Deep learning optimization of modular neutron beam shutters via Monte Carlo simulations

IF 2.8 3区物理与天体物理 Q3 CHEMISTRY, PHYSICAL

Radiation Physics and Chemistry Pub Date : 2025-09-05 DOI:10.1016/j.radphyschem.2025.113289

Li-Fang Chen

{"title":"Deep learning optimization of modular neutron beam shutters via Monte Carlo simulations","authors":"Li-Fang Chen","doi":"10.1016/j.radphyschem.2025.113289","DOIUrl":null,"url":null,"abstract":"<div><div>This work presents a novel data-driven methodology for optimizing modular neutron beam shutters by integrating Monte Carlo simulations (MCNP) with deep learning techniques. The novelty lies in combining physics-based simulations with machine learning to rapidly screen and identify high-performance shielding designs, reducing both radiation exposure and computational effort. The study targets a compact neutron science platform where a shutter on the thermal neutron line is required to minimize occupational dose during maintenance.</div><div>A dataset of 200 shutter configurations with different material sequences was generated using MCNP and used to train a fully connected neural network for neutron flux prediction. The trained model was then applied to 1,000 additional random designs, allowing rapid performance ranking. The top 20 candidates were re-evaluated by MCNP to verify accuracy.</div><div>Results show that the optimized design attenuates the neutron flux from 5.61 × 10<sup>9</sup> n·cm<sup>-2</sup>·s<sup>-1</sup> at the shutter entrance to 4.96 × 10<sup>5</sup> n·cm<sup>-2</sup>·s<sup>-1</sup> at the exit—achieving a reduction of four orders of magnitude. Compared to the best design among the initial 200 random cases, the machine learning–guided design further improves flux suppression by ∼13 %, while reducing MCNP evaluations by 82 %.</div><div>These findings highlight the application relevance of deep learning–assisted optimization in neutron system design, demonstrating its capability to lower radiation risks and accelerate the development of advanced shielding solutions.</div></div>","PeriodicalId":20861,"journal":{"name":"Radiation Physics and Chemistry","volume":"239 ","pages":"Article 113289"},"PeriodicalIF":2.8000,"publicationDate":"2025-09-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Radiation Physics and Chemistry","FirstCategoryId":"92","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0969806X25007819","RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"CHEMISTRY, PHYSICAL","Score":null,"Total":0}

引用次数: 0

Abstract

This work presents a novel data-driven methodology for optimizing modular neutron beam shutters by integrating Monte Carlo simulations (MCNP) with deep learning techniques. The novelty lies in combining physics-based simulations with machine learning to rapidly screen and identify high-performance shielding designs, reducing both radiation exposure and computational effort. The study targets a compact neutron science platform where a shutter on the thermal neutron line is required to minimize occupational dose during maintenance.

A dataset of 200 shutter configurations with different material sequences was generated using MCNP and used to train a fully connected neural network for neutron flux prediction. The trained model was then applied to 1,000 additional random designs, allowing rapid performance ranking. The top 20 candidates were re-evaluated by MCNP to verify accuracy.

Results show that the optimized design attenuates the neutron flux from 5.61 × 10⁹ n·cm^-2·s^-1 at the shutter entrance to 4.96 × 10⁵ n·cm^-2·s^-1 at the exit—achieving a reduction of four orders of magnitude. Compared to the best design among the initial 200 random cases, the machine learning–guided design further improves flux suppression by ∼13 %, while reducing MCNP evaluations by 82 %.

These findings highlight the application relevance of deep learning–assisted optimization in neutron system design, demonstrating its capability to lower radiation risks and accelerate the development of advanced shielding solutions.

查看原文本刊更多论文

基于蒙特卡罗模拟的模块化中子束百叶窗的深度学习优化

这项工作提出了一种新的数据驱动方法，通过将蒙特卡罗模拟（MCNP）与深度学习技术相结合，来优化模块化中子束百叶窗。其新颖之处在于将基于物理的模拟与机器学习相结合，以快速筛选和识别高性能屏蔽设计，减少辐射暴露和计算工作量。该研究的目标是一个紧凑的中子科学平台，在该平台上需要热中子线上的百叶窗，以尽量减少维护期间的职业剂量。利用MCNP生成了包含200种不同材料序列快门配置的数据集，并用于训练全连接神经网络进行中子通量预测。然后将训练好的模型应用到1,000个额外的随机设计中，从而实现快速的性能排名。前20名候选人由MCNP重新评估以验证准确性。结果表明，优化后的设计使中子通量从入口处的5.61 × 109 n·cm-2·s-1衰减到出口的4.96 × 105 n·cm-2·s-1，降低了4个数量级。与最初200个随机案例中的最佳设计相比，机器学习引导的设计进一步提高了~ 13%的通量抑制，同时将MCNP评估降低了82%。这些发现突出了深度学习辅助优化在中子系统设计中的应用相关性，展示了其降低辐射风险和加速先进屏蔽解决方案开发的能力。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Radiation Physics and Chemistry 化学-核科学技术

CiteScore

5.60

自引率

17.20%

发文量

574

审稿时长

12 weeks

期刊介绍： Radiation Physics and Chemistry is a multidisciplinary journal that provides a medium for publication of substantial and original papers, reviews, and short communications which focus on research and developments involving ionizing radiation in radiation physics, radiation chemistry and radiation processing. The journal aims to publish papers with significance to an international audience, containing substantial novelty and scientific impact. The Editors reserve the rights to reject, with or without external review, papers that do not meet these criteria. This could include papers that are very similar to previous publications, only with changed target substrates, employed materials, analyzed sites and experimental methods, report results without presenting new insights and/or hypothesis testing, or do not focus on the radiation effects.