Multi-objective genetic algorithm optimization of a directionally sensitive radiation detection system using a surrogate transport model

Abstract

Direction-sensitive and imaging radiation detection systems often rely on complicated mask geometries to reconstruct source positions and images. Optimizing the design of these masks is difficult, often non-intuitive, and computationally intensive due to the radiation transport involved. Advances in computational resources, efficient, stochastic optimization techniques, and improved variance reduction in radiation transport through hybrid deterministic-Monte Carlo methods have enabled researchers to consider methods to find formal optimal solutions for these systems. However, many systems span a complex and large design space making full radiation transport simulations to inform the optimization routine computationally intractable. This work applies a multi-objective genetic algorithm to a generalizable surrogate model, benchmarked through full Monte Carlo radiation transport simulations, to determine the optimal design for the rotating scatter mask directionally-sensitive detection system. Performing the equivalent optimization of 10,000 design evaluations with Monte Carlo radiation transport simulations would take 460 CPU-years. In contrast, the 10,000 surrogate design evaluations produced Pareto frontiers comparable to the Monte Carlo results, while reducing the computational cost by 99.998% to 4.1 CPU-days. The chosen optimal designs maintain high directional accuracy with respect to the original design, while improving the response basis similarity by 14% and increasing the efficiency by a factor of 40 at the cost of increasing the mask’s mass by a factor of 3.