A Transfer Learning method for deep drawing force prediction of sheet metal

The deep drawing process involves complex deformations with highly nonlinear interactions among geometry, physics, and boundary conditions. Deep Drawing Force (DDF) refers to the force exerted by the punch on the sheet during forming and is essential for mechanical design and the prediction of failure modes such as wrinkling or fracture. This study proposes a data-driven transfer learning framework to predict DDF under various process conditions, where knowledge learned from analytically generated data is transferred to a model trained on limited numerical or experimental data, effectively reducing the data requirement while maintaining high prediction accuracy. The transfer learning in this work is entirely data-driven and relies on analytically generated data to pre-train a base model. Specifically, a large synthetic dataset is generated for pre-training using theoretical forming force equations based on the energy method. This base model is then fine-tuned using a small number of high-fidelity Finite Element Method (FEM) results that have been experimentally validated, finally, the prediction performance is evaluated by testing on unknown FEM data. The pre-trained knowledge is also transferred to tasks involving parts with different geometries, demonstrating strong generalization capability. Experimental results show that the transfer learning framework significantly improves prediction accuracy while reducing data requirements, offering an efficient and cost-effective solution for deep drawing process modelling.