Deep Learning for Data-Driven Metamaterial Design Optimization

This work presents a data-driven framework for the design and optimization of elastic metamaterials composed of a hexagonal honeycomb unit cell with embedded cantilever-type resonators. A feed-forward neural network (FFNN) is adopted as a surrogate dynamic model to explore both direct and inverse modeling approaches, with the aim of integrating them into an effective design workflow. The surrogate is then employed to optimize the geometric parameters in order to maximize the bandgap width while enforcing a prescribed central frequency. To generate the training data, two analytical models are developed based on an orthotropic plate formulation: one treating the resonator as a uniform beam with a lumped tip mass, and the other representing it as a two-segment beam. A third dataset is obtained from extensive finite element simulations. The study compares the performance of the FFNN across the three datasets, highlighting how the underlying data source affects the accuracy and generalization of the surrogate model. The optimized design is fabricated using 3D printing and experimentally validated through laser scanning vibrometry, confirming the effectiveness of the proposed framework.