Scribd
Upload
29 views2 pages

Introduction

This document discusses using convolutional neural networks (CNNs) to automatically segment and quantify pore structures in scanning electron microscopy (SEM) images of cathode materials for lithium-ion batteries. A CNN encoder-decoder model is developed that can segment pores from SEM images with over 97% accuracy. Preprocessing techniques like hysteresis thresholding and histogram equalization are examined to optimize the CNN's performance. The porosity measured using this CNN-assisted 2D segmentation is found to be consistent with that measured from 3D electron tomography, demonstrating that CNNs can reliably quantify pore structures while reducing labor compared to conventional manual methods.

Uploaded by

QuiToRiQ
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
29 views2 pages

Introduction

This document discusses using convolutional neural networks (CNNs) to automatically segment and quantify pore structures in scanning electron microscopy (SEM) images of cathode materials for lithium-ion batteries. A CNN encoder-decoder model is developed that can segment pores from SEM images with over 97% accuracy. Preprocessing techniques like hysteresis thresholding and histogram equalization are examined to optimize the CNN's performance. The porosity measured using this CNN-assisted 2D segmentation is found to be consistent with that measured from 3D electron tomography, demonstrating that CNNs can reliably quantify pore structures while reducing labor compared to conventional manual methods.

Uploaded by

QuiToRiQ
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
You are on page 1/ 2

Introduction

Extensive research efforts have been made in recent decades to develop high-
performance durable Li-ion batteries with large capacity. Among promising
candidates, high-capacity Ni-based cathode materials, e.g., LiNi0.8Co0.1Mn0.1O2,
have been considered owing to their outstanding properties related to energy
density, cycling stability, and manufacturing cost (Shim et al. 2019). The
optimization of geometrical parameters, i.e., the porosity of the secondary
particles of the cathode materials, was suggested as an effective strategy to
further improve performance (Chen et al. 2016). X-ray computed tomography (XCT) is
used to characterize the porosity of electrode materials. However, this technique
is dedicated to the microscale characterization of porous electrode structures at
the cell level (Pietsch and Wood 2017). Cross-sectional image analysis or
tomography interpretation using scanning electron microscopy (SEM) combined with a
focused ion beam (FIB)-based serial milling was suggested for precise porosity
quantification owing to its nanoscale resolution down to a level of 10 nm (Cantoni
et al. 2014; Hong et al. 2022). However, statistical interpretation using 3D
tomographic reconstruction or 2D-image-based pore segmentation requires a large
number of image datasets and heavy human input for digital processing. In general,
a subjective decision can be made regarding the weak image contrast of pores in SEM
images, which can reduce the measurement veracity. Therefore, a reliable machine-
perspective automated segmentation analysis of pores in the SEM images of porous
cathodes is required. Digital segmentation methods were typically used to estimate
the pore content in a material based on SEM imaging (Andrä et al. 2013). These
digital methods first detect the edge of an image by applying a random threshold to
each pixel and then, extend the detection condition to the surrounding pixels using
a marker-based watershed. Furthermore, the traditional Otsu method (Zhang and Hu
2008) and Canny edge detection technique (Chen et al. 2015) were used for edge
detection. Alternatively, the expectation–maximization/maximization of posterior
marginals (Comer and Delp 2000) can be considered for edge detection. However,
multiple parameters for this digital processing must be optimized. Digital
segmentation using the aforementioned approaches is generally time-consuming and
requires human intervention for each digital processing step, thus complicating
measurements with statistical significance (Galbany et al. 2005; Gesho et al. 2020;
Roldán et al. 2021). To address the limitations of conventional digital approaches,
convolutional neural networks (CNNs) combined with relevant filters can be used to
devise a promising solution because CNNs are highly efficient in training images,
extracting features, performing semantic segmentation, and interpreting feature
characteristics (Krizhevsky et al. 2012; Long et al. 2015; Nanfack et al. 2018;
Garcia-Garcia et al. 2017). Recently, machine learning/deep learning algorithms
such as random forest classifiers and CNNs for feature segmentation were actively
adapted to reduce contrast artifacts in traditional SEM images and to improve
efficiency and reliability in FIB-SEM tomography for pore structure
characterizations (Zang et al. 2023; Osenberg et al. 2023). Considering the
advantages of the CNN-based approach, the geometrical characterization of high-
capacity Ni-based cathode materials can be rapidly and reliably performed, which
facilitates the understanding of the electrode structures and advances the design
strategy for high-performance electrodes (Finegan et al. 2022; Yang et al. 2022).

This study introduces a CNN-based model for the accurate feature segmentation and
statistical interpretation of geometrical factors, e.g., the pore content
distributed in high-capacity Ni-based cathode materials. To this end, a
convolutional encoder–decoder model was used to construct a deep learning model for
the pore segmentation of the SEM images of cathode materials, and the performance
of the model was evaluated by comparing it with the ground truth data based on
human inputs. The established model can automatedly expedite the segmentation
process for a large dataset of SEM images and effectively learn pore features from
SEM images, reliably determining porosity with an accuracy of > 97%. To optimize
the performance of the proposed model for the semantic detection and extraction of
geometrical pores from raw SEM images, the effects of preprocessing of raw image
data, i.e., hysteresis thresholding (Wang et al. 2008) or histogram stretching and
equalization (Luo et al. 2021; Abdullah-Al-Wadud et al. 2007), were examined. The
preprocessing was generally used to improve feature segmentation performance in
conventional FIB-SEM tomography. However, whether those digital filters are
effective has barely been tested in the CNN-driven automated feature segmentation.
Interestingly, the measured porosity of the material with this CNN-assisted 2D
segmentation process was revealed to be well consistent with the result measured
from the 3D electron tomography after the CNN-based workflow optimization. This
result shows that the labor-intensive effort required for pore structure analysis
of the cathode materials can be substantially reduced, while the human bias in the
segmentation task is avoided, warranting the measurement veracity corresponding to
the learning reliability of the used CNN model.

You might also like