A Deep Time Delay Neural Network with Generalized Activation Function for Microwave Transistor Modeling

This letter proposes a deep–narrow time delay neural network (TDNN) architecture for efficient microwave transistor modeling. The proposed approach incorporates a generalized activation function with four learnable parameters, enabling flexible adaption of the activation function into various forms, including standard sigmoid, adaptive sigmoid, swish, adaptive swish, and other variants tailored to meet specific modeling requirements. Such flexibility can further improve the modeling accuracy of the deep TDNN model. In addition, to efficiently train the deep TDNN model utilizing typical transistor measurement data, analytical formulations for direct current, small-signal S-parameter, and large-signal harmonic outputs of the proposed TDNN are derived. A three-stage training algorithm is also proposed to effectively train the deep TDNN model. The effectiveness of the proposed method is validated through modeling examples of microwave transistors and the application in power amplifier analysis.