Designing acoustic metalenses capable of customized beam control holds great potential for applications in focusing, imaging, and energy harvesting. However, conventional design methods often result in metalenses with limited functionality and poor adaptability due to the lack of structural adjustability freedom. Here, we propose a deep-learning-based design framework (DLDF) that enables rapid generation of high-degree-of-freedom lens structures with controllable sound emission patterns. The metalens is a pixelated structure in water with 50% of its pixels randomly occupied by cross-shaped scatterers. By integrating three neural network models—an autoencoder emission feature extraction, a variational autoencoder for structural representation encoding, and a deep neural network for latent-space mapping and optimization, the DLDF allows for a high-fidelity bidirectional mapping between the structural configuration and the resulting sound emission profile. Numerical simulations show that this approach can achieve metalenses with desired functionalities, including focusing, beam collimation, and emission angle control. Our work provides a scalable and generalizable strategy for the intelligent metalens design, stimulating related investigations and applications in other wave systems.
Wan et al. (2026) studied this question.