LamiNetDef: A novel physics-informed U-Net++ framework for predicting deflection shapes of laminated composite plates

This study introduces LamiNetDef, a novel physics-informed deep learning framework developed to accurately predict the mid-surface deflection patterns of laminated composite plates subjected to uniformly distributed loads. A comprehensive dataset is generated through an automated finite element modeling pipeline, integrating Abaqus and MATLAB via the Abaqus2Matlab toolbox. The model parameters encompass geometric dimensions, stacking sequence, ply orientations, and boundary conditions, sampled using a Latin Hypercube Sampling scheme to efficiently explore the design variable space. The finite element nodal displacements are subsequently normalized and processed to form structured data for training. The proposed LamiNetDef architecture, built upon a modified U-Net＋＋ network, incorporates a Boundary Condition Deflection Form that encodes physical boundary effects as a two-dimensional spatial input to better inform the encoder in learning physically consistent deformation modes. In parallel, design parameters including geometric and material descriptors are processed by Kolmogorov–Arnold Networks to derive adaptive conditioning parameters for the Feature-wise Linear Modulation layers. This mechanism enables dynamic coupling between physics and features, enriching the model's representational capacity. The network is trained under a physics-informed loss function that integrates both magnitude and gradient terms, ensuring smoothness and physical consistency in the predicted deformation fields. The LamiNetDef framework achieves outstanding predictive performance, with R² = 0.951 and MSSIM = 0.987 on independent test data, demonstrating excellent agreement with finite element results. Furthermore, the framework is deployed as a cloud-based interactive platform via Hugging Face, enabling rapid and reliable deflection prediction directly from user-defined design parameters.