A machine-learning virtual source model for nanoscale transistors

Neural networks (NNs) are powerful tools for modeling transistor characteristics from data, yet purely data-driven models are data-hungry, can yield unphysical results, and often fail to generalize. We present an integrated NN–virtual source (NN–VS) model that combines the physical rigor of the virtual source framework with the adaptive learning of neural networks. This end-to-end trainable hybrid model achieves higher accuracy, stronger data efficiency, and superior extrapolation power compared to pure NN models, while eliminating the manual parameter extraction required in the conventional VS models. Importantly, the NN–VS model demonstrates robust performance in data-limited scenarios. The model accurately captures experimental nanoscale transistor characteristics, including two-dimensional (2D) semiconductor field-effect transistors (FETs) and silicon FinFETs. The application of the model is illustrated in compute-in-memory circuit simulations with ferroelectric FETs.