Transformer-Based Neural Network Metamodel for Nearly Orthogonal Sampling in Conceptual Design of Multistage Space Launch Vehicles

This study introduces an innovative methodology for the efficient conceptual design of complex, multidisciplinary systems that involve computationally intensive analyses and a vast array of design variables. A novel nearly-orthogonal sampling strategy with superior space-filling characteristics is employed to extract maximal insights into system behaviour using a significantly reduced number of trial designs. The sampled dataset serves as input for a metamodel constructed using advanced artificial neural networks, augmented by Transformer Networks to enhance the metamodel’s capacity for capturing intricate dependencies and complex interactions within the data. Furthermore, a stage-wise interconnection of discrete neural networks is proposed for trajectory metamodeling, effectively mitigating the dimensionality challenges inherent in traditional neural architectures. The optimization process integrates a hybrid approach, leveraging a Genetic Algorithm for global optimization in tandem with Sequential Quadratic Programming for localized refinement utilizing exact disciplinary analyses. The efficacy of the proposed methodology is demonstrated through its application to the conceptual design optimization of a multistage solid-fuelled space launch vehicle. The results reveal exceptional accuracy in approximating highly nonlinear functions, a substantial reduction in overall computational time, and significant minimization of the reliance on exhaustive disciplinary analyses, underscoring the transformative potential of this approach.