Semantic Segmentation of SOFC Composite Electrode Images Incorporating Patch-Based Convolutional Neural Networks

Abstract

Quantitative microstructure analysis is an important task in many fields of material engineering. In solid oxide fuel cell (SOFC), the microstructure of porous electrodes determines electrochemical performance of the cell. Three dimensional analysis by focused ion beam - scanning electron microscopy (FIB-SEM) has been successfully implemented to evaluate the fabrication strategies and to identify degradation mechanisms. For precise evaluation of the electrode microstructures, segmentation of large number of SEM images is required. The conventional method for image processing of raw FIB-SEM dataset consists of semi-automatic filtering, thresholding, edge-detection, etc. Those semi-automatic methods usually require manual corrections in final stages of the processing to ensure highest accuracy, which greatly extends the processing time. Therefore, number of images, size and resolution of the obtained microstructures are limited not only by the physical capabilities of FIB-SEM hardware, but also by the post processing of datasets.In the present research, the framework for automated SOFC 3D microstructure reconstruction form a large dataset of FIB-SEM images is proposed. The framework is based on machine learning, namely convolutional neural networks (CNN). A simple yet effective deep neural network incorporating the patch-based convolutional layers (patch-CNN) organized in the encoder-decoder manner was designed and tested. Clear guidelines are given for the selection and labeling of training images, selection of network architecture, methods for the CNN training from a scratch and transfer learning. The most influential factors to assure short processing time and accurate segmentation are the selection of optimal training dataset and the proper network architecture. Teaching images have to be labelized prior to training, since minimizing their number greatly reduces the preprocessing time. Selections of optimal network architecture and patch size are crucial in providing enough spatial information and maintaining reasonable training time. The data augmentation and random patch extraction allow artificial enlargement of the training dataset. The augmentation techniques, including brightness and contrast random enhancements, rotation and reflection, are discussed in relation to the original features of different microstructures.The proposed methodology was tested for various electrodes including nickel - yttria-stabilized zirconia (Ni-YSZ) anode, nickel - gadolinium doped ceria pattern electrode (Ni-GDC), pure GDC and La0.9Sr0.1Cr0.5Mn0.5O3-δ - Gd0.1Ce0.9O2-δ (LSCM-GDC) nano-composite porous anodes. The tested data had various resolutions (5 nm – 30 nm), different image sizes (500 – 4000 pixels) and had very different measurements artifacts. The CNN-based approach succeeded in accurate binarization, ternarization and quaternarization of micro- and nano-structures as well as impurities detection. The achieved pixel-based accuracy was over 96%, and the 3D microstructural parameters such as phase fractions, connectivity, total and active TPB densities, surface area density and tortuosity factors calculated from ground truth data and CNN-prediction were identical.One of the challenges of image segmentation in handling FIB-SEM image set is the treatment of artifacts such as curtaining, shadows and charge-ups. Proposed patch-CNN can successfully process all of these artifacts. The detection of pores which are not filled with resin prior to the FIB-SEM measurement is another difficult task for the conventional image processing. The results achieved for the samples with a large number of un-filled pores are presented in Fig. 1. Reconstruction of this sample was not possible with the conventional algorithms based on the pixel brightness values. The two separated CNNs were trained to conduct artificial pore-filling operation (CNN1) and to ternarize the phases (CNN2). The amount of labelized data incorporated for training was 3 and 17 images for CNN1 and CNN2, respectively. The final segmentation was conducted for 935 images and the reconstructed volume size was 54 × 54 × 28 µm with resolution of 30 nm in all spatial directions.Deploying artificial intelligence methods into real problems in material engineering requires fairness, high accuracy, transferability and interpretability. The proposed automatic patch-CNN algorithm can significantly shorten the image processing time with maintaining accuracy, resolution and size of the reconstructed microstructures, which were not available by conventional methods. Proposed algorithm can be easily extended for any multiphase, porous materials and different 3-D imagining techniques.Fig. 1 Framework for automated SOFC 3D microstructure reconstruction form a large dataset of FIB-SEM images without pore filling. Figure 1