Population-based metaheuristic optimization in neutron optics and shielding design

TL;DR

This paper compares various population-based metaheuristic algorithms, including PSO, DE, ABC, and GA, for optimizing neutron optics and shielding design, introducing a new software tool and evaluating their performance through simulations.

Contribution

It develops a unified software platform for applying multiple metaheuristics to neutron design problems and provides a comparative analysis of their effectiveness.

Findings

Differential Evolution (DE) generally outperformed other algorithms.

Algorithm performance varied depending on specific problem scenarios.

The software facilitates easy comparison of metaheuristic algorithms in neutron design.

Abstract

Population-based metaheuristic algorithms are powerful tools in the design of neutron scattering instruments and the use of these types of algorithms for this purpose is becoming more and more commonplace. Today there exists a wide range of algorithms to choose from when designing an instrument and it is not always initially clear which may provide the best performance. Furthermore, due to the nature of these types of algorithms, the final solution found for a specific design scenario cannot always be guaranteed to be the global optimum. Therefore, to explore the potential benefits and differences between the varieties of these algorithms available, when applied to such design scenarios, we have carried out a detailed study of some commonly used algorithms. For this purpose, we have developed a new general optimization software package which combines a number of common metaheuristic algorithms within a single user interface and is designed specifically with neutronic calculations in mind. The algorithms included in the software are implementations of Particle-Swarm Optimization (PSO), Differential Evolution (DE), Artificial Bee Colony (ABC), and a Genetic Algorithm (GA). The software has been used to optimize the design of several problems in neutron optics and shielding, coupled with Monte-Carlo simulations, in order to evaluate the performance of the various algorithms. Generally, the performance of the algorithms depended on the specific scenarios, however it was found that DE provided the best average solutions in all scenarios investigated in this work.