A transfer learning method in press hardening surrogate modeling: From simulations to real-world

The introduction of data-driven surrogate models is a powerful solution to obtain a representation of a manufacturing system, overcoming the limitations of finite element simulations regarding complexity and time. Usually, data acquisition in real manufacturing plants is a very expensive task, and finite element simulations are employed to train Machine Learning-based surrogate models. However, the approximations of the finite element models may induce a deviation from reality that is transferred to the surrogate models. This paper proposes a methodology to combine AI-based surrogate modeling and transfer learning to create a trustworthy and efficient surrogate model of a real manufacturing process, using a low-fidelity finite element model as a source. In particular, the methodology has been demonstrated in a study involving press hardening of boron steel sheet in a pilot plant. Two deep neural networks have been trained with low-fidelity ABAQUS simulations, forming a baseline surrogate model that predicts the key outputs of the process. The use of few experimental real data of the process to perform transfer learning and adapt the original baseline surrogate model to the real environment shows remarkable results, surpassing other Variable-Fidelity Modeling approaches. The final transfer learning surrogate model provides fast and good predictions of the most relevant outputs of the real process with little training, and it removes completely the calibration stage or the need of a high-fidelity simulation model. Additionally, the presented methodology can be a trigger for creating efficient virtual manufacturing environments that can enable developing digital twins or reinforcement learning agents for process optimization.