Structural Analysis of Tow-Steered Composites Using Transfer Learning

Tow-steered composites offer enhanced structural performance by enabling the design of curvilinear fiber paths, making them highly attractive for aerospace applications. However, their widespread adoption is hindered by the high computational cost of exploring the large design space using traditional finite element analysis (FEA). While efficient surrogate models, such as neural network (NN) models, provide a promising alternative, training them typically requires extensive high-fidelity FEA data, which remain computationally intensive. To address this challenge, we propose a transfer learning (TL) framework that leverages sufficient low-fidelity data from coarse-mesh models and a limited set of high-fidelity data from fine-mesh models to train an NN model. Three case studies of tow-steered composite structures with different numbers of design variables are performed. The results show that TL can achieve accuracy comparable to that of conventional NN models trained using high-fidelity data while substantially reducing computational costs. Additionally, key factors affecting TL are systematically investigated, including the quality of low-fidelity data, the mixed ratio of the low- to high-fidelity data, and the architecture of the TL model. The results provide valuable insights and practical guidelines for applying TL to develop efficient NN models for structural analysis.