Answers are grounded in this paper and cite their sources.
To address the high structural dimensionality, complex electromagnetic response mapping, and difficult inverse search associated with high-degree-of-freedom pixelated metasurfaces, this work develops a pixelated metasurface inverse-design method based on a conditional variational autoencoder (CVAE) and a convolutional neural network (CNN) surrogate model. The framework uses the target electromagnetic response and the topology label as conditional inputs. The CVAE generates candidate pixelated unit cells, while the CNN surrogate model rapidly predicts their complex reflection responses for candidate screening. A reflection-type pixelated metasurface dataset is constructed from unit-cell structures, topology labels, and simulated x/y-polarized complex reflection responses. It contains 9000 samples covering cross-shaped, H-shaped, and split-ring unit cells. Each structure is encoded as a 32×32 binary matrix, and the corresponding complex reflection response is organized as a 200-dimensional feature vector. The response-matching performance is further evaluated separately for the two polarization channels, and the generated structures achieve mean squared errors of approximately 0.009 and 0.037 for the x- and y-polarized responses, respectively. The inversely designed unit cells are further used to construct a beam-deflecting metasurface, a focused vortex metasurface with l =+2, and a dual-polarization holographic metasurface, achieving beam deflection, second-order focused vortex generation, and polarization-dependent holographic imaging at 13 GHz. These results demonstrate the effectiveness of the proposed framework for pixelated unit-cell generation and multifunctional metasurface design.
No takes yet. Share an insight, caveat, or question.
li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: