Large language model-driven inverse design of acoustic valley-Hall insulators

Abstract Acoustic valley-Hall insulators (AVHIs) are an emerging class of acoustic topological materials that enable robust directional wave transport and show promise for applications in communication, sensing and wave manipulation. This study presents a large language model (LLM)-driven inverse design framework consisting of two collaborative agents. The Designer-LLM is fine-tuned on an AVHI database generated through reinforcement learning to perform bandgap-targeted inverse design. The Manager-LLM coordinates task planning, semantic parsing, simulation tool invocation and multimodal result generation. With this framework, multimodal design outputs, including structural parameters, band structures, eigenmode fields, and manufacturable model files, can be generated according to concise design requirements. Bandgap error analysis on test sets indicates that the fine-tuned Designer-LLM generates AVHI structures with reliable bandgap-matching accuracy. Overall, this study proposes a modular LLM-agent framework for the inverse design of AVHIs. The framework integrates semantic interaction, AVHI inverse design, numerical verification, and multimodal result generation into a unified workflow and can provide a useful reference for future extensions to other acoustic metamaterial design tasks.