| 研究課題/領域番号 |
22KF0210
|
|---|
| 補助金の研究課題番号 |
22F22708 (2022)
|
|---|
| 研究種目 |
特別研究員奨励費
|
| 配分区分 | 基金 (2023)
補助金 (2022) |
|---|
| 応募区分 | 外国 |
|---|
| 審査区分 |
小区分60100:計算科学関連
|
|---|
| 研究機関 | 京都大学 |
|---|
研究代表者 |
鹿島 久嗣 京都大学, 情報学研究科, 教授 (80545583)
|
|---|
| 研究分担者 |
BARBOT ARMAND 京都大学, 情報学研究科, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2023-03-08 – 2025-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
1,600千円 (直接経費: 1,600千円)
2024年度: 400千円 (直接経費: 400千円)
2023年度: 500千円 (直接経費: 500千円)
2022年度: 700千円 (直接経費: 700千円)
|
|---|
| キーワード | マテリアルズインフォマティクス / 機械学習 / 人工知能 |
|---|
| 研究開始時の研究の概要 |
Nucleation of dislocations (defects in crystals responsible of plasticity) are essential to understand the deformation of nanocrystals. The objective of this project is to use machine learning (ML) to improve simulation methods of crystal plasticity at the mesoscale.
|
| 研究実績の概要 |
This research enhances mesoscopic simulation of dislocation nucleation by integrating machine learning models trained on atomistic data. Initially flawed, the approach was refined into three parts: (1) A deterministic model predicts the first nucleation based on system shape and size. (2) A second model outputs strain interval distributions, informing subsequent nucleation timing. (3) A third model evaluates the likelihood of nucleation considering strain and energy, ensuring realistic simulation outcomes. These advancements were presented at MRS 2023 in Boston and MRM 2023 in Kyoto.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
A set of elemental technologies has been obtained through the research to date, and these have been presented at several international conferences. It is expected that the remaining research period will lead to final results integrating these technologies.
|
| 今後の研究の推進方策 |
The fellow is preparing a paper for a peer-reviewed journal and collaborating with the counterpart lab in France to implement a mesoscopic simulation model. Concurrently, they're developing a Physically Informed Neural Network (PINN) to generate potential energy-strain curves for more effective nucleation criteria. This PINN model inputs system shape and size and outputs potential energy-strain curves, using them as nucleation criteria. The PINN approach, by applying physical constraints, significantly reduces data training needs. It assumes constant slope segments in potential energy-strain curves, specific to each system. Although the current model sometimes overestimates nucleations for certain shapes and sizes, it shows promise, and improvements are underway to enhance its accuracy.
|
