2023 Fiscal Year Final Research Report
Development of experimental data analysis methods using deep learning for LHC and ATLAS
Project/Area Number |
21K13936
|
Research Category |
Grant-in-Aid for Early-Career Scientists
|
Allocation Type | Multi-year Fund |
Review Section |
Basic Section 15020:Experimental studies related to particle-, nuclear-, cosmic ray and astro-physics
|
Research Institution | The University of Tokyo |
Principal Investigator |
Saito Masahiko 東京大学, 素粒子物理国際研究センター, 助教 (70865162)
|
Project Period (FY) |
2021-04-01 – 2024-03-31
|
Keywords | 素粒子物理学 / ATLAS実験 / 機械学習 / 深層学習 |
Outline of Final Research Achievements |
This study focused on and solved several problems in the application of deep learning techniques to the data analysis in high-energy collider experiments. (1) A graph neural network that can reflect the structure of the input data was introduced, improving the performance of physics data analysis. (2) By connecting multiple deep learning models in a way that improves information propagation and introducing a multitask learning method, the performance of the target task was improved while maintaining the interpretability of the overall task. (3) A parameter scan method for physics models using normalising flows was developed to enable efficient search in the parameter space.
|
Free Research Field |
素粒子実験物理学
|
Academic Significance and Societal Importance of the Research Achievements |
本研究では大規模化・複雑化する実験データを解析する上での重要な知見が得られた。このことは、物理データ解析への深層学習技術応用を促進させ、素粒子物理の新たな知見を生み出す可能性を高めることに繋がる。また、本研究のテーマである複数タスク・複雑な入力データ構造や深層学習モデルの解釈可能性の必要性は産業応用上も必要とされることがあり、本研究で得られた知見は深層学習技術の浸透に伴って今後より重要になっていくと期待される。
|