This paper proposes an advanced inverse neural network modeling technique for the high-accuracy and rapid electromagnetic optimization of passive devices. This method integrates the transfer function with a convolutional neural network to address the challenge of high-dimensional input data in wideband device modeling. The transfer function effectively reduces input dimensionality, while the convolutional neural network, with enhanced feature extraction capabilities, captures the complex nonlinear relationship between the transfer function coefficients and the geometric variables. Once trained, the model allows for direct prediction of geometric values without the need for iterative optimization algorithms. This inverse modeling approach is successfully applied to the multi-objective design of a wideband filter, enhancing the overall design efficiency of terahertz frequency multipliers.