Quantitative evaluation of process-induced line-edge roughness in FinFET: Bayesian regression model

Abstract

With the aggressive scaling down of the minimum feature size of advanced metal–oxide–semiconductor devices, it has become imperative to design and fabricate process-variation-immune devices. Technology computer-aided design simulations are typically used to test thousands of devices for process-variation immunity, but the process is computationally expensive. In this work, we propose a novel approach to simulate and predict the current–voltage characteristics of fin field-effect transistor devices with process-induced line-edge roughness (LER), within a few seconds. We exploit the Bayesian linear regression model to estimate the mean and standard deviation of the drain-to-source current (I DS) for an arbitrary gate voltage (V GS) and LER profile. We evaluate the prediction accuracy in terms of the mean absolute percentage error (MAPE) and root mean square error (RMSE). The MAPEs for the mean and standard deviation of I DS are <1% and <20%, respectively, and the corresponding RMSEs are 0.0804 and 0.0263, respectively. Once the I DS–V GS distribution is estimated by means of this novel approach, the distributions of other device metrics such as the threshold voltage and off-state leakage current can be estimated.

Quantitative Evaluation of Line-Edge Roughness in Various FinFET Structures: Bayesian Neural Network With Automatic Model Selection

“…The overall prediction performance was improved, as compared against the previous work [1] which is the Bayesian linear regression (BLR) model (see Table II). Note that the HS-BNN model for this comparison is only for the FinFET with Lg = 14 nm, Wfin = 7 nm, Hfin = 30 nm because the BLR model made predictions only for the corresponding FinFET structure.…”

“…As mentioned in the previous study [1], the distribution of 𝑙𝑜𝑔(𝐼 𝐷𝑆 ) for a given 𝑉 𝐺𝑆 can be considered as a normal distribution in terms of small kurtosis and skewness values. From this perspective, the mean and standard deviation of 𝑙𝑜𝑔(𝐼 𝐷𝑆 ) (denoted as 𝜇𝐼 𝐷𝑆 and 𝜎𝐼 𝐷𝑆 , respectively) were selected as target variables.…”

“…For this reason, a novel approach for less time-consuming simulation is necessary. In this context, we proposed a linear regression model in [1]; however, this linear regression model was limited in dealing with nonlinearity. Conversely, artificial neural network (ANN) is the most popular machine-learning model because it shows a powerful ability to find patterns as well as to detect trends in complex real-world data.…”