Data-driven approach for prediction of band diagram of locally resonant sonic crystal

Abstract

The present study focuses on the prediction of the band diagram of locally resonance sonic crystal (LRSC) using machine learning models from structural, lattice material parameters. The data set for the study is generated from the numerical simulations of LRSC for wide ranges of structural, lattice material parameters using COMSOL multiphysics solver. Three machine learning algorithms namely, support vector machine (SVM), artificial neural network (ANN), and random forest regression (RFR), are employed for the study. A total of 23040 samples are simulated and included in the training of the models. Hyperparameter tuning and data set size optimization are carried out to find the optimum machine learning model given the minimum required data set. The R2 score, root mean square error (RMSE), computational time, and inference time are used to evaluate the performance of machine learning algorithms. The feature importance scores and partial dependence plots (PDPs) are used to understand each input feature’s critical role in predicting the band diagram. Among all the machine learning models studied, RFR outperformed the ANN and SVM models. RFR with 50 decision trees performed well. However, RFR with 10 decision trees also performed well with slight higher RMSE offering lower inference time compared to RFR with 50 decision trees. Due to the importance of the first two band gaps in noise attenuation, the width and center frequency of the first two bandgaps are predicted through the band diagram. The feature importance scores and PDPs of RFR models show that the structural and lattice parameters play a crucial role in predicting the first and second bandgap, aligning with physical significance.