Development of machine learning-based nuclear data estimator and its application to nuclear data evaluation

Iwamoto, Hiroki   

Nuclear data are indispensable for the simulation of particle transport when estimating radiation doses, radioactive isotope production, and safety parameters of nuclear systems. The quality and quantity of nuclear data in the evaluated nuclear data libraries, such as JENDL, ENDF, and JEFF, have been steadily enhanced owing to physicists' strenuous efforts; however, discrepancies remain between experimental data and the physics model calculations particularly at high incident energies. Although it is necessary to improve the physics models implemented in nuclear data evaluation code or Monte Carlo-based physics models, conventional approaches require significant time and effort. To address this issue, we have developed a machine learning-based nuclear data estimator, GHyND. G-HyND estimates cross-sections with Gaussian processes from a training dataset composed of experimental data and analytical data based on a physics model calculation. For evaluation of nuclear data in JENDL, proton-induced nuclide production cross-sections were estimated for energies from threshold energy to 3 GeV using G-HyND. As the training dataset, experimental data compiled in the EXFOR database and those recently measured at J-PARC, and analytical data calculated by the Monte Carlo-based nuclear reaction model (INCL4.6 and GEM) were employed. Part of the estimated nuclear data for incident energies up to 200 MeV were compiled in the next release of JENDL.