engleski

Application of Support Vector Regression on Neutron Buildup Factors

In this paper the machine learning technique named Support Vector Regression (SVR) method application on neutron buildup factors determination is investigated. When applied on gamma-rays, SVR method shows very good results, leading to the conclusion that there is a potential of using this approach for practical gamma-ray buildup factors determination within established point kernel codes, like QAD-CGGP. A possibility of SVR application on neutron buildup factors estimation has yet to be determined. Neutrons can interact with shielding materials in many ways, including scattering and absorption processes, as well as production of secondary neutrons, protons, gamma-rays or alpha particles. The mentioned by-products can further interact with shielding materials and therefore significantly increase the measured quantity. Additionally produced neutrons can be described with neutron buildup factors. For this study, neutron buildup factors are calculated using approximate formula and are strongly dependent on the shielding material, thickness, and the incident neutron energy range. Typical shielding materials like iron and concrete are observed for neutron energy range 0.5 eV - 14 MeV, and 1 - 10 mean free paths (mfp) thicknesses. Hence, the input vector is multidimensional, comprised of the atomic number of the material, shielding thickness, and the incident neutron energy. Output vector is one-dimensional, presenting target buildup factors. Due to high complexity of neutron transport through shielding material, building the SVR model for neutron buildup factor determination is very time and computer hardware demanding. When regarding the size of the training set, one is confronted with a vast number of potential training points. In order to minimize the number of training points and speed up the training process, active learning measures are applied. By combining various active learning methods, the training set is composed of most informative points, leading to good generalization properties of the model with adequate accuracy.

support vector regression; active learning; neutron buildup factor; reactor shielding

nije evidentirano