Session: 02-01 Mechanical Behavior of Ceramics and Composites

Paper Number: 127092

127092 - An Integrated Neural Network and Finite Element Method for Defect Characterization and Strength Prediction of Unidirectional Composites 

Unidirectional composite materials are highly regarded for their exceptional mechanical properties, rendering them indispensable in a wide array of engineering applications. Due to their straightforward geometric structure, these materials are notably vulnerable to defects, which can exert a profound influence on their mechanical performance. To address this significant challenge, the study presents an integrated approach, wherein the synergistic power of neural networks and finite element analysis is harnessed to devise a robust method for recognizing defect characteristics and providing accurate predictions of mechanical strength in unidirectional composite materials. In response to the critical role of defects in these materials, ABAQUS, a potent finite element analysis tool, is utilized to simulate random distributions of fibers and pores, along with variations in fiber-matrix mechanical properties. Extensive numerical simulations, drawing from the Monte Carlo method, produce authentic mechanical performance data, which serves as the output label for the neural network. Two methods for the recognition of geometric features are employed. In the first method, image recognition is performed using a two-point cross-correlation algorithm, which is subsequently refined through the integration of techniques such as feature selection and dimensionality reduction, including principal component analysis (PCA). This facilitates the more efficient extraction of essential geometric features while reducing feature dimensionality. Extracted geometric structural information is employed as input features for the construction of predictive neural network models. In the second method, convolutional neural networks (CNN) are utilized for image recognition, with attention mechanisms integrated into the CNN to improve the capturing and recognition of geometric features. Additionally, descriptors such as kurtosis, skewness, and eccentricity have been incorporated as supplementary attributes within the neural network, enabling a more comprehensive analysis of the influence of pore morphology on mechanical performance. Concealed patterns and correlations are unveiled through the application of deep learning techniques, leading to the development of highly precise mechanical property models. This introduced integrated approach for defect characterization in unidirectional composite materials is an innovative and novel method, offering rapid and accurate results, with predicted tensile strength errors consistently remaining under 10%.

Presenting Author: Bo Zhang Beihang university

Presenting Author Biography: Zhang Bo, male, born in March 1999, 24 years old, of Han ethnicity, a member of the Communist Party of China, currently pursuing a Ph.D. in the School of Energy and Power Engineering at Beihang University with a major in Power Machinery and Engineering. His research primarily focuses on the structural strength design and lifetime analysis of ceramic matrix composites.

Authors:

Bo Zhang Beihang university
Changqi Liu Beihang University
Duoqi Shi Beihang University
Xiaoguang Yang Beihang University