Deep-learning-enabled inverse design of large-scale metasurfaces with full-wave accuracy

Recent advances in meta-optics have enabled diverse functionalities in compact optical devices; however, conventional forward design approaches become inadequate as device complexity and scale grow. Inverse design offers a powerful alternative but often requires massive computational resources and neglects mutual coupling effects. Here, we propose and experimentally validate a deep-learning-enabled framework for rapid inverse design of large-scale, aperiodic metasurfaces with full-wave accuracy.The framework integrates an inverse design network responsible that maps target near-field responses to metasurface geometries in a non-iterative and scalable manner. A lightweight forward prediction network, integrated as a full-wave solver surrogate within the framework, enables efficient end-to-end training of the inverse design network while capturing mutual coupling effects by considering both local and neighboring geometries.The framework’s effectiveness is experimentally verified through a multi-foci metalens and a holographic metasurface. This framework enables the inverse design from micrometer to centimeter scales (> 20kλ), with near-field responses discrepancies less than 3% compared to full-wave solvers at subwavelength (< λ/10) resolution.Moreover, it is generalizable to metasurfaces of arbitrary size and operates efficiently without high-performance resources, overcoming the computational bottlenecks of previous inverse design methods.

💡 Research Summary

This paper presents a groundbreaking deep-learning framework for the rapid and accurate inverse design of large-scale, aperiodic metasurfaces, effectively overcoming the critical limitations of computational cost and neglected mutual coupling effects in conventional methods.

The core innovation lies in a dual-network architecture. First, a lightweight forward prediction network, based on a Multilayer Perceptron (MLP), is developed. It takes the geometric parameters of a target meta-atom and its five surrounding layers of neighbors as input and predicts the target’s local near-field electromagnetic response. Crucially, by explicitly including neighbor information, this network accurately captures mutual coupling effects—a factor often overlooked in simplified periodic approximations. This network is trained on a dataset of over 10,000 locally coupled meta-atom configurations, generated via rigorous Finite-Difference Time-Domain (FDTD) simulations.

Second, an inverse design network is trained to map a desired near-field distribution for an entire metasurface directly to the corresponding geometric parameters of all constituent meta-atoms. The key to its efficient training is the integration of the pre-trained forward prediction network as a fixed, surrogate full-wave solver within the framework. This allows for end-to-end training of the inverse network without the prohibitive cost of iterative FDTD simulations.

The framework demonstrates exceptional performance across multiple dimensions:

• Unprecedented Scale and Generalizability: It enables inverse design from the micrometer to centimeter scale (>20,000 wavelengths in diameter), representing an area expansion of five orders of magnitude compared to prior coupling-aware studies. A single trained model can handle metasurfaces of arbitrary size.
• Full-Wave Accuracy: The predicted near-field responses of designed metasurfaces show discrepancies of less than 3% compared to high-fidelity FDTD simulations, even at subwavelength (< λ/10) resolution.
• Computational Efficiency: The design process is non-iterative. Once trained, the inverse network generates designs for large-scale metasurfaces in tens of seconds on a standard four-core CPU, eliminating the need for high-performance computing resources.
• Experimental Validation: The practical efficacy of the framework is conclusively proven through the fabrication and testing of two complex devices: a multi-foci metalens and a holographic metasurface. The experimental results show excellent agreement with both the framework’s predictions and full-wave simulations.

In summary, this work successfully breaks the longstanding trade-off between simulation accuracy, design scale, and computational tractability in metasurface inverse design. By leveraging a physics-aware, locally-coupled forward model and a direct inverse mapping strategy, it establishes a scalable and efficient pipeline for creating large-area, high-performance meta-optical devices, paving the way for their adoption in demanding real-world applications such as advanced imaging, augmented reality, and laser systems.