Parallel and Scientific Computing Laboratory, Electrical Engineering and Computer Science International Graduate Program, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
Auto-generated profile, based on publicly available information
Rajat Butola
Parallel and Scientific Computing Laboratory, Electrical Engineering and Computer Science International Graduate Program, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
Machine learning (ML) is poised to play an important part in advancing the predicting capability in semiconductor device compact modeling domain. One major advantage of ML-based compact modeling is its ability to capture complex relationships and patterns in large datasets. Therefore, in this paper a novel design scheme based on dynamically adaptive neural network (DANN) is proposed to develop fast and accurate compact model (CM). This framework constitutes a powerful yet computationally efficient methodology and exhibits emergent dynamic behaviors. This paper demonstrates that the compact model based on ML can be designed to replicate the performance of conventional compact model for nanodevices. For this work, gate-all-around (GAA) nanosheet (NS) device characteristics are comprehensively analyzed for process variability sources using the proposed model. The device geometry parameters such as channel length, nanosheet width and nanosheet thickness are fed as input features to the DANN model. The adaptive neural network learns dynamically by updating weights of the model in accordance with the input features and achieves accurate neural weight convergence. The proposed model predicted the electrical characteristics of NS devices with less than 1% error rate. The model is also implemented and validated for the simulations of digital circuit designs such as inverter, and logic gates.
26 August 2022 in IEEE Transactions on Microwave Theory and Techniques
Advanced field-effect transistors (FETs), such as gate-all-around (GAA) nanowire (NW) and nanosheet (NS) devices, have been highly scaled; therefore, they are critically affected even by a microscopic fluctuation. As the GAA NS device is considered a promising candidate beyond 5-nm technology, it is essential to analyze the effects of these fluctuations on dc and analog/radio frequency (RF) characteristics for future applications. In this article, we for the first time demonstrate that the machine learning (ML)-aided numerical device simulation approach can be used to model the effects of various fluctuations on the characteristics of GAA NS FETs (NSFETs). Among various fluctuations, we mainly focus on work function fluctuation (WKF), random dopant fluctuation (RDF), and interface trap fluctuation (ITF). The independent and combined effects of these fluctuations on the characteristics of NSFETs are studied. Except for transfer and output characteristics, analog and RF parameters, such as gate capacitance, transconductance, cutoff frequency, 3-dB frequency, and transconductance efficiency, are analyzed in detail. The main aim of this work is to show the capability and generality of ML in modeling various electrical characteristics of the explored NSFETs. The results show that the ML-based technique is fast and efficient, which accelerates the overall process and gives engineering acceptable accurate results.
18 April 2024 in IEEE Transactions on Nanotechnology
In this study, we report the process variation effect (PVE) including the work function fluctuation (WKF) on the DC/AC characteristic fluctuation of stacked gate-all-around silicon complementary field-effect transistors (CFETs). The PVE affects characteristic fluctuation significantly; in particular, for the variability of off-state current. Owing to the bottom channel of a fin-type, the P-FET suffers from the worst off-state current fluctuation (more than 200% variation) compared to the N-FET. The device variability induced by the WKF is marginal because of amorphous-type metal grains. As input features to an artificial neural network (ANN) model, low and high work function values, as well as parameters of PVE that have prevalent effects on CEFT transfer characteristics are further considered and modeled. The estimated values of R 2 -score prove that the ANN model properly grasps information from the dataset successfully; thus, it can be used to model emerging CFETs for circuit simulation.
24 December 2022 in Computers and Electrical Engineering
In this work, a dynamic weighting-artificial neural network (DW-ANN) methodology is presented for quick and automated compact model (CM) generation. It takes advantage of both TCAD simulations for high accuracy and SPICE simulations for cost-effectiveness. This methodology is developed for gate-all-around (GAA) silicon (Si) nanosheet (NS) complementary field effect transistor (CFET), a potential candidate for future CMOS technology due to its innate properties. The critical process variation (PV) sources that severely degrade the CFET performance are estimated using DW-ANN. It reduces the computation cost and predicts the effects of PV sources with less than 1% error. Furthermore, a compact DW-ANN-based Verilog-A model has been developed that captures the dc as well as transient behavior accurately for circuit-level analysis. CFET-based circuits such as inverter, 6T-static random access memory (SRAM), and ring oscillator (RO) are characterized and implemented using DW-ANN-based CM. The overall average error of the model is reported as less than 2%. Therefore, the proposed device and circuit modeling approach provides a feasible solution for the rapid compact modeling of new emerging devices with good convergence and high accuracy.