PHIMO-NN: Compact Modeling by Fusing Device Physics and Neural Networks for One-Shot Parameterization

The complexity of compact device modeling has grown significantly in advanced technologies, including an increase in the number of model parameters, the time-consuming process of parameter extraction, and significant efforts required for model updating with device technology evolutions. To address these challenges, a new methodology by fusing device physics and tiny neural networks (PHIMO-NN) is developed. With the capability of neural networks in the universal approximation to assist compact modeling, e.g., replacing manual fitting functions and/or binning equations, the number of model parameters is reduced while a global coverage is achieved. The well-behaved derivatives of PHIMO-NN outputs to its parameters lead to a desired one-shot parameterization using a back-propagation (BP) training. For FinFET technology data of a 14/16 nm PDK model comprising 5 bins covering the entire gate lengths, PHIMO-NN parameterization takes 20 minutes, resulting in a significant speedup compared with traditional model extractions. Further, model transferring for technology updates is made easier with the continuous learning capability. For example, a PHIMO-NN trained for FinFETs with an error of 0.22% is easily extended for nanosheet GAAFETs with an error of 0.31%. PHIMO-NN combines the advantages of physics-based models (physics-awareness and simplicity) and AI models (parameterization efficiency and accuracy), promising an agile device modeling methodology for advanced/emerging technologies.