Large-scale aerospace structures often exploit high stiffness-to-weight ratios and the anisotropic advantages of composites through a two-step optimization process. In the first step, the lamination parameters (a set of 12 continuous variables) are optimized using efficient gradient-based methods. In the second step, stacking sequences are derived to match the optimized stiffness properties. However, lamination parameters are interdependent, and only a bounded convex subset represents realizable stiffness properties. The feasibility bounds that relate all lamination parameters remain unknown. We broadly classify methods for determining feasibility bounds into three categories based on our literature survey. Computational methods enumerate stacking sequences to generate large datasets of feasible points, but their computational cost grows exponentially with dimension. Analytical and semi-analytical methods bypass enumeration using mathematical formulations and geometric insights, yet fail to capture the complete set of lamination parameters. We propose a novel methodology that integrates deterministic boundary-point generation with neural networks to determine feasibility bounds. Our approach extends Diaconu’s method [1], which identifies a tangent hyperplane to the feasibility bounds. To obtain a representative dataset for training neural networks, we incorporate discrete angles in the layerwise optimization step, enabling locating vertices (points of tangency) at reduced computational cost. The preliminary 2D and 3D results confirm that neural networks can accurately capture feasibility bounds, pointing to a promising path for higher-dimensional scenarios [2]. In conclusion, our method using neural networks can encode a large number of bounding hyperplanes within their architectures, providing a more scalable solution for determining feasibility bounds