| 研究課題/領域番号 |
23K13261
|
|---|
| 研究種目 |
若手研究
|
| 配分区分 | 基金 |
|---|
| 審査区分 |
小区分19020:熱工学関連
|
|---|
| 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) |
2023-04-01 – 2025-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
4,680千円 (直接経費: 3,600千円、間接経費: 1,080千円)
2024年度: 1,690千円 (直接経費: 1,300千円、間接経費: 390千円)
2023年度: 2,990千円 (直接経費: 2,300千円、間接経費: 690千円)
|
|---|
| キーワード | Solid Oxide Fuel Cell / Degradation / Machine Learning / Unsupervised learning / Microstructure / Unsupervised Learning |
|---|
| 研究開始時の研究の概要 |
Electrochemical devices like fuel cells, electrolyzers and batteries undergo microstructural changes during long-term operation which causes performance deterioration. Predicting these changes is challenging as the underlying mechanisms are not fully understood and there is limited experimental data available. Here, a new machine learning method is proposed to predict microstructure modification during the reduction-oxidation (redox) process of nickel-based Solid Oxide Fuel Cell anode by incorporating physics-constrained unsupervised neural networks.
|
| 研究実績の概要 |
Electrochemical devices such as fuel cells, electrolyzers, and batteries undergo microstructural changes during long-term operation, leading to performance deterioration. Predicting these changes is challenging due to incomplete understanding of underlying mechanisms and limited experimental data availability.
In FY2023, a machine learning framework was proposed for predicting microstructure modifications during the reduction process of nickel-based Solid Oxide Fuel Cell (SOFC) anodes. The framework integrates unsupervised neural networks (UNIT) and conditional unsupervised neural networks (C-UNIT). These algorithms were tested on simplified toy problems and for predicting microstructural changes in real SOFC anodes, achieving not only a high visual agreement with real experimental data but also fitting characteristic microstructural parameters.
Moreover, experimental studies were conducted to understand the influence of process conditions on the reduction and oxidation of Ni-yttria stabilized zirconia (YSZ) and Ni-gadolinium doped ceria (GDC) anodes. The investigated conditions included temperature (ranging from 500°C to 1000°C), reduction time, and the number of redox cycles.
The proposed method has the potential to simulate not only microstructural changes during redox cycling but also various degradation processes.
|
| 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-YSZ and Ni-GDC SOFC fuel electrodes were successfully fabricated, and redox tests and morphological characterizations were conducted. Machine learning algorithms for the prediction of morphology changes from a limited experimental dataset were developed. The achieved pixel-wise prediction accuracy for the toy problems was over 98%, and good agreement was achieved for the microstructural parameters between real and artificially degraded microstructures.
The feasibility of the conditional C-UNIT algorithm was tested on the reduction of Ni-YSZ anode conducted at different temperatures. The C-UNIT was able to predict various dominant processes depending on the external conditions, resulting in drastic differences in the final microstructures.
|
| 今後の研究の推進方策 |
The planned research involves the application of developed machine learning algorithms for the quantitative predictions of SOFC electrode redox processes depending on various external conditions. In particular, it is planned to incorporate the reduction time as a C-UNIT condition, enabling time-dependent predictions. Additionally, it is planned to investigate the effect of the initial sample microstructure on reduction and redox processes. The samples with various microstructures will be prepared by modifying the initial powder composition to influence phase fractions, isostatic pressing of the screen-printed anodes to influence porosity, and increasing the sintering temperature to influence particle size.
In the scope of algorithm development, an extension to physics-informed UNIT (PINN-UNIT) is planned to include physical constraints in the loss function, transforming the black-box machine learning approach into a grey-box model. Additionally, further study on the 3D versions of UNIT, C-UNIT, and PINN-UNIT is currently undergoing research.
|
