Machine Learning Surrogate Modeling for Accurate Prediction of GAA ‐ MOSFET Electrical Characteristics

Abstract

Machine learning regression models were employed to predict the electrical characteristics of GAA‐MOSFETs using a dataset of 459 simulation points. The models considered input features including gate voltage, drain voltage, channel length, silicon thickness, and metal work functions, with the logarithm of the drain current as the target. Various algorithms—including MLPRegressor, GradientBoostingRegressor, XGBRegressor, and ensemble tree‐based models—were trained with cross‐validation and evaluated on an independent test set. The MLPRegressor achieved the highest predictive performance ( R 2  = 0.9990, RMSE = 0.0946, MAE = 0.0612), closely reproducing simulation results for ON‐ and OFF‐state currents and threshold voltage. Feature importance analysis identified gate voltage and metal work function as the most influential parameters. The model accurately captured the effects of channel length and silicon thickness on device behavior, demonstrating its potential as a fast and reliable surrogate for computationally intensive numerical simulations, enabling rapid design and optimization of nanoscale transistors. In excess of DC behavior, the study is extended to small‐signal RF analysis and the linearity of the cylindrical GAA MOSFET: the transconductance ( g m ), output conductance ( g d ), transconductance generation factor (TGF), unity‐gain cutoff frequency ( f T ), and two‐port Y/S parameters. The simulated device exhibits a maximum transconductance ( g m ) of 107 μS, an intrinsic gain g m / g d of 54.5, and a cutoff frequency f T of 139 GHz. The same MLP model reproduces these RF performance metrics with a coefficient of determination R 2  > 0.97, enabling rapid design and optimization of nanoscale RF transistors.