Dissertation

Applications of Deep Convolutional Neural Networks for Energy-Based Materials Characterization on Digital Images

Tang, Kunning

2023

Overview

Energy-based materials, such as hydrogen fuel cells and metal ore, play an essential role in reducing greenhouse gas emissions, facilitating the clean energy transition, and designing renewable energy devices. Image techniques, such as x-ray microcomputed tomography (micro-CT), scanning electron microscopy (SEM), and quantitative evaluation of materials by scanning electron microscopy (QEMSCAN), provide a means to characterize the physical and chemical properties of porous materials. Conventionally, 2-dimensional (2D) image techniques are used to evaluate the bulk properties of a material’s surface, such as QEM-SCAN for mineral liberation analysis. However, 2D analysis of 3D materials causes sampling error and stereological effects, and therefore is insufficient for accurate materials characterization. Micro-CT provides a 3-dimensional (3D) vision of the internal material’s structure at the length scale at which materials interfaces are well resolved, and therefore is used to extract detailed information by coupling with other 2D imaging techniques. However, due to the limits of micro-CT, including resolution limits, scan artifacts, similar attenuation values for different materials, etc. Image processing techniques are required to accurately quantify the information from micro-CT. However, traditional image processing methods, such as thresholding and gradient-based methods for image segmentation, are insufficient to fully characterize the 3D porous material structure and in addition contain significant human-judgement.Machine learning (ML) methods are utilized to alleviate the limitations associated with traditional methods. Firstly, for iron ore sinter bed particles that contain various minerals with different textures and geometry, a supervised ML workflow is proposed to segment micro-CT image sinter images. Instead of voxel-level segmentation methods that cannot segment the particles containing various minerals; the geometric, texture, and grayscale features of each particle are generated as input to a ML model. XGBoost and LightGBM as two gradient descent ensemble ML models are utilised to classify particles based on these features. Accuracy over 90 % is achieved for iron ores that are morphologically domain-distinct in their feature space, while lower accuracy in the order of 40-80 % is achieved between particles that derive from different mine sources. Additionally, several CNN architectures are proposed to perform multi-phase segmentation on sandstone, iron ore, and a hydrogen fuel cell. A novel network called EfficientU-Net and its variance EfficientU-Net-cGAN with an additional image-to-image GAN is proposed that demonstrates a higher pixelwise and physical accuracy than the commonly used U-ResNet. With the assistance of these networks, materials characterization is extended from 2D to 3D without user judgement by coupling 3D micro-CT with other 2D image techniques, such as QEMSCAN and micro-XRF. Meanwhile, two GAN-based networks, PH-GAN and CycleGAN, are presented to increase the generalizability of ML techniques when dealing with a large number of images under diverse scanning conditions. Large 3D images are downscaled using PH-GAN in a way that preserves the topological feature. Lastly, CycleGAN executes unpaired domain transfer to eliminate the domain inconsistency of various scanning situations.Overall, this dissertation demonstrates how ML methods and workflows, along with their coupling with imaging techniques are useful for 3D full-feature characterization of energy-based materials, surpassing the capabilities of conventional characterization techniques and minimizing the influence of scanning conditions for cross-image characterization.

• Publisher: ProQuest Dissertations & Theses
• Subject: Bioinformatics / Biomedical engineering / Coal / Connectivity / Engineering / Iron compounds / Minerals / Neural networks / Permeability / Probability distribution / Quartz / Visualization
• ISBN: 9798304906357