TY - GEN
T1 - Dual-Loss Trained Compact Cascaded ANN for Custom-Range BSIM-CMG Parameter Extraction
AU - Vazarkar, Sidhant V.
AU - Cachhi, Abdullah M.H.
AU - Ghosh, Shreyan
AU - Pahwa, Girish
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - This paper presents a novel deep learning framework that combines dual-loss training with floating normalization for BSIM-CMG model parameter extraction. The framework addresses the fundamental challenge of non-unique parameter-to-characteristic mappings through a cascaded forward-inverse neural network architecture trained with both parameter supervision and curve reconstruction constraints. A flexible floating normalization scheme enables user-defined extraction ranges, allowing multiple valid parameter sets to be extracted from identical I-V characteristics. The dual-loss training mechanism provides superior convergence by simultaneously enforcing parameter accuracy and forward-inverse consistency. Validated with 200 K Monte Carlo samples across 3 nm, 7 nm, and 14 nm FinFET technology nodes, the framework achieves RMSE below 3.75% while offering unprecedented flexibility compared to conventional fixed-range extraction methods.
AB - This paper presents a novel deep learning framework that combines dual-loss training with floating normalization for BSIM-CMG model parameter extraction. The framework addresses the fundamental challenge of non-unique parameter-to-characteristic mappings through a cascaded forward-inverse neural network architecture trained with both parameter supervision and curve reconstruction constraints. A flexible floating normalization scheme enables user-defined extraction ranges, allowing multiple valid parameter sets to be extracted from identical I-V characteristics. The dual-loss training mechanism provides superior convergence by simultaneously enforcing parameter accuracy and forward-inverse consistency. Validated with 200 K Monte Carlo samples across 3 nm, 7 nm, and 14 nm FinFET technology nodes, the framework achieves RMSE below 3.75% while offering unprecedented flexibility compared to conventional fixed-range extraction methods.
KW - BSIM-CMG
KW - deep learning
KW - dual-loss training
KW - floating normalization
KW - inverse design
KW - parameter extraction
UR - https://www.scopus.com/pages/publications/105040793083
U2 - 10.1109/EDTM65772.2026.11497494
DO - 10.1109/EDTM65772.2026.11497494
M3 - Conference contribution
AN - SCOPUS:105040793083
T3 - 10th IEEE Electron Devices Technology and Manufacturing Conference: Emerging Semiconductor Devices and Manufacturing Technologies, EDTM 2026
BT - 10th IEEE Electron Devices Technology and Manufacturing Conference
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 10th IEEE Electron Devices Technology and Manufacturing Conference: Emerging Semiconductor Devices and Manufacturing Technologies, EDTM 2026
Y2 - 1 March 2026 through 4 March 2026
ER -