| 研究課題/領域番号 |
23K13702
|
|---|
| 研究種目 |
若手研究
|
| 配分区分 | 基金 |
|---|
| 審査区分 |
小区分32010:基礎物理化学関連
|
|---|
| 研究機関 | 北海道大学 |
|---|
研究代表者 |
|
|---|
| 研究期間 (年度) |
2023-04-01 – 2026-03-31
|
| 研究課題ステータス |
交付 (2023年度)
|
| 配分額 *注記 |
4,550千円 (直接経費: 3,500千円、間接経費: 1,050千円)
2025年度: 1,040千円 (直接経費: 800千円、間接経費: 240千円)
2024年度: 780千円 (直接経費: 600千円、間接経費: 180千円)
2023年度: 2,730千円 (直接経費: 2,100千円、間接経費: 630千円)
|
|---|
| キーワード | NN Potentials / Reaction Path Search / Machine Learning / Artificial Intelligence / Theoretical Chemistry / Kinetics simulation / NNP / Reaction Design / MLFF / Reaction path network |
|---|
| 研究開始時の研究の概要 |
Innovative chemical reactions are essential tools for humanity's upcoming challenges. This project aims to combine the latest Machine Learning technology (Neural Network Potentials) with efficient reaction path search algorithms, to accelerate the design and discovery of novel chemical reactions.
|
| 研究実績の概要 |
While general-purpose Neural Network Potentials (NNP) can provide very accurate predictions at a much lower cost, these NNPs were found to suffer from robustness issues when applied to force-induced reaction path search. Therefore, combining them with approximate but robust models was demonstrated to produce combined models that are robust, fast and accurate. In particular, the semi-empirical xTB potential was found to yield an appropriate robustness/cost ratio to design combined models adapted for reaction path search.
Using these models, an automated framework was designed, implemented and tested for the AI-driven acceleration of reaction path search, by combining the Artificial Force Induced Reaction (AFIR) method with AI models based on NNPs. Combined with a newly designed training procedure, this recently implemented NNP-AFIR framework now achieves an effective few hundred times acceleration of traditional reaction path searches, with similar accuracy. This development now enables us to perform extensive reaction path searches on much larger systems, allowing to carefully study the influence of substituents on the chemical reactivity, instead of considering only the reaction centers.
As a consequence, this NNP-accelerated AFIR framework is being applied to study several chemical systems that were previously inaccessible.
|
| 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
An NNP-AFIR framework was implemented and deployed on the clusters of the host institute, using a dual CPU server (for DFT calculations) + GPU server (for Neural Network training) design. It appeared that using only Equilibrium States and Transition States, obtained during previous searches, provides enough diversity for training the NNP. The iterative learning scheme (with an improved sweep rehearsal mechanism) developed now leads to a ~500x acceleration compared to traditional DFT-based reaction path search. Thanks to this performance, the NNP-AFIR framework is now being applied on several chemical systems with up to 200+ atoms.
|
| 今後の研究の推進方策 |
A major focus will be the publication of the ongoing applications of the developed NNP-AFIR framework in the context of this project. In particular, the methodology of the NNP-AFIR method will be showcased in a study to solve the ongoing controversy on the Passerini reaction, which has been one of the earliest application of NNP-AFIR, despite its high sensitivity to the AFIR search options. As planned, efforts will be pursued to adapt and improve the NNP-AFIR framework for the current applications as well as the potential new collaborations. Especially, heterogeneous catalysis was recently identified as an application target that could greatly benefit from the NNP-AFIR approach.
|
