On-demand prediction of low-frequency average sound absorption coefficient of underwater coating using machine learning

This study proposes an underwater coating with sound absorption ability in the middle-to-low frequency range and establishes an acoustic theoretical model combining the equivalent medium theory and the transfer matrix method. The sound absorption coefficient, surface characteristic impedance, equivalent volume longitudinal wave modulus, and equivalent sound velocity are calculated and solved. Using the preset 20 sensitive parameters and the hypercube sampling method, this study establishes 100,000 random sound absorption coefficient curves in the frequency range of 1 Hz-1,000 Hz. Further, deep neural networks are employed to predict the average value of the sound absorption coefficient curve. The overall loss function is derived by combining the mean square error between the expected average sound absorption coefficient and its predicted value and the network- optimized loss function to ensure that the 20 sensitive parameters that meet the acoustic performance can be predicted. Finally, two randomly selected sound absorption curves are used for prediction tests. The verification results indicate that the error between the expected average absorption coefficient and the predicted average absorption coefficient corresponding to the 20 sensitive parameters is only 0.026 % and 0.33 %. The proposed method can be extended to predict the average absorption coefficient value for any acoustic structure, which could be beneficial for the performance development of acoustic functional devices.