Achieving precise prediction of sound absorption performance for composite acoustic metamaterials utilizing machine learning

Abstract

To facilitate rapid and precise estimation of the acoustic performance of composite structures, this paper employs Deep Neural Networks (DNNs) within the realm of machine learning to tailor the design based on three key characteristics of the sound absorption performance of composite acoustic metamaterials: the frequency and magnitude of the maximum absorption peak, and the average absorption coefficient. Initially, a database comprising 100,000 randomly generated absorption curves was established, with 90 % of the data allocated for training and the remaining 10 % for test named data subset A. Subsequently, the database subjected to five-fold cross validation demonstrated a considerable level of prediction accuracy on data subset A and various ranges of data subset B. Finally, this paper randomly selected 10 sets of three sound-absorption characteristic parameters and conducted inverse prediction of the 28 geometric parameters for the corresponding composite acoustic metamaterials, using a fixed data subset A for each parameter set. These predicted geometric parameters were then used to derive the predicted sound absorption characteristics for the ten sets. When compared to the given values, the results exhibited a maximum relative error of 4.110 %, a minimum of 0.000 %, with the majority of errors falling within 0.100 %. This demonstrates that the DNN model presented in this paper can achieve accurate and swift predictions of the primary acoustic characteristics of acoustic composite structures, offering direct benefits in reducing the development cycle and saving labor and time costs.