Data-driven lay-up design of a type IV hydrogen storage vessel based on physics-constrained generative adversarial networks (PCGANs)

This paper proposes a deep learning model that can quickly design the lay-up scheme of a type IV hydrogen storage vessel, which is called physics-constrained generative adversarial networks (PCGANs). A third-party neural network is built to evaluate the model's training output, and a tailored penalty is appended to the generator's loss function to introduce the physical constraint. This penalty term occurs when the metrics of lay-up design fail to meet the requirement, thus changing the training direction of the model. The datasets are generated randomly with the finite element method, and the training process can be visualized by converting one-dimensional lay-up information into a two-dimensional matrix. The results demonstrate that PCGANs has achieved an impressive effect; not only can it generate designs that meet the target, but also it has good accuracy in the prediction. The optimal proposal is selected from the different design options provided by PCGANs, and the burst pressure is 158 MPa, which is 2.25 times greater than the working pressure. The optimal design has 33 hoop layers, the weight of which is reduced by 16.33 % when compared with traditional netting theory, and the maximum fiber accumulation thickness of 24.68 mm, which is reduced by 42.11 %.