A machine learning (ML) -based inverse hybrid design method (IHDM) is proposed to efficiently design multilayer frequency-selective surfaces (FSSs). The method inputs S11 and the number of layers for the FSS and outputs a topological shape matrix and a structural parameter matrix for the FSS design without iterative procedures. Two numerical examples validate the ML-based hybrid modeling method both within and outside the training frequency range. Compared to mainstream machine learning frameworks such as fully connected networks (FCNs), convolutional neural networks (CNNs) and attention networks, the proposed method is proven effective with higher design accuracy.