Sensitivity Analysis of Process Parameters Based on Adjoint Partial Derivative Network

With the development of RF microwave system to modularization, miniaturization and high frequency, the RF performance of RF microwave integrated interconnection transmission structures, such as gold wire bonding wires, RF interconnection holes, transition transmission lines and other passive transmission structures, is significantly affected by various factors such as material property parameters, dielectric thickness, and process parameters. At the same time, because of the complex nonlinear coupling relationship between the physical parameters and RF performance, it is difficult and time-consuming to analyze the influence of each parameter. Based on the multi-input and multi-output multi-layer full-connected neural network, this paper proposes a method to analyze the influence of physical parameters by using the ad-joint partial derivative network (APDN) of the neural network. Firstly, the behavioral relationship between the physical parameters and the RF scatter parameters of the RF passive transmission structure is roughly modeled through the neural network. Then the partial derivatives of each physical parameter in the entire parameter space are calculated based on the APDN; Finally, the parameters that dominate the complex nonlinear coupling relationship are extracted through partial derivatives. The proposed method is helpful to quickly build the sensitivity analysis model of the RF passive transmission structure, and help to achieve the effective control of the sensitive parameters in the microwave integration process.