The inverse problem of designing materials to achieve desired acoustic functionality while strictly adhering to acoustic principles remains an unresolved scientific challenge. This work introduces a generative adversarial network based on the wave equation (Wave-GAN), in which the governing equation is directly embedded into the training process to iteratively optimize the material density distribution. The physics-guided framework enables the model to learn acoustic patterns directly from sound and generate material distributions with specified functionalities. As a verification example, a speaker recognition task was conducted. The results demonstrate that Wave-GAN produces physically consistent materials, achieving a recognition accuracy of 95.6%. This opens a promising direction for fully physics-driven material design at the interface of acoustics and machine learning.
Wu et al. (Fri,) studied this question.