Phase-Informed Deep Learning for Automated nTOF Unfolding
Ronghao Yang¹, Yihan Wang², Long Zeng¹, Bolun Chen², Zifeng Song², Yuchi Wu^2
1 Department of Engineering Physics, Tsinghua University, Beijing 100084, China
² Research Center of Laser Fusion, China Academy of Engineering Physics, Mianyang 621900, China
yrh21@mails.tsinghua.edu.cn

A physics-guided deep learning method is developed for automated unfolding of neutron time-of-flight (nTOF) spectra at inertial confinement fusion (ICF) facilities such as Shenguang (SG). Accurate spectrum unfolding is essential for reliable extraction of hot-spot ion temperature and evaluation of fusion implosion performance. Conventional methods mainly rely on forward fitting [1] or frequency-domain deconvolution [2]. While computationally efficient, deconvolution tends to amplify high-frequency noise, and the subsequent manual selection of filtering thresholds introduces subjective uncertainty, especially under low-signal-to-noise conditions.

Accordingly, an objective unfolding framework based on deep learning is developed. To overcome the lack of ground truth data, a physics-driven synthetic dataset covering both DD and DT reactions is constructed. Preliminary results show that purely data-driven models trained only with amplitude-spectrum mean squared error exhibit limited generalization due to severe aliasing between signal and noise.
(Please refer to the uploaded PDF for Figure1)
Addressing this, prior physical knowledge from the ideal Gaussian neutron spectrum [3] is incorporated into the learning objective. Specifically, the first-order phase difference of valid signals varies smoothly, whereas deconvolution-induced high-frequency noise produces irregular oscillations. Guided by this, a physics- constrained loss function is designed to train a network incorporating the phase spectrum. Evaluations indicate that integrating the local variance of the phase difference helps reduce the boundary ambiguity faced by amplitude-only models and achieves a more rational balance between signal fidelity and noise reduction. Compared with manual processing, the proposed method provides a more consistent and automated solution for nTOF spectrum unfolding, showing strong potential for application to data analysis at SG and related ICF facilities.

[1] R. Hatarik et al.,  Rev. Sci. Instrum. 86, 073503 (2015).
[2] Q. Tang et al., " High Power Laser Part. Beams 25, 3153–3157 (2013); E.-F. Guo et al., J. Inst. 14, P05018 (2019).
[3] H. Brysk, Plasma Phys. 15, 611–617 (1973).
 

Neutron Time-of-Flight,Deep Learning,Deconvolution,Phase Spectrum Analysis