Deep learning approach-based prediction of transfer functions in structure-acoustic coupling models

This study proposes a novel deep learning framework including an autoencoder and a convolutional neural network to predict transfer functions in the structure-acoustic-coupled model. The primary objective is to reduce the computational expenses typically associated with finite element method (FEM) simulations. The FEM model utilized in this study comprises a structure-acoustic system with 400 points, spanning a frequency range of 5 Hertz (Hz) to 2000 Hz. This study aims to mitigate the significant computational cost imposed by conventional FEM when simulating multiple transfer functions. This study enables the rapid and precise prediction of transfer functions by applying deep learning techniques. The findings reveal a significant decrease in computational time, with the deep learning model requiring only 30 min to predict transfer functions at 40 specific points within the structure-acoustic model, a stark contrast to the 1453 min required using the FEM. Not only does the proposed method accelerate the process, but it also delivers enhanced accuracy compared with conventional interpolation methods. The predicted magnitude results achieved an error rate of 4%, outperforming the 12% error rate associated with the conventional interpolation approach. Similarly, for phase results, the proposed method achieved a 13% error rate, which was significantly lower than the 33% error rate from interpolation. These improvements make deep learning a promising tool for optimizing acoustic system design, with potential applications in noise control, structural acoustics, and industrial acoustic analysis. • The CAE-CNN model achieves a mean magnitude error of 4% and phase error of 13%, significantly outperforming conventional interpolation-based methods, especially at high frequencies. • A novel convolutional autoencoder transforms 1D transfer functions into a 2D latent space, enhancing the learning of spatial-spectral patterns in structure-borne acoustics. • The deep learning framework accurately reconstructs resonance peaks and sharp phase shifts across a broad frequency range (5 Hz to 2 kHz), demonstrating physical interpretability. • The proposed method generalizes well to unseen spatial positions and geometries, offering a scalable tool for noise prediction, diagnostics, and real-time acoustic applications.