Optimal design of one-dimensional elastic metamaterials through deep convolutional neural network and genetic algorithm

Abstract

Deep learning (DL) has found extensive application in elastic metamaterial design. However, these methods are often limited to fixed design parameters and single structural configuration, lacking versatility. To overcome these, we propose a novel data-driven design framework combining deep convolutional neural network (DCNN) and genetic algorithm (GA) to design vibration-isolating metamaterial structures in foundations. This framework is highly adaptable, accommodating diverse design parameters by identical models. We discretize different structural configurations into a uniform tensor format, and generate multiple datasets. Two DCNN-based forward prediction models are then developed to accurately capture bandgap distributions for all body waves. Additionally, we incorporate multiple populations to enhance the parallel search capacity of GA, and integrate it with the trained DCNN models to simultaneously determine multiple optimal structures. The optimal design result for different parameters shows that the designed structures can achieve the target bandgap, and proves the effectiveness and generality of our method. Finally, through the design of two structural configurations, we find that four-layer structure exhibits wider low-frequency bandgaps and superior vibration reduction performance compared to the three-layer structure, under the same material composition and usage.