Embedding composite failure mechanics into neural networks: A physics-constrained framework for dual-criteria failure assessment of Type IV pressure vessels

The transition toward hydrogen-based mobility demands lightweight, high-pressure storage systems with guaranteed structural integrity. Type IV composite overwrapped pressure vessels fulfil this role, yet their design optimisation remains constrained by the computational cost of high-fidelity finite element analysis. This study presents a physics-constrained neural network framework for rapid dual-criteria failure prediction in Type IV vessels. A finite element model is validated against experimental burst tests on a glass/epoxy vessel, achieving prediction errors of 0.63% and 2.48% for the Puck and Hashin criteria, respectively, and cross-validated against four independent burst-pressure cases spanning 20 to 70 MPa, with mean absolute errors of 5.0% (Hashin) and 2.8% (Puck). A dataset of 2,000 carbon-fiber-reinforced polymer configurations trains the surrogate, whose architecture embeds the Hashin and Puck criteria as gradient-based penalty terms within the loss function through a dual-output network with adaptive loss weighting. The surrogate achieves a coefficient of determination of 0.941 and a root-mean-square error of 0.066, with a computational speedup of approximately 1.8 million times relative to the finite element simulation. Leave-one-material-out cross-validation confirms that physics-based regularisation halves performance degradation under extrapolation compared with conventional neural networks (7.1% versus 14.7%). The framework is presented as a proof-of-concept for methodology transfer; certification-grade deployment requires a dedicated experimental campaign.