| 研究課題/領域番号 |
22F22333
|
|---|
| 配分区分 | 補助金 |
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
ヴァギンズ マーク 東京大学, カブリ数物連携宇宙研究機構, 教授 (90509902)
|
|---|
| 研究分担者 |
XIA JUNJIE 東京大学, カブリ数物連携宇宙研究機構, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2022-11-16 – 2025-03-31
|
| キーワード | Deep learning / Neural networks / Water Cherenkov / Neutrino / Particle Physics / Cosmology |
|---|
| 研究実績の概要 |
I have modified the pipeline of current deep learning based water Cherenkov event generative neural network (CRinGe, arXiv:2202.01276v1). Thus the computational speed is improved by a factor of 5 and the training process can be finished within days instead of weeks. Different architectures and loss functions have been implemented for this neural network to achieve better numerical stability and physical robustness. In the meantime I am developing the Monte-Carlo simulation and analysis pipeline for the upcoming Water Cherenkov Test Experiment in CERN, for which the first test run will start in July 2023.
Collaborating with M.Mandal from Poland National Centre for Nuclear Research, I have developed, verified, and implemented a new event selection criterion for the T2K and Super-Kamiokande-T2K (SK-T2K) joint neutrino oscillation analysis to exclude the potential neutron background events in the new SK detector with Gadolinium. We constrained the impurity in the selected events to less than 1%. Meanwhile I am also fulfilling my responsibility of updating and maintaining the T2K data taking and reduction pipeline to adapt to the new SK detector and more powerful T2K neutrino beam.
Besides, I have established a collaboration with the cosmologists from both domestic and foreign institutions to investigate the possibility of common deep learning techniques for particle physics and cosmology research. Taking this chance I have contributed to the foundation and inauguration of the new Center of Data-Driven Discovery (CD3, https://cd3.ipmu.jp/) at Kavli IPMU, the University of Tokyo.
|
| 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
The original plan was to develop a deep learning based event reconstruction method for water Cherenkov experiments and test it with the data from Water Cherenkov Test Experiment (WCTE) at CERN, which is expected to be available by mid-2024. So far various improvements have been achieved in the model and the preparation of WCTE is underway. Besides, the interaction with machine learning experts in other fields has inspired new ideas for this project.
|
| 今後の研究の推進方策 |
WCTE is a 50-ton water Cherenkov experiment at CERN with rich physics potential. One of the critical contributions it can bring to the neutrino physics community is to understand the hadronic interactions of charged pions in water, especially those low energy ones below the Cherenkov threshold. This contributes to one of the dominant uncertainties in neutrino oscillation experiments including SK and T2K. Thus I am working to apply a generative neural network model for this reconstruction and define an analysis pipeline. It will also serve as a validation for the application of deep learning techniques into larger water Cherenkov detectors such as SK and Hyper-K.
Meanwhile, further developments in this generative neural network architecture and analysis pipeline that can include different networks are required. The current generative model is a simple convolutional encoder and decoder architecture. More advanced architectures such as Bayesian probabilistic models or differential models will be especially interesting and beneficial since it is easier to understand the physical causal relationships and analysis uncertainties with these models.
Furthermore, given the productive history of deep learning techniques in other fields such as cosmology, I look forward to continuing the collaboration with the experts to develop better deep learning techniques that can be capable of various scientific tasks.
|
