A deep learning-based approach for the inverse design of the Helmholtz resonators

This article discusses the use of artificial neural networks in the design of Helmholtz resonators. A large database is constructed analytically by using the classical approach for computing the transmission loss and resonance frequencies of the Helmholtz resonators. In DNN, four geometric parameters, neck radius (rn), corrected neck length (ln) (derived from neck length (hn)), cavity radius (rc), and cavity height (hn) of the Helmholtz resonator, are the final output of the designed model which are mapped with the transmission loss and resonance frequency of the Helmholtz Resonators through our proposed neural network. A Feed-forward deep neural network (DNN) based on pre-transfer learning approach is used to map feature variables to target variables. The training follows the three major steps i.e., (a) Generalised pre-training in unsupervised manner, (b) Decoder pruning and regressor head training and (c) End to End regressor training using full backpropagation. This modularized approach removes the chances of overfitting, by effectively tuning the weights at each layer of the network. Each step focuses on creating a more structured and organized model. The best combination of the weight and biases is used for the prediction of the geometric parameters. Also, the finite element study of the transmission loss and resonance frequency supports the predicted geometric parameter of the randomly chosen sample for testing with the true value of the sample. It has been found that the accuracy of the model can be improved by training in a modular way. The approach discussed in this article can be useful to bypass the complex wave analysis approach for designing the Helmholtz resonators.