GenPnCFormer: Latent diffusion transformer for inverse design of variable 1D phononic crystals

초록

This study proposes GenPnCFormer, a generative inverse design framework for one-dimensional phononic crystals (1D PnCs) that integrates a latent diffusion transformer (LDT) and a conditional variational autoencoder (CVAE) to match target band gaps and/or defect bands. Conventional optimization-based design approaches are susceptible to local optima and require iterative numerical evaluations, resulting in high computational costs. Existing generative models, on the other hand, are limited by fixed structural representations, which restricts their applicability to variable unit-cell configurations and defect patterns. GenPnCFormer overcomes these limitations with two key components: a transformer-based CVAE that employs adaptive padding and self-attention to encode variable-length structural sequences into a unified latent space, and an LDT that, conditioned on a structured band-mask representation, leverages AdaLN-Zero and classifier-free guidance for controlled and diverse generation, effectively resolving the one-to-many mapping inherent in inverse design. Extensive evaluations demonstrate that the proposed framework consistently achieves high-fidelity matching in both periodic (band gaps) and aperiodic (defect bands) scenarios while maintaining robust performance across diverse and previously unseen design conditions. By eliminating iterative optimization, GenPnCFormer enables rapid and scalable wave engineering, providing a versatile solution for potential applications such as vibration isolation, waveguiding, and structural health monitoring. © 2026 Elsevier Ltd

키워드