Input-adaptive shear representation for composite laminated plates using a physics-informed DeepONet framework

초록

This study presents a Deep Operator Network (DeepONet)-based framework for identifying input-adaptive shear representations for the static bending analysis of laminated composite plates, aiming to improve the prediction of transverse shear deformation and equilibrium-reconstructed stress responses within the considered benchmark domain. A training dataset is constructed from through-thickness displacement and transverse shear stress distributions of simply supported square and rectangular plates computed using representative higher-order shear deformation (HSDTs), including the third-order shear deformation theory (TSDT) and trigonometric shear deformation theory (TrSDT). The DeepONet architecture employs a branch network to encode the sampled field data and a trunk network to query the normalized thickness coordinate z/h. Two variants are developed: a data-driven DeepONet-based shear deformation model (DSDM) and a physics-informed DeepONet-based shear deformation model (PIDSDM). Within the training domain, both models reproduce 3D elasticity benchmark trends for deflection and stresses. In the unseen sandwich case, PIDSDM gives more stable responses than DSDM within the adopted HSDT-ESL framework. The proposed framework should therefore be interpreted as an operator-learning-based computational model for identifying case-dependent effective shear representations within the considered HSDT-ESL response space, rather than as a new independent shear deformation theory.

키워드