For the problems of large-scale non-linear finite element simulations, which often involve long computation times and difficulty in convergence, this paper presents an optimized method with good computational accuracy and applicability. First, a sufficient amount of data is obtained through finite element analysis or experiment, which is called a snapshot, and the principal components of the snapshot are extracted by using the proper orthogonal decomposition (POD) method. A low-order model of the physical problem is then established, and a low-order surrogate model of the complex problem is obtained by combining the radial basis function (RBF) method. This method can quickly predict the response with input data, which effectively shortens the computation time with good efficiency. The proposed method provides a new method in the computation of soft actuators.