Inverse design of acoustic topological insulators based on physics-informed neural network

Research on acoustic topological insulators has advanced, with their topologically protected edge states enabling dissipationless sound transmission and showing potential in anti-interference acoustic devices and communication. However, precise and efficient design remains a challenge due to the complexities of geometric configurations and the sporadic nature of phase transitions. This study proposes a Topological-evaluation Network, a physics-informed neural network framework that integrates physical band structure information into the network as an embedding structure. Through convolutional operations for feature separation, it enables the inverse design and mechanistic analysis of acoustic topological insulators. A symmetry-related Core and Perturbation Design Strategy was discovered through this framework. Compared to traditional trial-and-error methods based on numerical simulation, this strategy simplifies the acquisition of acoustic structures with topological insulation properties. It effectively optimizes the design efficiency for systems with complex physical mechanisms. By integrating physical band information, the framework achieves a 12-fold reduction in training data volume. The proposed model achieves a 95 % prediction accuracy in identifying topological phases, significantly accelerating the design of complex acoustic insulators. This approach offers extensive applications for future analysis and design of complex acoustic topological systems.