A hybrid genetic-Levenberg Marquardt algorithm for automated spectrometer design optimization

• Published in: Ultramicroscopy Vol. 202; pp. 100 - 106
• Main Authors: Cheong, Kang Hao, Koh, Jin Ming
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.07.2019
• Subjects: Computational optimization; Electron optics; Energy analyzer; Genetic algorithm; Hybrid algorithm; Computational optimization; Energy analyzer; Electron optics; Genetic algorithm; Hybrid algorithm
• ISSN: 0304-3991; 1879-2723; 1879-2723
• DOI: 10.1016/j.ultramic.2019.03.004

•Our paper explores a novel hybrid optimization framework that adaptively switches between constituent genetic and Levenberg–Marquardt algorithms.•Our method yields superior convergence performance than either individually.•It solves a longstanding optimization problem in spectrometer design. Advancements in computational tools have driven increasingly automated, simulation-centric approaches in the design and optimization of spectroscopic electron-optical systems. These augmented methodologies accelerate the optimization process, and can yield better-performing instruments. While classical gradient-based methods had been explored, modern alternatives such as genetic algorithms have rarely been applied. In this paper, we propose a novel fully-automated hybrid optimization method for use on electron-optical systems. An adaptive switching scheme between a Levenberg–Marquardt and a genetic sub-algorithm enables the simultaneous exploitation of the computational efficiency of the former and the robustness of the latter. The hybrid algorithm is demonstrated on two test examples—the parallel cylindrical mirror analyzer, and the first-order focusing parallel magnetic sector analyzer—and is found to outperform both the Levenberg–Marquardt and genetic algorithms individually. Our work is significant as a versatile tool for parallel energy spectrometer design, and can greatly aid the development of mechanically-complex parallel energy analyzers, which are expected to be of utility to the semiconductor industry in the near future.