Automated Design and Inverse Modelling of Metasurfaces using Modified V-shaped Binary Particle Swarm Optimization and Convolutional Neural Networks

This paper presents a fully automated framework for the design and inverse modeling of polarization-insensitive metasurfaces. A theoretical formulation of V-shaped transfer functions in Binary Particle Swarm Optimization (BPSO) is first established to analyze their stability and dynamics, from which a Modified V-shaped transfer function is analytically derived. This modification enhances the balance between exploration and exploitation with a ratio of 1.03, nearest to unity among existing methods, achieving an average mean-squared error of approximately 0.6% in binary electromagnetic (EM) optimization. The MVBPSO is integrated with a Python-CST automation pipeline that efficiently generates simulation data for training a lightweight Convolutional Neural Network (CNN). The trained CNN predicts metasurface geometries from target responses with sub-1% error using 50-60% fewer samples than conventional approaches. The combined MVBPSO-CNN framework enables one-shot optimization and inverse design. Experimental realization of polarization-insensitive bandpass and bandstop metasurfaces at 28 GHz and 79 GHz are performed. The results demonstrate a reliable, data-efficient, and general approach for EM optimization and inverse design.