[paper] DL Physics-Driven MOSFET Modeling

Ming-Yen Kao, H. Kam, and Chenming Hu, Life Fellow, IEEE

Deep-Learning-Assisted Physics-Driven MOSFET Current-Voltage Modeling

in IEEE Electron Device Letters

DOI: 10.1109/LED.2022.3168243

Abstract: In this work, we propose using deep learning to improve the accuracy of the partially-physics-based conventional MOSFET current-voltage model. The benefits of having some physics-driven features in the model are discussed. Using a portion of the Berkeley Short-channel IGFET Common-Multi-Gate (BSIM-CMG), the industry-standard FinFET and GAAFET compact model, as the physics model and a 3-layer neural network with 6 neurons per layer, the resultant model can well predict IV, output conductance, and transconductance of a TCAD-simulated gate-all-around transistor (GAAFET) with outstanding 3-sigma errors of 1.3%, 4.1%, and 2.9%, respectively. Implications for circuit simulation are also discussed.

Fig: (a) Model implementation for circuit simulations, without the relative gm and gds errors terms in the cost function, Model shows larger prediction error in (b) gm and (c) gds.

Acknowledgements: This work was supported by the Berkeley Device Modeling Center, 

UCB, CA (USA)