Optimal non-guideline blended composite laminate design using generative stacking sequence and optimal thickness prediction models

This article introduces a novel optimization approach for solving the non-guideline ply-drop composite laminate optimization problem using a Goal-Directed Generation framework consisting of two models: a Generator model called Generative Stacking Sequence (GSS) and a Scorer model called Optimal Thickness Prediction (OTP). The discrete design variables for the optimization problem can be divided into two groups: Stacking Sequence and regional thicknesses. The GSS model generates Stacking Sequence samples, which are then fed to the OTP model to predict the optimal feasible regional thickness for those Stacking Sequence samples regarding structural integrity. By training the GSS model toward optimizing the objective function, the present approach replaces discrete metaheuristic algorithms in traditional optimization approaches with Neural Network models with more robustness and versatility. Furthermore, the proposed framework transforms the discrete design variable space of the optimization problem into the smooth latent space of the models’ weights, which enables the usage of gradient-based optimizers for discrete design variables. A non-guideline variant of the 18-panel horseshoe blending optimization problem is utilized to benchmark the proposed approach against previous studies in the literature. The obtained results highlighted the optimization performance of the proposed fully neural optimizer framework as competitive with that of metaheuristic algorithms and hybrid approaches in the literature.