Parichehr Khalilzadeh , S. Farhad Masoudi , Hassan Yousefnia , Fatemeh S. Rasouli
{"title":"基于机器学习的医学相关溴放射性同位素质子诱导反应截面预测","authors":"Parichehr Khalilzadeh , S. Farhad Masoudi , Hassan Yousefnia , Fatemeh S. Rasouli","doi":"10.1016/j.rinp.2025.108357","DOIUrl":null,"url":null,"abstract":"<div><div>Accurate prediction of nuclear reaction cross-sections is essential for optimizing the production of medically relevant radioisotopes. In this study, we investigate proton-induced reactions on selenium isotopes leading to the formation of bromine radioisotopes (<sup>75</sup>Br, <sup>76</sup>Br, <sup>77</sup>Br, and <sup>80m</sup>Br) using machine learning (ML) techniques. A total of 499 experimental data points were extracted from the EXFOR database and processed using Savitzky–Golay filtering and cubic spline interpolation to reduce noise and augment the dataset. Feature engineering incorporated both categorical and physics-informed variables to enhance model learning. Three ML models—Random Forest (RF), LightGBM, and Gaussian Process Regression (GPR)—were trained and evaluated using RMSE, R<sup>2</sup>, and 10-fold cross-validation. The results show that RF consistently outperformed other models, achieving an RMSE of 16.73 mb and R<sup>2</sup> of 0.99 post-preprocessing. LightGBM also performed robustly, while GPR showed limited stability in sparse or nonlinear regions. Model performance was further analyzed by reaction type and radioisotope. The study highlights the effectiveness of ML and preprocessing in cross-section prediction under data-scarce conditions and suggests future directions including deep learning and hybrid ensemble models.</div></div>","PeriodicalId":21042,"journal":{"name":"Results in Physics","volume":"75 ","pages":"Article 108357"},"PeriodicalIF":4.6000,"publicationDate":"2025-07-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Machine learning-based prediction of proton-induced reaction cross-sections for medically relevant bromine radioisotopes\",\"authors\":\"Parichehr Khalilzadeh , S. Farhad Masoudi , Hassan Yousefnia , Fatemeh S. Rasouli\",\"doi\":\"10.1016/j.rinp.2025.108357\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Accurate prediction of nuclear reaction cross-sections is essential for optimizing the production of medically relevant radioisotopes. In this study, we investigate proton-induced reactions on selenium isotopes leading to the formation of bromine radioisotopes (<sup>75</sup>Br, <sup>76</sup>Br, <sup>77</sup>Br, and <sup>80m</sup>Br) using machine learning (ML) techniques. A total of 499 experimental data points were extracted from the EXFOR database and processed using Savitzky–Golay filtering and cubic spline interpolation to reduce noise and augment the dataset. Feature engineering incorporated both categorical and physics-informed variables to enhance model learning. Three ML models—Random Forest (RF), LightGBM, and Gaussian Process Regression (GPR)—were trained and evaluated using RMSE, R<sup>2</sup>, and 10-fold cross-validation. The results show that RF consistently outperformed other models, achieving an RMSE of 16.73 mb and R<sup>2</sup> of 0.99 post-preprocessing. LightGBM also performed robustly, while GPR showed limited stability in sparse or nonlinear regions. Model performance was further analyzed by reaction type and radioisotope. The study highlights the effectiveness of ML and preprocessing in cross-section prediction under data-scarce conditions and suggests future directions including deep learning and hybrid ensemble models.</div></div>\",\"PeriodicalId\":21042,\"journal\":{\"name\":\"Results in Physics\",\"volume\":\"75 \",\"pages\":\"Article 108357\"},\"PeriodicalIF\":4.6000,\"publicationDate\":\"2025-07-10\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Results in Physics\",\"FirstCategoryId\":\"101\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S2211379725002517\",\"RegionNum\":2,\"RegionCategory\":\"物理与天体物理\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"MATERIALS SCIENCE, MULTIDISCIPLINARY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Results in Physics","FirstCategoryId":"101","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2211379725002517","RegionNum":2,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"MATERIALS SCIENCE, MULTIDISCIPLINARY","Score":null,"Total":0}
Machine learning-based prediction of proton-induced reaction cross-sections for medically relevant bromine radioisotopes
Accurate prediction of nuclear reaction cross-sections is essential for optimizing the production of medically relevant radioisotopes. In this study, we investigate proton-induced reactions on selenium isotopes leading to the formation of bromine radioisotopes (75Br, 76Br, 77Br, and 80mBr) using machine learning (ML) techniques. A total of 499 experimental data points were extracted from the EXFOR database and processed using Savitzky–Golay filtering and cubic spline interpolation to reduce noise and augment the dataset. Feature engineering incorporated both categorical and physics-informed variables to enhance model learning. Three ML models—Random Forest (RF), LightGBM, and Gaussian Process Regression (GPR)—were trained and evaluated using RMSE, R2, and 10-fold cross-validation. The results show that RF consistently outperformed other models, achieving an RMSE of 16.73 mb and R2 of 0.99 post-preprocessing. LightGBM also performed robustly, while GPR showed limited stability in sparse or nonlinear regions. Model performance was further analyzed by reaction type and radioisotope. The study highlights the effectiveness of ML and preprocessing in cross-section prediction under data-scarce conditions and suggests future directions including deep learning and hybrid ensemble models.
Results in PhysicsMATERIALS SCIENCE, MULTIDISCIPLINARYPHYSIC-PHYSICS, MULTIDISCIPLINARY
CiteScore
8.70
自引率
9.40%
发文量
754
审稿时长
50 days
期刊介绍:
Results in Physics is an open access journal offering authors the opportunity to publish in all fundamental and interdisciplinary areas of physics, materials science, and applied physics. Papers of a theoretical, computational, and experimental nature are all welcome. Results in Physics accepts papers that are scientifically sound, technically correct and provide valuable new knowledge to the physics community. Topics such as three-dimensional flow and magnetohydrodynamics are not within the scope of Results in Physics.
Results in Physics welcomes three types of papers:
1. Full research papers
2. Microarticles: very short papers, no longer than two pages. They may consist of a single, but well-described piece of information, such as:
- Data and/or a plot plus a description
- Description of a new method or instrumentation
- Negative results
- Concept or design study
3. Letters to the Editor: Letters discussing a recent article published in Results in Physics are welcome. These are objective, constructive, or educational critiques of papers published in Results in Physics. Accepted letters will be sent to the author of the original paper for a response. Each letter and response is published together. Letters should be received within 8 weeks of the article''s publication. They should not exceed 750 words of text and 10 references.