Influence of Algorithm Parameters of Bayesian Optimization, Genetic Algorithm, and Particle Swarm Optimization on Their Optimization Performance

• Published in: Advanced theory and simulations Vol. 2; no. 10
• Main Authors: Wang, Zhi‐Lei, Ogawa, Toshio, Adachi, Yoshitaka
• Format: Journal Article
• Language: English
• Published: 01.10.2019
• Subjects: Bayesian optimization; data‐driven material design; genetic algorithm; inverse analysis; particle swarm optimization
• ISSN: 2513-0390; 2513-0390
• DOI: 10.1002/adts.201900110

In response to modern materials research, a data‐driven properties‐to‐microstructure‐to‐processing inverse analysis is proposed for use in material design. In the present work, machine learning optimization algorithms of Bayesian optimization, genetic algorithm, and particle swarm optimization are used to perform inverse analysis with a maximum property search. The use of machine learning algorithms readily involves careful tuning of learning parameters, which is often carried out by a trial‐and‐error method requiring expert experience or general guidelines, and the choices of such parameters can play a critical role in attaining good optimization performance. Thus, the influence of various parameters on the optimization performance of the aforementioned algorithms are systematically investigated to provide a protocol for selecting adequate algorithm parameters for a given optimization problem in data‐driven material design. Experiment‐based materials research is becoming increasingly inefficient for discovering new materials because of its time‐consuming and expensive trial‐and‐error methods. Data‐driven properties‐to‐microstructure‐to‐processing inverse analysis performed by machine‐learning optimization algorithms is proposed for use in material design to accelerate the material discovery process.