Inverse design of bending channel sound-absorbing structures with porous material by two-stage deep neural network model

Abstract Acoustic metamaterials with embedded porous are widely used in noise reduction applications. The acoustic properties of these structures are usually characterized by geometrical parameters of the structure and physical parameters of the porous material. In order to realize the perfect acoustic absorption unit at a specific frequency, the structural parameters need to be adjusted precisely, which brings a complicated workload to the researchers. To solve this problem, this paper proposes a machine learning method, Through the sensitivity analysis of sound absorption unit, the effective variables for constructing machine learning database are extracted, and an autoencoder two-stage deep neural network model (TSDNN) is constructed. Based on database distribution characteristics, the performance parameters can be divided into a non-uniform learning area in the mid-low frequency band (350 Hz–1050 Hz), a transfer learning area with sparse data distribution (1050 Hz–1300 Hz), and a uniform learning area in the high frequency band (1300 Hz–4500 Hz). Selecting four acoustic absorption units in each of the three regions to analyze the performance of the target value and the predicted value. High accuracy of matching target and predicted values in non-uniform and uniform learning areas. There is a large deviation between the target value and the predicted value in the transfer learning area. Finally, one unit in each of the three studied frequency band regions is selected for experimental testing to calculate the acoustic performance of the units and to further validate the effectiveness of the TSDNN model. Machine learning improves on-demand design efficiency and accuracy for acoustic metamaterials and has great potential for application in noise reduction.