[Paper Review] Predicting Mechanical Properties from Microstructure Images in Fiber-reinforced Polymers using Convolutional Neural Networks

Yixuan Sun, Imad Hanhan+2 more|arXiv (Cornell University)|Oct 7, 2020

Non-Destructive Testing Techniques36 references19 citations

TL;DR

This study proposes a fully convolutional neural network (FCNN) based on StressNet to predict stress fields in 2D microstructure slices of fiber-reinforced polymers from segmented tomography images. Trained on finite element (FE) simulation data of a single specimen, the model predicts stress distributions in seconds—compared to 92.5 hours for FE—accurately capturing high-stress regions on fibers, suggesting potential for rapid damage site identification.

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.

Motivation & Objective

• To accelerate mechanical response prediction in fiber-reinforced polymers by replacing computationally expensive finite element analysis (FEA) with machine learning.
• To develop a data-driven model that learns the mapping from microstructure images to stress fields without requiring explicit physics-based modeling.
• To explore whether trained neural networks can identify incipient damage sites based solely on microstructural features.
• To evaluate the feasibility of using deep learning for fast, accurate stress field prediction in non-linear, fiber-reinforced polymer composites.
• To assess the generalization capability and limitations of a model trained on a single microstructure specimen under specific loading conditions.

Proposed method

• A fully convolutional neural network (FCNN) with an encoder-decoder architecture was adapted from StressNet, originally designed for linear elastic materials, to handle non-linear finite element (FE) simulations.
• The network was trained on 2D slices of segmented microstructure images and corresponding stress field outputs from high-fidelity FE simulations of a single 3D fiber-reinforced polymer specimen.
• Input data consisted of segmented micrographs (fiber and matrix phases), while output was the predicted stress field, specifically σzz (normal stress in the z-direction).
• Training used a single specimen with 44.5 million elements, and the model was evaluated via a single forward pass on test microstructure images.
• The model was evaluated using R² and visual comparison between predicted and FE-simulated stress fields, with emphasis on stress concentration near fibers.
• Future extensions are proposed using conditional GANs to generalize across boundary conditions and 3D convolutions to preserve volumetric spatial information.

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict the stress field in fiber-reinforced polymer composites from 2D microstructure images alone?
• RQ2 How does the performance of the FCNN compare to full finite element simulations in terms of accuracy and inference speed?
• RQ3 Can the trained model identify regions of high stress that correlate with experimentally observed incipient damage sites?
• RQ4 What are the limitations of a model trained on a single microstructure specimen in terms of generalization to other composites or loading conditions?
• RQ5 To what extent can the model’s predictions be extended to different boundary conditions or 3D microstructures?

Key findings

• The trained FCNN predicted stress fields with an R² of 0.69 on the test set, explaining 69% of the variance in the true stress field from FE simulations.
• The model accurately captured stress concentration on fibers, especially in regions corresponding to experimentally observed incipient damage sites.
• Prediction inference took less than a second on a standard laptop, compared to 92.5 hours required for a full FE simulation on a high-performance cluster.
• The model performed best when predicting σzz (normal stress in the z-direction) from microstructure slices in the xy-plane.
• Despite strong visual agreement on fiber stress concentrations, the model showed limitations in capturing global stress field variations due to single-specimen training.
• The study suggests a corollary that such models may be used to identify potential damage initiation sites in fiber-reinforced polymers based on microstructure alone.