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

Abstract

To design a device that is robust to process-induced random variation, this study proposes a machine-learning-based predictive model that can simulate the electrical characteristics of FinFETs with process-induced line-edge roughness. This model, i.e., a Bayesian neural network (BNN) model with horseshoe priors (Horseshoe-BNN), can significantly reduce the simulation time (as compared to the conventional technology computer-aided design (TCAD) simulation method) in a sufficiently accurate manner. Moreover, this model can perform autonomous model selection over the most compact layer size, which is necessary when the amount of data must be limited. The mean absolute percentage error for the mean and standard deviation of the drain-to-source current (I DS ) were ~0.5% and ~6%, respectively. By estimating the distribution of the current-voltage characteristics, the distributions of the other device metrics, such as off-state leakage current and threshold voltage, can be estimated as well.INDEX TERMS line edge roughness (LER), process-induced random variation, Bayesian neural network, automatic model selection.

Pelgrom-Based Predictive Model to Estimate Metal Grain Granularity and Line Edge Roughness in Advanced Multigate MOSFETs

“…The slope of the Pelgrom plots (A v ) increases linearly with the GS for the three architectures (see Fig. 4) validating (2). Note that, all devices with the same architecture and metal on the gate, even having different shapes (see Fig.…”

“…Next generations of advanced transistors require an improvement in the gate control, together with the reduction of variability effects [1]. Some of the most relevant sources of variability are the line edge roughness (LER) [2], [3], [4], produced by lithographic processes; the metal grain granularity (MGG) [4], [5], [6], caused by the metal gate deposition; or the random dopant fluctuations (RDF) [7], that are due to the presence of dopants in the channel.…”

“…These two factors notably increase the computational cost of these variability studies. Therefore, complementary techniques to predict variability were developed to lower the computational time, such as those based on machine learning (ML) [2], [8], [9], [10], [11], [12], the fluctuation sensitivity map (FSM) [13], the impedance field method [14], or the statistical model reported in [15].…”