Convolutional neural networks and fully connected neural networks for acoustic super-focusing

In this work, we design neural network architectures, including fully connected neural networks (FC) and convolutional neural networks (CNN), to solve forward design problems. We show that FCs and CNNs can predict the acoustic response based on the positions of cylindrical scatterers. These networks develop regressor models for approximating the absolute acoustic pressure at a focal point. A comparative study between the two networks is conducted against exact results from multiple scattering theory. We implement the FC and CNN models to simulate acoustic multiple scattering from cylindrical structures. While CNNs are computationally intensive, they perform better by leveraging spatial correlations between scatterers. This study tests the model's applicability to our problem. Numerical examples of pressure amplitude prediction at the focal point using random planar configurations of cylindrical rigid scatterers in water are presented. Results from deep learning models are compared to semi-analytical solutions from multiple scattering theory, with computations performed on MATLAB.