Sound absorption performance prediction of multi-dimensional Helmholtz resonators based on deep learning and hyperparameter optimization

The problem of low-frequency noise is becoming increasingly severe and measuring the sound absorption performance of acoustic metamaterials (AMs) using accurate absorption coefficients is of great interest in low-frequency noise control engineering. Conventional calculation methods such as Finite Element Method (FEM) simulations and the theoretical analysis methods (TAM) have specific limitations. Deep learning (DL) models provide new perspectives for studying AMs acoustic performance. However, the prediction performance of DL models is highly dependent on the proper tuning of hyperparameters. As far as is known, existing literature has not systematically explored the impact of hyperparameter tuning on DL models in the context of acoustic performance studies. The present paper designed a multi-dimensional Helmholtz resonator (MDHR) consisting of a 4 × 4-type continuous parallel arrangement, while a dataset was established via FEM. Furthermore, a deep neural network (HPO-DNNs) model based on hyperparameter optimization (HPO) was proposed to predict the acoustic performance of the MDHR. Random search (RS), Bayesian optimization (BO), Simulated annealing (SA), and genetic algorithm (GA) were introduced to optimize the hyperparameters (learning rate, weight decay, optimizer, and batch size) of the DNNs. The mean square error (MSE), coefficient of determination (R2) of the testing dataset and the optimization time were used as the evaluation metrics, GA was selected for further study based on the comparison results (MSE = 0.00177, R2 = 0.98151) of the optimization efficiency and predictive precision of DNNs from the four HPO algorithms. Finally, the prediction performance of the GA-DNNs model was evaluated in single-, multi-, and broadband conditions in practical applications, demonstrating high precision and stability and providing a new approach for acoustics performance studies.