Enhancing Acoustic Metamaterial Performance and Fabrication Robustness with Deep Learning

Abstract

Conventional acoustic materials, such as porous or fibrous absorbers, are limited in their ability to achieve low-frequency sound attenuation within compact volumes. Acoustic metamaterials (AMMs) offer a promising alternative, enabling extreme subwavelength resonances and tunable acoustic performance. However, designing AMMs that are simultaneously high-performing, fabrication-robust, and physically interpretable remains a major challenge. Additive manufacturing (AM) allows complex geometries to be realised, but introduces imperfections that can degrade acoustic performance, while traditional design methods often provide limited insight into the underlying structure�performance relationships.

In this thesis, robustness is defined as the ability of an AMM design or predictive model to maintain consistent acoustic performance in the presence of manufacturing tolerances, modeling uncertainties, and experimental variability. Interpretability refers to the capacity to extract physically meaningful insights from data-driven models, enabling designers to understand how geometric features influence acoustic response rather than treating models as black boxes.

This thesis presents a comprehensive framework for the design, optimisation, and robust realisation of AMM absorbers using deep learning (DL) approaches, with an emphasis on interpretability and real-world manufacturability. An AMM absorber, termed the Hybrid Resonant Acoustic Meta-Absorber (HRAM), consisting of an Archimedean spiral embedded in a hexagonal cavity with a perforated top plate, was developed and successfully validated through analytical modeling, high-fidelity numerical simulations, and experimental testing. The analytical model, while sufficient for the subsequent data-driven studies, showed observable discrepancies when compared to experimental results especially in terms of bandwidth characteristic. The high-fidelity numerical simulation also shows observable bandwidth mismatches relative to experimental curves. These observable deviations were largely attributed to unmodeled physical effects in the absorber behaviour, manufacturing issues and the unrealistic assumption of sound-hard boundaries in numerical simulations. Furthermore, due to the computational demand of the full-wave (FW) numerical simulation, a simplified numerical model (SNM) was developed which significantly reduced the simulation time from from the order of hours to minutes, and this simplified model was subsequently used in all references to numerical results.

Using the validated analytical model, a representative spectrum�geometry dataset was generated and employed to compare traditional optimisation methods, exemplified by a generalised pattern search (GPS), with deep learning (DL) approaches. A structured latent autoencoder (SLAE) was trained for both forward and inverse design tasks, and its performance was enhanced using on-the-fly data augmentation (artificially increasing the size and diversity of the training dataset by applying spectral transformations to existing data) and a transfer learning (TL) strategy incorporating sparse experimental data to capture manufacturing variability. The resulting SLAE-TL model demonstrated efficient, robust, and predictive performance, bridging the gap between theoretical design and real-world behaviour.

Beyond predictive capability, a novel design-space segmentation approach was developed to provide interpretability, revealing the relationship between geometric features and acoustic performance. This framework was applied to the design of a broadband AMM absorber for room acoustics, demonstrating both practical utility and scientific insight.

Collectively, the work establishes a unified methodology for the design of robust, interpretable, and high-performance AMMs. By integrating physics-based modelling, deep learning, and experimental validation, this research addresses long-standing challenges in AMM design and paves the way for real-world deployable solutions, advancing both the scientific understanding and practical implementation of AMMs.