Inverse design of terahertz metasurfaces for multi-channel holographic imaging based on modified gradient descent method

In recent years, research on terahertz metasurfaces has made significant progress and is evolving towards multifunctional integration. Traditional forward design methods have gradually exposed efficiency bottlenecks in multifunctional design due to their high computational complexity. The proposal of the inverse design paradigms has brought revolutionary breakthroughs in the design efficiency of multifunctional metasurfaces and greatly reduced the design threshold. As an efficient inverse design optimization algorithm, the gradient descent method can rapidly optimize a phase distribution driven by multi-objective tasks. However, the loss function in this traditional gradient descent (TGD) method only considers image loss, leading to large gradients between adjacent phases in the phase distribution, which in turn results in low efficiency of the metasurface. To address this issue, we propose a modified gradient descent (MGD) method. By introducing gradient loss, this method effectively enhances the smoothness of phase distribution. As simulation verification, we used the MGD method to design a geometric phase terahertz metasurface, successfully achieving nine-channel dispersive meta-holography, which demonstrates the effectiveness of the gradient loss term. We believe the proposed MGD method has broad application prospects in multifunctional metasurface design.