标题： 基于扩散模型的衍射超表面的逆向设计 显示英文标题

标题： Inverse Design of Diffractive Metasurfaces Using Diffusion Models

摘要： 超表面是由工程化的亚波长结构组成的超薄光学元件，能够精确控制光。 其逆向设计——确定产生所需光学响应的几何形状——由于结构与光学特性之间的复杂非线性关系而具有挑战性。 这通常需要专家调整，容易陷入局部最小值，并且涉及大量的计算开销。 在本工作中，我们通过将扩散模型的生成能力整合到计算设计工作流中来解决这些挑战。 使用RCWA模拟器，我们生成由超表面几何结构及其对应的远场散射图案组成的训练数据。 然后，我们训练一个条件扩散模型，从指定波长下连续支持带中采样的目标空间功率分布预测元原子的几何形状和高度。 一旦训练完成，该模型可以生成误差较低的超表面，可以直接使用RCWA引导的后验采样，或者作为传统优化方法的初始值。 我们在空间均匀强度分束器和偏振分束器的设计中展示了我们的方法，两者均在30分钟内以低误差生成。 为了支持数据驱动的超表面设计的进一步研究，我们公开发布了我们的代码和数据集。 显示英文摘要

摘要： Metasurfaces are ultra-thin optical elements composed of engineered sub-wavelength structures that enable precise control of light. Their inverse design - determining a geometry that yields a desired optical response - is challenging due to the complex, nonlinear relationship between structure and optical properties. This often requires expert tuning, is prone to local minima, and involves significant computational overhead. In this work, we address these challenges by integrating the generative capabilities of diffusion models into computational design workflows. Using an RCWA simulator, we generate training data consisting of metasurface geometries and their corresponding far-field scattering patterns. We then train a conditional diffusion model to predict meta-atom geometry and height from a target spatial power distribution at a specified wavelength, sampled from a continuous supported band. Once trained, the model can generate metasurfaces with low error, either directly using RCWA-guided posterior sampling or by serving as an initializer for traditional optimization methods. We demonstrate our approach on the design of a spatially uniform intensity splitter and a polarization beam splitter, both produced with low error in under 30 minutes. To support further research in data-driven metasurface design, we publicly release our code and datasets.