2023 Fiscal Year Research-status Report
水チェレンコフ実験のための新しい機械学習ベースのイベントの再構築と分析
| Project/Area Number |
22KF0113
|
|---|
| Allocation Type | Multi-year Fund |
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
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)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
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
|
|---|
| Project Period (FY) |
2023-03-08 – 2025-03-31
|
| Keywords | Deep Learning / Neural networks / Water Cherenkov / Neutrino / Particle Physics / Cosmology |
|---|
| Outline of Annual Research Achievements |
Accurately quantifying systematic uncertainties in water Cherenkov experiments poses a significant challenge. The conventional physics analysis pipeline often relies on empirically derived assumptions, leading to separate calibrations targeting various potential effects. While this approach has yielded insights, its time-consuming nature can limit the timely implementation of analysis upgrades. Moreover, it lacks the necessary adaptability to accommodate discrepancies arising from asymptotic inputs and factorized physics processes.
I am building a differentiable physics emulator to enhance the estimations of detector systematic uncertainties and advance physics inference. By developing a novel analysis pipeline based on AI/ML technique, I constructed a physics-based detector model that is optimizable with calibration data in contrast to the traditional Monte Carlo simulation. It can infer convoluted detector effects using a single differentiable model, informed by basic yet robust physics knowledge inputs. The scalability of an AI/ML model surpasses that of conventional methods.
I am also working with cosmologists from domestic and foreign institutions to construct a cosmology simulation for the possible realizations for both the Large Scale Structure (LSS) and Cosmic Microwave Background (CMB). In this work I am attempting multiple statistical inference methods to predict the possible radio emission foreground based on the observed galactic catalogs, which is essential for the determination of the CMB “B-mode” polarization. This work is expected to be concluded in 2024.
|
| Current Status of Research Progress |
Current Status of Research Progress
4: Progress in research has been delayed.
Reason
The original plan was to develop a deep-learning-based event calibration/reconstruction method for water Cherenkov experiments and test it with the data from the Water Cherenkov Test Experiment (WCTE) at CERN, which was expected to be available by mid-2024. However, there was an unforeseeable delay on the construction of WCTE due to the prolonged manufacturing procedure of the water tank in Spain. It is still unclear now whether WCTE can take data in 2024. Instead, I have focused more on software development and exploring different ideas while waiting for the WCTE data to test the tools under development. I am also participating in the construction and commissioning of WCTE to help with the already delayed schedule.
|
| Strategy for Future Research Activity |
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 the various novel detector calibration techniques that will be tested and applied to the next generation experiment like Hyper-K. Another is the development and testing of new particle reconstruction algorithms thanks to the tagged CERN beam. Thus I am working to apply a novel neural network model for the calibration and reconstruction and define a new analysis pipeline. The key of this novel work is that it is fully differentiable and has the initial model input optimizable, for instance constants for the detector condition, by the gradient-descent method using calibration data. It will also serve as a validation for the application of deep learning techniques into larger water Cherenkov detectors such as SK and HK.
Next, I plan to expand this neural network to a broader application across various neutrino experiments and construct a first foundation model for experimental neutrino physics. The current model is a simple multi-layer-perceptron architecture, with differentiable activation functions in the full phase space. It uses a supervised learning strategy with labeled data, and thus cannot solve the problem of “unknown bias” in the training inputs, e.g. the Monte Carlo simulation. More advanced techniques such as unsupervised learning, widely used in the Large Language Models in recent years, will make it possible to understand the physical causal relationships and detect anomalies in the input simulation.
|
| Causes of Carryover |
This small amount of leftover funds, 11,260 yen from FY23, was the result of imprecision in anticipating travel expenses; airline ticket prices have been fluctuating wildly in the aftermath of COVID. It will be put toward travel expenses in the current fiscal year.
|
