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

Paper Number: 138536

138536 - Structural Analysis of Tow-Steered Composite Structures Using Physics-Guided Neural Networks 

Tow-steered composites are new material systems that can be tailored to achieve curvilinear fiber orientations for optimal load paths to enhance structural performance. Comparing to the unidirectional fiber-reinforced composites (UDFRCs), tow-steered composites have shown improved buckling load, aeroelastic properties and reduced weight. One challenge for developing tow-steered composite structures is the computational cost for optimizing fiber paths and layups considering the huge design space. In particular, the current studies mostly focus on exploring innovative fiber paths in simple composite laminates (e.g., flat panel) while the computational cost for optimizing fiber paths in realistic aerospace structures could be computationally prohibitive.

In recent years, machine learning (ML) models have been widely developed as ultra-efficient surrogate models to address many computational challenges in the modeling of composite materials and structures. However, advanced ML models, such as neural network (NN) models, often require a significant amount of training datasets to meet the required accuracy. For the analysis of composite structures, training data are mostly generated from finite element (FE) simulations, where each simulation could take several minutes. As a result, the computational cost for generating required training data from FE-based simulations is also very expensive. In this paper, we will present a physics-guided NN models to compute the structural responses with reduced the training costs. Instead of using design variables of fiber paths as the input, we will first compute the ABD matrix of the laminate in the structural, and then transfer the ABD matrix into a standardized domain. As a result, the input for the NN model is 6 images containing the distributions of ABD matrix in a structure. The output of the NN model is the structural responses, such as a critical buckling load. A plate structure and a cylinder shell structure will be used as examples to showcase the proposed NN models. The training data will be generated using FE software Abaqus and an Abaqus plug-in called the Design Tool for Advanced Tailorable Composites (DATC). Leveraging the DATC, the training data for the NN models can be effectively generated. To demonstrate the improved efficiency of the physics-guided NN models, we will also develop NN models that take the design variables as the input and predict the structural responses. The proposed work will investigate the potential of including known mechanics into ML models for composites modeling, resulting in reduced training cost and better physical interpretability.

Presenting Author: Xin Liu University of Texas at Arlington

Presenting Author Biography: Dr. Xin Liu is an Assistant Professor in the Mechanical and Aerospace (MAE) Department at the University of Texas at Arlington. He is also a member of the Institute for Predictive Performance Methodologies (IPPM) at the UTA Research Institute. Dr. Liu obtained his PhD in 2020 and Master of Engineering in Aeronautical and Astronautical Engineering in 2016 from Purdue University. His expertise is in data-driven multiscale modeling of composite materials and structures.