Acoustic Lens Design Using Machine Learning

Abstract

This thesis aims to contribute to the development of a novel approach and efficient method for the inverse design of acoustic metamaterial lenses using machine learning, specifically, deep learning, generative modeling, and reinforcement learning. Acoustic lenses can focus incident plane waves at the focal point, enabling them to detect structures non-intrusively. These lenses can be utilized in biomedical engineering, medical devices, structural engineering, ultrasound imaging, health monitoring, etc. Finding the global optimum through a traditional iterative optimization process for designing the acoustic lens is challenging. It may become infeasible due to high dimensional parameter space and the compute resources needed. Machine learning techniques have been shown promising for finding the global optimum. Generative modeling is a powerful technique enabling recent advancements in drug discoveries, organic molecule development, and photonics. We combined generative modeling with global optimization and an analytical form of gradients computed by means of multiple scattering theory. In addition, reinforcement learning can potentially outperform traditional optimization algorithms. Thus, in this thesis, the acoustic lens is modeled using two machine learning techniques, such as generative modeling, using 2D-Global Topology Optimization Networks (2D-GLOnets), and reinforcement learning using the Deep Deterministic Policy Gradient (DDPG) algorithm. Results from the aforementioned methods are compared with traditional optimization algorithms.