Tiny Neural Network Representing Mosfet Physical Effect Sub-Model

Tiny neural network (NN) modeling physical effects inside MOSFET is demonstrated in this work. BSIM-IMG and HSPICE are used to generation datasets, which samples above 20000. Physical effects, Threshold Voltage Roll-off (TVR) and DIBL, are set as NN modeling goal and achieved >99% and >92% fitting accuracy respectively (@ 5-100-100-42 FCN, 15 training epochs). The conception of cooperation between multi NN-based sub-model and BSIM core-model is also verified (accuracy >95%).