Design of subwavelength wide bandwidth sound absorbers by inverse convolutional neural networks

Abstract

Microperforated panel sound absorber metamaterials are crucial for noise reduction in various applications. This study leverages a convolutional neural network (CNN) machine learning model to optimise these metamaterials for maximum absorption strength and bandwidth range. The model allows for inverse optimisation of sound absorption performance. A desired absorption response can be supplied as input, and the network returns the necessary geometry parameters to achieve the target characteristic.

Metamaterials were optimised to provide over 90 % absorption at target frequencies between 0–1000 Hz. Theoretical predictions were validated experimentally via impedance tube testing. The model achieved no less than 70 % absorption over a 923 Hz range (548–1471 Hz) with a material thickness of 41 mm, and 70 % absorption over 1000 Hz (470–1470 Hz) with a thickness of 57 mm. A case study for an automotive/energy application targeted 50 % absorption between 500–1000 Hz at a thickness of less than 25 mm. Experimental results showed 50 % absorption between 506–1032 Hz at 23 mm thickness. These findings demonstrate the potential of CNN models in optimising sound absorber metamaterials, offering significant improvements in noise reduction with minimal material thickness. The proposed methodology offers significant potential for lightweight applications in various noise-reduction scenarios, including automotive, aerospace, energy, and architectural acoustics.