Non-Quasi-Static Modeling of Neural Network-Based Transistor Compact Model for Fast Transient, AC, and RF Simulations

We develop a charge deficit-based non-quasi-static (NQS) model that is compatible with neural network-based transistor compact models for transient, AC, and RF simulations. The charge deficit model calculates the deficient or surplus charge in the channel to model the NQS effect. We introduce physics-based parasitic charges to extract intrinsic channel charges from trained quasi-static (QS) NN models. An improved loss function is also proposed to obtain physical charge values from capacitance-only training data. A charge deficit subcircuit is applied to calculate the NQS currents. We demonstrate the model’s accuracy in transient, AC, s-parameter, and RF mix-signal simulations. The proposed model can be easily integrated with QS NN-based compact models without the loss of model efficiency. The speed benefit of the proposed model is more than 14X faster than the standard compact model.