Scholarlink

Abstract <p>Radiation monitoring outside accelerator shielding is performed using the Bonner multi-sphere spectrometer, which detects neutrons across a wide energy range. Since neutron spectra cannot be directly measured, computational unfolding methods are required. In this study, neutron spectrum deconvolution models were developed using the LightAutoML and FEDOT automated machine learning frameworks, optimizing hyperparameters and ensemble combinations of multiple algorithms. For training and validation a synthetic spectra were generated as weighted combinations of thermal, epithermal, fast and high-energy component. The models were evaluated on a test set of 375 real spectra, with uncertainty assessed via Monte Carlo simulations incorporating random input errors. Reconstruction quality was measured using Spearman and Pearson correlation, cosine similarity, cross-entropy, Wasserstein distance, Kullback–Leibler divergence, maximum mean discrepancy and <i>R</i><sup>2</sup>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated Machine Learning for Neutron Spectrometry

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • K. A. Chizhov,
  • A. A. Bely,
  • M. D. Starikovskaya

摘要

Abstract

Radiation monitoring outside accelerator shielding is performed using the Bonner multi-sphere spectrometer, which detects neutrons across a wide energy range. Since neutron spectra cannot be directly measured, computational unfolding methods are required. In this study, neutron spectrum deconvolution models were developed using the LightAutoML and FEDOT automated machine learning frameworks, optimizing hyperparameters and ensemble combinations of multiple algorithms. For training and validation a synthetic spectra were generated as weighted combinations of thermal, epithermal, fast and high-energy component. The models were evaluated on a test set of 375 real spectra, with uncertainty assessed via Monte Carlo simulations incorporating random input errors. Reconstruction quality was measured using Spearman and Pearson correlation, cosine similarity, cross-entropy, Wasserstein distance, Kullback–Leibler divergence, maximum mean discrepancy and R2.