Reactor core design is inherently a multi-objective problem which spans a large design space, and potentially larger objective space. This process relies on high-fidelity models to probe the design space, and sophisticated computer codes to calculate the important physics occurring in the reactor. In the past, the design space has been reduced by individuals with extensive knowledge of reactor core design; however, this approach is not always available. In this paper, we utilize a set of high-fidelity models to generate a reduced-order model, and couple this with a genetic algorithm to quickly and effectively optimize a preliminary design for a prototypical sodium fast reactor. We also examine augmenting the genetic algorithm with physical programming to generate the fitness function(s) that evaluates the degree to which a core has been optimized. Physical programming is used in two variations of multi-objective optimization and is compared with a traditional weighting scheme to examine the solutions present on the Pareto front.
Optimization on the reduced-order model produces a set of solutions on the Pareto front for a designer to examine. The uncertainty for the objective functions examined in the reduced-order model is less than 7% for the given designs, and improves as additional data points are employed. Utilizing a reduced-order model can significantly reduce the computation time and storage to perform preliminary optimization. Physical programming was shown to reduce the objective space when compared with a traditional weighting scheme. It also provides an intuitive and computationally efficient way to produce a Pareto front that meets the designer’s objectives.
Experimental design for estimating unknown hydraulic conductivity in an aquifer using a genetic algorithm and reduced order model
