An accurate neural network model to study threshold voltage variability due to metal grain granularity in Nanosheet FETs

Nanosheet FETs are currently considered one of the preferred architectures for the next technology nodes. Due to the expensive manufacture of new devices, other solutions, such as technology-aided computer design (TCAD), are needed to evaluate the impact of variability on future transistors. However, the realistic simulation of these devices is computationally demanding. Therefore, exploring new techniques such as the Pelgrom-based predictive model or the application of machine learning techniquesis essential. We present a multi-layer perceptron (MLP) neural network (NN) to estimate the impact of metal grain granularity (MGG), one of the most harmful sources of variability, on the threshold voltage (Vth) of a 12 nm gate length nanosheet (NS) FET. We demonstrated that this NN could obtain an accuracy of R2 = 0.937/0.972 by using only the 20/60% of the training dataset, reducing the computational time 5.0×/1.6× with respect to the total train dataset. Also, once the NN is trained, it can accurately predict the impact of realistic MGG variability on VT h with no further simulations.

keywords: Machine Learning, TCAD, Nanosheet FET, metal grain granularity (MGG), Variability