The accurate prediction of structural responses under complex loading is crucial for ensuring structural integrity in engineering systems. Conventional finite element method (FEM) analyses provide high-fidelity solutions but become computationally expensive for iterative evaluations, whereas purely data-driven deep-learning models exhibit limited extrapolation capability and reduced physical consistency. This study proposes a hybrid physics-informed neural network surrogate model for a three-dimensional oil-water separator pressure vessel subjected to internal pressure. The model integrates FEM-generated displacement and stress data with soft physics constraints from linear elasticity and boundary conditions. To improve the training stability in high-dimensional settings, a two-stage transfer learning strategy comprising data-driven pretraining followed by physics-informed fine-tuning is adopted. Validation results indicate the accurate prediction of displacement and von Mises stress fields as well as reliable performance under unseen pressure conditions, including extrapolation at 4.0 MPa, highlighting the potential of the proposed model for rapid and efficient structural assessment.
Choi et al. (Fri,) studied this question.