Abstract
Background: The development of computing resources has seen machine learning techniques and models integrated with evolutionary algorithms being successfully applied to solve a vast of engineering problems. And the advance in elastic metamaterial research, the identification of band structure, which reflects the physical property of Elastic Metamaterial, holds the key to the design of wave-controlled devices.
Objective: In order to conduct bandgap analysis on two specific metamaterial structures, machine learning models that are integrated with the evolutionary algorithm are proposed to predict band structure.
Methods: This paper proposes two integration models with a new loss function for predicting elastic metamaterial’s band structure. The self-defined loss function composed of mean square error and concordance correlation is designed to ensure the numerical eigenfrequency values but also the position of each band.
Results: The results of the integration models indicate the MLPs-PSO and RBFs-PSO models indeed have relatively satisfying performances on such pattern recognition tasks with respect to the numerical values of the error measurements. The performances of the machine learning models could be outstandingly improved by the Particle Swarm Optimisation algorithm.
Conclusion: In short, the well-trained machine models are able to predict the band structure and therefore contribute to bandgap enlargement study.
Keywords: Elastic metamaterial, supervised learning models, particle swarm optimisation, finite element method, loss function, machine learning
Graphical Abstract
1