Target-driven inverse design of periodic grillage-stiffened panels for buckling resistance using C-VAE and GNN

The design of stiffened shells for optimal buckling resistance is a challenging combinatorial optimization problem. The primary challenges stem from the vast design space of discrete stiffener layouts and the high computational cost of high-fidelity simulations. To overcome these barriers, this study proposes an integrated data-driven framework that combines parametric design, automated finite element analysis, deep learning-based performance prediction, and generative inverse design. A compact tensor-based encoding scheme uniquely represents stiffener unit cell configurations within a periodic lattice. An automated simulation pipeline constructs a dataset of critical buckling loads, which is used to train a Graph Neural Network (GNN) surrogate that achieves accurate buckling strength prediction (R²=0.9882, MAPE = 3.75%). Building upon this predictive model, a Conditional Variational Autoencoder (C-VAE) learns a probabilistic mapping from target performance to latent design space, enabling generation of stiffener layouts that meet specified buckling loads. A screening strategy employing the pre-trained GNN ranks generated designs by predicted performance and structural mass. The results demonstrate that this generation-screening framework provides fast, accurate performance predictions and successfully generates structurally efficient designs from target specifications, offering a promising approach for accelerating the discovery of advanced lightweight structures.