Predicting Mechanical Properties from Microstructure Images in Fiber-reinforced Polymers using Convolutional Neural Networks

TL;DR

A fully convolutional neural network modified from StressNet is extended here for a non-linear finite element (FE) simulation to predict the stress field in 2D slices of segmented tomography images of a fiber-reinforced polymer specimen and shows promise in using ML techniques to conduct fast structural analysis.

Abstract

Evaluating the mechanical response of fiber-reinforced composites can be extremely time consuming and expensive. Machine learning (ML) techniques offer a means for faster predictions via models trained on existing input-output pairs and have exhibited success in composite research. This paper explores a fully convolutional neural network modified from StressNet, which was originally for lin-ear elastic materials and extended here for a non-linear finite element (FE) simulation to predict the stress field in 2D slices of segmented tomography images of a fiber-reinforced polymer specimen. The network was trained and evaluated on data generated from the FE simulations of the exact microstructure. The testing results show that the trained network accurately captures the characteristics of the stress distribution, especially on fibers, solely from the segmented microstructure images. The trained model can make predictions within seconds in a single forward pass on an ordinary laptop, given the input microstructure, compared to 92.5 hours to run the full FE simulation on a high-performance computing cluster. These results show promise in using ML techniques to conduct fast structural analysis for fiber-reinforced composites and suggest a corollary that the trained model can be used to identify the location of potential damage sites in fiber-reinforced polymers.

Figures (5)

• Figure 1: The segmented specimen microstructure is shown in (A), where the blue state represents polymer and the yellow state represents fibers, and the FE simulation result is shown in (B), where the Von Mises stress is plotted.
• Figure 2: Convolutional neural network architecture with an encoder-decoder structure. This network takes microstructure images of size 32 $\times$ 32 as input and outputs the corresponding stress field of the same size. The red highlighted blocks are Squeeze-Excitation Residual blocks hu2018squeeze and the rest are plain 2D convolution layers with MaxPooling.
• Figure 3: Mean squared error of training and testing losses (top) and curves of coefficient of determination of model prediction on the training and testing sets (bottom) for each of the three orthogonal planes: (A,B) xy-plane, (C,D) xz-plane, and (E,F) yz-plane.
• Figure 4: Visualization of the predicted stress fields from the CNN and the true stress fields from the FE simulation on the 9 data points randomly selected from the testing set. The first column shows the input microstructure; the second column is the corresponding predicted stress fields; the third column shows the true stress field obtained from FE simulation; the fourth and fifth columns show the predicted and true stress within the fibers, respectively. All stress metrics correspond to the normal stress relative to the loading direction, $\sigma_{zz}$.
• Figure 5: (first column) 3D phase reconstruction of the microstructure. Reconstructed stress fields (normal stress along the loading axis) from (second column) the predicted by the CNN model and (third column) the FE simulation dataset representing the training and testing sets. (A) The stress field over the entire sample volume of the composite. (B) The stress field within the discontinuous glass fibers.