[论文解读] Physics-Informed Deep Neural Network Design of Reactively Loaded Metasurfaces

Malik Almunif, Jade Le+1 more|arXiv (Cornell University)|Mar 16, 2026

Metamaterials and Metasurfaces Applications被引用 0

一句话总结

该论文提出一个 tandem 深度神经网络，联合学习用于超表面的反应性负载分布，使其来自目标远场图案，并利用一个可微分的微波网络前向求解器来快速预测辐射并指导训练。

ABSTRACT

A tandem deep neural network approach is presented for the inverse design of reactively loaded metasurfaces with prescribed far-field radiation characteristics. The proposed approach integrates a deep neural network (DNN) with a physics-based microwave network forward solver. The DNN maps target far-field patterns to distributions of reactive loads across the metasurface unit cells. The predicted distribution of reactive loads is evaluated by the forward solver to compute the resulting radiation pattern and guide the learning process through a cosine-similarity loss function. The forward solver enables a fast evaluation of the metasurface's electromagnetic response, significantly reducing the computational cost required for training. The proposed approach is applied to a metasurface with aperture-coupled unit cells loaded with reactances. Several design examples are presented to demonstrate the accurate synthesis of shaped and steered radiation patterns. Full-wave electromagnetic simulations are performed to validate the accuracy of the designed beamforming metasurfaces.

研究动机与目标

• Motivate inverse design of reactively loaded metasurfaces for shaped and steered beams.
• Develop a tandem architecture combining a CNN inverse network with a physics-based forward solver.
• Reduce training data and computational cost by using a microwave network solver for forward evaluation.
• Ensure predicted reactive loads yield realizable, capacitive loads within a practical range.

提出的方法

• Use a CNN with residual blocks to map a target far-field pattern to a 15x15 distribution of reactive loads.
• Employ a microwave network forward solver to compute the resulting far-field pattern from the predicted loads (Equations 1–4).
• Train end-to-end with a cosine-similarity loss between target and predicted far-field patterns plus a reflection constraint on S11.
• Train on randomly generated far-field patterns via array theory, with data augmentation to increase diversity.
• Implement differentiable forward propagation in TensorFlow to enable backpropagation through the network equations.

实验结果

研究问题

• RQ1 Can a tandem CNN–microwave-network framework accurately synthesize reactive-load distributions for target far-field patterns?
• RQ2 Does integrating a physics-based forward solver during training reduce the need for large paired datasets and accelerate convergence?
• RQ3 Are the designed metasurface patterns close to target patterns while maintaining acceptable input reflection (S11) levels?
• RQ4 What ranges of reactive loads are necessary to realize the predicted patterns within practical constraints?

主要发现

• The inverse network accurately synthesizes reactive-load distributions that reproduce target far-field patterns across multiple design examples.
• Forward evaluation via the microwave-network solver reduces computation time from ~12.2 hours (full-wave) to ~1.2 seconds per evaluation.
• The learned designs achieve close agreement with targets; table I shows competitive S11 values between predicted and simulated results.
• Cosine-similarity loss effectively aligns pattern shapes while the S11 constraint mitigates excessive reflections.
• Full-wave HFSS validation confirms the metasurface designs achieve the intended beam patterns.