Bi-Level Optimization and Dimensionality Reduction of Ply Configurations in Composite Panels via Autoencoders

The next generation, high performance and high aspect--ratio wings require composite panel designs that balance minimum mass with stringent stress, structural performance, and manufacturability related constraints. The paper proposes a bilevel optimization framework that learns a latent space from high dimensional ply‑count and embeds it into a surrogate‑assisted d genetic algorithm (GA) using deep autoencoders. At the top level, number of plies in standard orientations (0, 90, 45, and $-$45\textdegree) serves as design variables; a GA coupled with surrogate models accelerates the global search while enforcing strain, buckling, and ply orientation percentage constraints.The trained encoder maps the ply-count values into a reduced latent space, while the decoder reconstructs the corresponding ply-count values after optimization. Once a high--fidelity surrogate is established, optimization is performed in the latent space, dramatically reducing dimensionality and computational cost. Decoding maps optimal latent vectors back to number of ply-counts, which are then resolved into manufacturable stacking sequences at the bottom level using a permutation GA that enforces blending and stacking sequence rules while matching top--level lamination parameter values. The proposed workflow demonstrates improved optimization efficiency and scalability, enabling a comprehensive population search and a robust exploration of non-‐conventional structural configurations. This approach facilitates rapid iteration between the surrogate evaluation with a lower dimension, making it well suited for the design of a next generation light weight composite wing.