| 研究課題/領域番号 |
21K14090
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
Sciazko Anna 東京大学, 生産技術研究所, 特任助教 (30898945)
|
|---|
| 研究期間 (年度) |
2021-04-01 – 2023-03-31
|
| キーワード | solid oxide fuel cell / machine learning / 3D microstructure / FIB-SEM / GAN network / semantic segmentation / super-resolution |
|---|
| 研究実績の概要 |
The multi-sized porous electrodes are crucial for the electrochemical performance of solid oxide fuel cells (SOFC) and batteries, but their quantitative characterization is a challenging task. The common method to analyze 3-D structures with feature size of 0.01-30 microns is focus ion beam-scanning electron microscope (FIB-SEM), but its applicability to multi-scale porous media is limited.
In FY2021, a framework for automated SOFC microstructure reconstruction from asymmetric-resolution FIB-SEM was developed. The framework enables microstructure reconstruction with large volume and high resolution. Deep neural networks consisting of patch-VDSR residual network for the super-resolving the in-depth direction and patch-CNN in the encoder-decoder configuration for semantic segmentation were incorporated. Moreover, the algorithm for fabrication of artificial 3D microstructure models directly from 2D SEM image was developed. The developed method is based on the generative adversarial network with 3D generator and 2D discriminator.
In order to investigate the correlation between multiphase porous electrode morphology and the SOFC electrochemical performance, the anodes with controlled volume fraction of nickel - gadolinium doped ceria (Ni-GDC) and porosity were fabricated. The porosity was controlled by initial packing density. The study revealed the importance of enhanced ceramic phase composition and decreased porosity.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The Ni-GDC SOFC fuel electrodes with the controlled porosity and phase fractions were successfully fabricated. The electrochemical tests and morphological characterizations were conducted. The machine learning algorithms for the automatic processing of large high-resolution datasets of Focus Ion Beam-Scanning Electron Microscope (FIB-SEM) images were developed. The semantic segmentation algorithm achieves over 98% processing accuracy. The developed asymmetric super-resolution algorithm allows for reconstructing FIB-SEM samples with up to 8 times difference in SEM and FIB resolution. The artificial 3D microstructure models were fabricated with the generative adversarial network basing of the 3D and 2D training data.
|
| 今後の研究の推進方策 |
The planned research involves application of developed machine learning algorithms to the quantitative evaluation of fabricated SOFC electrode samples. In the scope of experimental research, further FIB-SEM measurements are planned. The samples with reduced particles and pore sizes are under preparation to investigate the influence of diffusion mechanism in porous electrode. In the scope of the algorithms development, the artificial microstructure generation will be further investigated for the influence of the minimal volume representative element. Additionally, the methods to fabricate artificial microstructures with predefined gradient and layered structures are desired.
|
| 次年度使用額が生じた理由 |
The reduced spending in FY2021 comparing with the initial budget are due to the modification in the workstation specification and release of a newer NVIDIA GPU with improved computing capabilities. Additionally, as the international travels were restricted the participation in the international conferences was shifted to virtual event.
The usage plan in FY2022 include the purchase of additional GPU card for speeding-up the 3D microstructural calculations and participation in the additional international conference (the Materials Science & Technology (MS&T22) technical meeting and exhibition).
|
