Rapid design method for metamaterial lens based on transfer learning and convex optimization

Abstract This paper proposes a fast design method for electromagnetic focusing lenses based on transfer learning and convex optimization. By constructing a three-layer symmetric structure model and conducting CST-MATLAB joint simulation, 40000 sets of data covering 360° phase are obtained. A ResNet34 network with transfer learning is used to classify the phase. The classification reaches an accuracy of 89.94%. The predicted phase and amplitude are used as input to a convex optimization model. This model calculates the target phase distribution for the entire metamaterial. Based on the calculated results, the method synthesizes the lens structure. Simulations and measurements show that the lens achieves effective far-field focusing at 2.6 GHz. The sidelobe level remains below −12 dB. Compared with traditional empirical design methods, this approach improves accuracy and reduces development time. It is suitable for applications such as electromagnetic imaging, radar detection, and wireless energy transmission.