Physics-Aware Machine-Learning-Driven Inverse Design of Broadband Ultra-Open Acoustic Metamaterials

Area of Science:

• Acoustics and Materials Science
• Metamaterials Design
• Computational Physics

Background:

• Balancing sound attenuation, bandwidth, openness, and thickness in acoustic silencers is a complex design challenge.
• Advanced noise control requires silencers that offer both high performance and efficient ventilation.

Purpose of the Study:

• To introduce a physics-aware machine learning-driven inverse design framework for broadband ultra-open acoustic silencers (UAS).
• To enable rapid exploration and identification of optimized silencer architectures by decoupling the design space and employing a hybrid-objective inverse strategy.

Main Methods:

• Utilized Green's function parameterization to physically decouple the design space into spectral and radial components.
• Developed a two-stage coarse-to-fine surrogate model to capture broadband and resonant acoustic features.
• Employed a population-based, hybrid-objective parallel (PHP) inverse strategy for efficient design space exploration.

Main Results:

• Identified optimized UAS architectures with broadband bandwidth exceeding 830 Hz within the 1000-2000 Hz range.
• Achieved ultra-thin silencer profiles (0.1-0.2λ, 4-8 cm) with 80% ventilation.
• Validated multiple architectures, including monolithic (UAS-2), multi-mode interference (UAS-3), and parallel-composite (UAS-4) designs.

Conclusions:

• The physics-aware machine learning framework effectively designs high-performance, ultra-open acoustic silencers.
• The study reveals intrinsic relationships between silencer thickness, ventilation, and spectral response, uncovering linear design rules.
• This data-driven paradigm advances the discovery of design principles for functional metamaterials in noise control.