior by gradually learning on different number of real data points as well as simulation data points.