Data-driven inverse design framework for broadband mode-selective elastic metasurfaces

Elastic waves in structural materials typically propagate as a combination of multiple coupled modes, each exhibiting complex frequency-dependent behaviour. This inherent multimodal nature poses a major challenge for precise wave manipulation. While as a subwavelength artificial interface, metasurfaces have shown strong potential in controlling elastic waves, most existing approaches are limited to single-mode or single-frequency operation, constraining their effectiveness in more realistic, multimodal scenarios. Moreover, conventional metasurface design methods based on sophisticated theories, strict phase constraints, along with complex topology optimization are computationally demanding and time-intensive. Here, we propose a data-driven inverse design framework for broadband mode-selective elastic metasurfaces, integrating the strengths of artificial intelligence. A Physics-Informed-Embedded Multi-Branch Convolutional Neural Network (PIMB-CNN) is developed to rapidly and accurately predict the transmission coefficients of both A0 and S0 modes across a broad frequency range. This surrogate framework is integrated with the Sparrow Search Algorithm (SSA) to enable an efficient inverse design workflow, capable of generating optimal metasurfaces to achieve customized elastic-wave mode in under a couple of minutes. Numerical simulations and experimental validations confirm that the designed metasurfaces can effectively suppress undesired modes within a broadband range while enabling the transmission of target modes with minimal loss. This work offers a rapid, artificial-intelligence-assisted (AI-assisted) design paradigm for elastic metasurfaces of customized mode-selection, highlighting a promising route toward intelligent wave control in complex multimodal physical scenarios.