A hybrid global surrogate modeling software for nuclear reactor cross section estimation

Nuclear fuel cycle (NFC) simulators track the amount and composition of materials as they move through facilities such as mines, fuel fabrication plants, and nuclear reactors. A major task of a NFC simulator is to calculate the evolution of compositions of batches of nuclear materials as they are transmuted in reactors, decay, and are blended with other batches to create reactor fuel or be reprocessed or disposed. Codes used for NFC simulation that utilize intermediate data saved in databases which are calculated ahead of time are attractive since their fidelity can be improved by investing more resources in expanding their databases. Shifting the computational work ahead of the reactor simulation like this allows the fidelity to be improved without sacrificing runtime computational cost. This dissertation describes a method that attempts to maximize the fidelity increase per unit time invested during this precomputation step. Unlike previous work in the reactor simulation field, this methodology does not limit the number and type of runtime simulation inputs. NUDGE (NUclear Database GEneration software) is an implementation of this methodology. The methodology has two main steps where new data is added to databases. First is exploration, where inputs to the database are selected to be as uniformly distributed as possible within the problem input domain. Second step is exploitation, where output information is utilized to inform the selection of the next point to run. An improvement to exploitation, named Voronoi Cell Adjustment, is described in this dissertation and implemented in NUDGE. This improvement has been shown to benefit the average fidelity increase during database building. A study of the scaling of the methodology, a comparison of error metrics, and an exploration of optimal values for several key parameters in the methodology are presented. NUDGE has also been used to create a global surrogate model of a NFC simulation software (named XSgen). This model shows better performance compared to models generated by other established methods under equal constraints.