The theory and methods of inverse problems, via comprehending experimental measurements and numerical simulation, serve as a new pathway for high accuracy analysis and design of complex mechanical structures. Because of the ill-condition of the inverse problem, it cannot be solved as easily as the forward problem. To tackle the intricacies due to insufficient samples, unclear modeling mechanism, high-dimensional multi-source uncertainties with strong correlation, high computation complexity due to multilayer-nested inversion, this paper systematically explores the two-way neural network inverse method. The two-way neural network consists of a forward network and an inverse network. The forward network constructs the functional model between the parameters to be solved and the response parameters. By analyzing the sensitivity of the distribution of parameters, such as the activation function and the condition number of the weights, on the solution when the forward network trains the forward problem, parameter dimension is reduced, the activation function is revised and regularization is supplemented. The solution error to the inverse network is thus substantially reduced. Compared with other methods, expensive inverse problem solving is avoided. Experiments show that the two-way neural network can improve the solution accuracy and efficiency of the inverse problem through exemplification with the inverse problem of solving the composite material parameters and the stiffness of the robot joints.