Seismic metamaterial design prediction based on joint neural network

SMs (SMs), artificial periodic composites utilized to mitigate seismic hazards by attenuating seismic waves within specific frequency bands, have garnered significant research interest in recent years. To expedite the determination of optimal structures within a limited design space, the joint neural network (JNN) incorporating a Depth Feedforward Network (DFN) and an Undercomplete Autoencoder (UAE) in series was devised. A dual-component SMs dataset was generated using curve functions for material parameter analysis. The UAE was trained to discern crucial features of SMs configurations. Dispersion curves for the sample dataset were computed using finite element method (FEM). Employing interval merging algorithm and normalized frequency labeling, data underwent dimensionality reduction and feature extraction, revealing that the pre-trained JNN exhibited an error less than 0.2% compared to FEM, with a design time of merely 40 s. Adhering to the principles of optimal bandgaps and periodic symmetry, SMs were designed, combined, and subjected to frequency domain analysis, achieving an ultra-wide bandgap of 4.9-20 Hz. Inputting Helena Montana-02 and Chi-Chi seismic waveforms demonstrated reductions in peak seismic amplitudes by 52.6% and 72.2% respectively, validating the efficacy of the design model.