A deep learning approach to solving the peen forming inverse problem

Peen forming is a cold forming process where surface plastic deformations are created by high-velocity projectile impacts. The resulting local expansions of the treated surface lead to local curvature variations. The effect of the peening treatment can be represented as a thermal expansion on a bi-layered plate, making this problem identical to bilayers in soft material physics. The motivation for this work is the development of automated robotic peen forming, which will increase the repeatability, accuracy, and efficiency of the process. With our goal being full process automation, we are interested in solving the inverse problem of calculating the required peening trajectories to form a part into the desired shape. To achieve this, we present an efficient method using deep learning to recognize patterns linking deformed plates to the required peening trajectories. The proposed model is trained on a large dataset to infer the required peening trajectories. Our results show that the predicted patterns create deformed shape deviations in the order of 100 microns. We consider that this approach can be used as part of a feedback loop to achieve full process automation.