Session: 01-05-02: Applications of Artificial Intelligence/Machine Learning for Aerospace Structures 2

Paper Number: 162461

162461 - Neural Network-Assisted Design Optimization With Adaptive Sampling for Tow-Steered Composite Structures 

Tow-steered composite structures offer exceptional potential in aerospace applications by enabling the design and fabrication of curvilinear fiber paths that can improve structural performance and reduce weight. Compared to conventional unidirectional fiber-reinforced composites (UDFRCs), tow-steered composites provide greater design flexibility and spatially tailored material properties. However, this flexibility also expands the design space of tow-steered composites, introducing significant computational challenges. Finite element analysis (FEA)--based optimization of aerospace tow-steered composite structures becomes prohibitively expensive due to the structural complexity and high-dimensional design space. As a result, conventional finite element (FE)-based optimization processes can demand hours to days of computational time.

Machine learning (ML) models, particularly neural networks (NN), have emerged as powerful tools to address these challenges by serving as surrogate models for the FE model. NN is highly effective in capturing complex, non-linear behaviors of tow-steered composites and handling high-dimensional design space. Despite these advantages, its application in tow-steered composite design faces two limitations. First, the high dimensionality of design inputs, involving parameters for defining curvilinear fiber paths and ply configurations, demands a large amount of training data to achieve reliable predictive accuracy. Second, the accuracy of NN-based optimization is inadequately validated in the existing studies, raising concerns about the reliability of the reported performance improvements.

To tackle these limitations, this work introduced an adaptive sampling approach. The key idea is to sample more data around the critical regions in the design space. Thanks to the superior performance of NN models in handling non-uniform training data, the proposed approach enables the NN model to achieve better optimization accuracy with much-reduced training data. In the adaptive sampling approach, we identify critical regions in the design space based on an initial sampling with limited training data. After that, we gradually increased the training data in these critical regions so that the NN model could better capture more critical structural responses. Two different algorithms were developed to sample the data in the critical regions. Two case studies using cylinder tow-composite structures with different numbers of design variables were performed to validate the proposed approach. In these two cases, both adaptive sampling algorithms identify critical design regions and significantly improve optimization accuracy with less training data, achieving reliable structural improvements with limited computational resources.

This study made two contributions. First, we evaluated the accuracy of the NN-assisted optimization for tow-steered composites using two different tow-steered composite examples. The discussion on accuracy provides critical insights into the challenges and potential pitfalls of the NN-assisted optimization methods. Second, we proposed and validated a practical adaptive sampling strategy that enhances optimization accuracy while reducing the training costs of the NN models. The proposed adaptive sampling is very general and can be applied to other NN-assisted optimization problems. By focusing on critical regions within the design space, this approach can improve the optimization accuracy while reducing the requirements on training data, particularly the design problems with high-dimensional design space. Future work will focus on the study of the effects of different parameters in the adaptive sampling method on the accuracy of NN-assisted optimization.

Presenting Author: Bangde Liu The University of Texas at Arlington

Presenting Author Biography: Bangde Liu is a PhD candidate in the Industrial, Manufacturing and Systems Engineering at the University of Texas at Arlington. His research focuses on machine learning assisted multiscale modeling and design of tow-steered composite materials and structures.