Two-decade strides in metasurfaces have enabled a multitude of theoretical breakthroughs and experimental discoveries that outride established comprehension. Deep learning has recently found favor for expediting metasurface design and unearthing complex light-matter interactions, in contrast to resource-intensive numerical simulations. However, most “black box” algorithms lack enough discernment on parsing internal physical connections. Here, we propose a physical adversary channel to be complementary to gradient descent channel by embedding the Kramers-Kronig (KK) relations into a neural network, quantifying the inherent spectral contradictions at the output side. We evaluate the superiority of the KK-driven neural network in forward prediction and inverse metasurface design by modifying loss function and shaping probability distribution in latent space, respectively. The exceptional outcome suggests that the similarity between output and given spectra reaches up to 99.8% and maintains an extremely high fidelity even in a mutant band. Our work provides a physically explicable perspective to explain “black box” models, possibly reviving intelligent metasurface applications.

Close

PrevNext

View figure in article

• Inverse problems
• Artificial neural networks
• Metamaterials
• Metasurfaces
• Deep learning
• Electromagnetic wave theory