Deep learning-based prediction of line edge roughness-induced variability in dual-gate FinFETs

Abstract

The continuous scaling of semiconductor devices into the sub-5-nm regime has intensified the influence of Line Edge Roughness (LER) on the electrical characteristics and reliability of Dual-Gate FinFETs, particularly when integrated with high-k gate dielectric materials. LER presents random variations in the fin geometry and gate structure which cause major variations in the key measures of performance like Static Noise Margin (SNM), Threshold Voltage (VTC), and current drivability. Although it is correct, conventional Technology Computer-Aided Design (TCAD) simulations are computationally expensive and cannot be used to model statistical variation on a large scale. To deal with this challenge, the Shallow Quaternion Reverse Generative Artificial Protozoa Adversarial Attention Networks (SQRA-P2AN) model is suggested, which is a combination of Quaternion Generative Adversarial Networks (QGAN) and a Shallow and Reverse Attention Network (SRAN). The Artificial Protozoa Optimizer (APO) is used to optimize the network parameters to enhance the rate of convergence and reduce the prediction error. The hybrid deep-learning framework is an efficient way of representing nonlinear associations between LER parameters and variation of FinFET performance, and it gives better predictive power than traditional models. The prediction results and accuracy assessment of the hybrid SQRA-P2AN model and TCAD simulation indicate that the model is highly consistent with the WMAPE value of 0.38–0.86, RSNM value of 149.05–150.12 mV, and WSNM value of 293.65–296.43 mV, as the proposed model is measured using the same parameters of LER (0.2–0.7 nm, 20–90 nm, 2) The SQRA-P2AN model decreases the average error in prediction by about 32 percent compared to the results of TCAD, and it also increases the stability of the voltage prediction by 28 percent and decreases the overall time of computation by almost 35 percent.