Advanced Liquid-Metal-Cooled Fast Reactor Core Design Using Modern Optimization Algorithms

Abstract

When performing core loading pattern design in fast spectrum reactors, it is often assumed that the larger neutron mean free path in fast reactors makes the core loading pattern less significant than in thermal reactors. Due to this assumption, the literature often includes homogeneous core designs for liquid-metal-cooled fast reactors (LMFRs). In this paper, heterogeneous loading patterns are investigated using modern LMFR multiphysics analysis. It was found that the figures of merit (FOMs) used to design cores are very sensitive to the core loading pattern, and better core designs (measured by the FOM) can be obtained from heterogeneous designs. Based on this investigation, a need for a modern LMFR core loading pattern design methodology is identified and developed. The methodology is demonstrated through the optimization of the core loading pattern of a sodium-cooled fast reactor. The reactor design used in this demonstration is based on the Super Power Reactor Innovative Small Modular (SPRISM) core, and an optimized loading pattern is obtained by searching for the fuel enrichment and locations of the driver and blanket assemblies. The objectives of the search are to reduce the fuel cost and peaking factors, while meeting the design constraints. To calculate the fuel cost, a preliminary cost model is developed and applied for transuranium fuel loading. Upon establishing the methodology, six optimization algorithms are tested for their effectiveness in solving the LMFR core loading pattern problem. Four of the algorithms are population based: Gray Wolf Optimization, Salp Swarm Optimization, Whale Optimization Algorithm, and the Moth Flame Optimization. A reinforcement learning algorithm, called the proximal policy optimization, was also selected. Finally, the differential evolution (DE) algorithm was selected as the choice of an evolutionary-based algorithm. All algorithms showed a competitive converged design. However, the DE showed more favorable performance since it was able to converge to a superior design compared to the rest of the algorithms with a reasonable number of design sample evaluations, in addition to avoiding local minima entrapment.