PC-CNN: A Physics-Augmented Framework for Reliable Electromagnetic Prediction in Metasurface Engineering

Metasurfaces, as a new type of electromagnetic control device, have shown great potential in wavefront manipulation and polarization control, etc. However, the traditional design methods based on numerical simulation suffer from computational efficiency bottlenecks and face the challenge of high-dimensional parameter space computational costs. Existing data-driven methods such as convolutional neural networks often lead to prediction results violating basic physical laws due to the lack of physical constraints. To address this key issue, this paper proposes a Physics-Constrained CNN (PC-CNN) model, innovatively embedding the discretization conditions of Maxwell’s equations, the law of energy conservation, and the continuity requirements of field components through an adaptive regularization mechanism into the network training process. Experiments show that compared with traditional data-driven models, PC-CNN achieves an average error reduction of 4.3% in electromagnetic response prediction for typical metasurface unit prediction tasks. This model maintains the advantage of millisecond-level response speed while significantly enhancing the physical rationality of sub-wavelength structure electromagnetic characteristic prediction, providing a new paradigm for intelligent electromagnetic computing.