LayupFormer: a deep generative model for composite laminate layup design

Abstract

Conventional laminate layup design relies on search-and-evaluate strategies that become intractable as ply counts grow, offering limited guarantees of feasibility, interpretability, and efficiency. While surrogate modelling and optimisation using artificial intelligence have accelerated composite design, most approaches still explore the design space via a search-centric manner. Generative methods offer an alternative by incorporating performance criteria, but often operate with restricted orientation sets, lack interpretability, or require auxiliary tools to ensure feasibility. This article introduces LayupFormer, a physics-informed Transformer framework that reformulates laminate layup design as an inverse sequence generation problem. The solution embeds mechanics through laminate parameters derived from Tsai's invariants and constrains designs with a domain-specific grammar over ply orientations. The framework couples a high-fidelity predictor, which regresses load and stiffness, with a generator that directly produces requirement-compliant layups efficiently across data scales. Attention analyses reveal that LayupFormer captures long-range through-thickness interactions and internalises laminate principle, providing interpretable insights into the generation process. Experimental validation confirms that LayupFormer-designed layups achieve superior bearing performance and reduced variability compared with empirical baselines. LayupFormer establishes a unified physics-informed generative framework that transforms laminate design from search-based optimisation into an interpretable and data-efficient inverse-design process, paving the way for scalable and automated composite design.