Machine Learning Based Device Simulation Using Multi-variable Non-linear Regression to Assess the Impact of Device Parameter Variability on Threshold Voltage of Double Gate-All-Around (DGAA) MOSFET

For the first time, the machine learning approach is proposed for the analysis of device parameter variability impact on the threshold voltage of silicon-nanotube-based double gate-all-around (DGAA) MOSFET using multi-variable non-linear regression with five input variables. Interior-point algorithm is implemented in MATLAB and used for training the hypothesis. Algorithm is supplied with 2000 random initial guesses to ensure global minima in optimization for healthier accuracy at comfortable simulation time. The worst-case accuracy of 96.33 percent is achieved in prediction with an average percentage prediction error of 0.83 percent. Study revealed that, with proper training and optimization of the hypothesis (fitness) function it is possible to predict the threshold voltage of the device by keeping good trade-off between computation time and accuracy.