表題番号：2025R-032
日付：2026/04/02
研究課題Exploring Cosmic Ray Propagation with Machine Learning
| 研究者所属（当時） | 資格 | 氏名 | |
|---|---|---|---|
| （代表者） | 理工学術院 国際理工学センター（理工学術院） | 教授 | モッツ ホルガー マルティン |
| （連携研究者） | 理工学術院 理工学術院総合研究所 | 主主任研究員（研究院准教授） | 赤池 陽水 |
| （連携研究者） | 先進理工学部応用物理学科 | 学部４年生 | 櫻井 隆登 |
| （連携研究者） | 先進理工学部物理学科 | 学部４年生 | 平本 亮 |
- 研究成果概要
- This project focused on using machine learning techniques in two tasks:
(1) understand the high dimensional parameter spaces of complex cosmic ray propagation models
(2) identifying and classifying signatures of nearby SNRs in the cosmic-ray electron+positron spectrum
(1) Numerical calculation is used to predict the cosmic ray spectra observed at Earth based on the assumed spectra of accelerated cosmic rays at supernova remnants and the propagation conditions such as the diffusion coefficient which form a large parameter space. Since the numerical calculation is very time consuming, it is difficult to find suitable parameters yielding spectra agreeing to those observed at Earth by systematic or random variation of the input parameters. By training on already calculated results, machine learning algorithms could predict suitable parameters for input in the numerical calculation. As planned, a study of the neural network hyper-parameters was undertaken to improve the prediction precision, showing that the input parameters can be predicted at few percent accuracy.
It ICRC2025, the CALET collaboration showed preliminary results on the cosmic-ray proton/Helium ratio, exhibiting a break around 10 TeV rigidity, which may indicate that the primary cosmic ray sources are divided into two categories with different helium abundances (e.g. early and late phase SNRs). A model based on this two source category hypothesis was implemented in the numerical CR propagation code DRAGON and the parameters optimized making use of the developed machine learning algorithms. It was shown that this model can simultaneously explain the break in the p/He ratio and the most important CR nuclei spectra and secondary to primary ratios, with this result being presented at the JPS spring meeting [4].
(2) While using machine learning in the study of CR propagation was the original focus of the project, it was extended to using it for nearby SNR signature search, since the analysis of CALET electron data revealed a strong dipole anisotropy towards the Vela SNR [2,3]. Identification of a signature from a nearby individual SNR through spectral and anisotropy features is a main goal of the CALET project, with Vela being the prime candidate. To translate the finding in the data into a concrete significance for it being a signature of Vela, previously an unbinned spectral fit in combination with a directional anisotropy search had been developed and evaluated [1]. To supplement and possibly improve over this method, it was investigated if neural networks trained on simulated data could identify the signature. It was shown that the identification works in principle, paving the way for future development in this direction.