Transformer-based inverse-design model for optimal multilayer microperforated panels

Multilayer microperforated panels (MPPs) are highly efficient noise absorbers with simple structures, making them an increasingly popular alternative to porous material-based solutions. However, as the number of MPP layers increases to achieve desirable and broader absorption bandwidths, the complexity of the design process grows significantly, making traditional optimization algorithms inefficient and extremely time-consuming for inversely designing long-sequence MPP structures. In addition, existing inverse-design methods have many drawbacks because they either fail to adapt to different design targets quickly or are difficult to suit different types of structures, e.g., designing for different topologies at each layer. In this work, we introduce AcoustoGPT (Acousto Generative Pretrained Transformer), a transformer-based model for the quick inverse design of long-sequence MPPs to realize a target absorption spectrum, and solve all these drawbacks simultaneously. Our approach employs a transformer encoder to capture a compact latent representation of the absorption spectrum, while a decoder generates the design parameters for each MPP layer. Finally, we integrate a bidirectional long short-term memory (LSTM) model with the transformer to enhance the design accuracy of the absorption spectrum. The trained model is versatile and can easily adapt to various design conditions, such as varying the number of layers and incident angles. This method has potential in designing other sequential or layered structures, including series Helmholtz resonators and higher-dimensional metastructures.