Understanding how acoustic waves interact with complex structures lies at the heart of inverse design for acoustic metamaterials. While numerous analytical and computational techniques have advanced the design of periodic and homogenizable metamaterials, the characterization of finite metamaterials remains challenging due to diffraction and complex scattering patterns that emerge at their boundaries. Recently, AI/ML techniques have shown promise in estimating effective properties of metamaterial structures, but few address the inverse problem of inferring material geometry directly from acoustic responses—a crucial step for data-driven inverse design. In this talk, we present a data-driven framework that leverages Convolutional Neural Networks (CNNs) to infer the geometry of finite 3-D homogeneous objects of known effective material properties from scattered ultrasound echoes. Our approach involves generating synthetic echoes from various object shapes and training shape-specific CNNs to classify real, measured echoes obtained from physical experiments. This simulation-to-reality learning paradigm demonstrates the ability to infer geometric information from complex acoustic responses, without requiring extensive real-world data. These capabilities offer a promising foundation for AI-driven design and characterization of unit cell geometries and finite acoustic metamaterials.
Patil et al. (Wed,) studied this question.