Intelligent BSIM4 Model Parameter Extraction for Sub-100 nm MOSFET Era

Abstract

We present, in this paper, an intelligent extraction technique for obtaining a set of optimal model parameters of the Berkeley short-channel insulated gate field effect transistor model 4 (BSIM4) for sub-100 nm metal-oxide-semiconductor field effect transistors (MOSFETs). Based on the genetic algorithm (GA), the monotone iterative Levenberg-Marquardt (MI-LM) method, and the neural network (NN) algorithm, this novel approach can perform simultaneous BSIM4 parameter extraction with more than 16 sub-100 nm MOSFETs in a global sense. Before extraction, all input measured I-V data are preprocessed by statistical reduction and sampling procedures. The GA and MI-LM method are then applied to calculate all parameters. The NN algorithm is used to trace the errors of I-V curves and their first derivatives, and also to inspect the variations of physical quantities. Once parameters are found, the postprocess will identify a sensible searching path so that the solution engine continues evolution until parameters reach specified stopping criteria. Good accuracy is obtained for the 90 nm NMOSFETs by several testing cases. Compared with manual fitting processes, this methodology overcomes the conventional inconvenience and is cost-effective in automatic parameter extraction. It bridges device fabrication technology and system-on-a-chip (SOC) design.

Parallel genetic algorithm for SPICE model parameter extraction

“…When the GA obtains a solution, the monotone iterative Levenberg-Marquardt method is activated to search for the nearby local optima, and the neural network algorithm suggests proper searching directions according to the current result and physical constrain. The detailed description of this extraction system is reported somewhere else [1].…”

“…The simulation program with integrated circuit emphasis (SPICE) models, such as BSIM, HiSIM, and PSP models characterize very large scale integration (VLSI) device's electrical characteristics (e.g., currentvoltage (I-V) curves), which are associated with a set of optimized parameters [1]- [5]. For the problem of the SPICE model parameter extraction, it usually refers to several hundred I-V points.…”

“…For the problem of the SPICE model parameter extraction, it usually refers to several hundred I-V points. It forms a multidimensional nonlinear optimization problem; therefore, model parameter extraction of the VLSI device is a time consuming task, and requires engineering expertise to find a set of proper parameters with reasonable physical meanings [1]- [6]. Many researches for model parameter extraction, such as pure GA or numerical optimization methods have been reported [1], [6]- [11], and most of them were performed separately.…”

“…Unfortunately, such methods may not work efficiently for VLSI devices in the regime of deep-submicron. To overcome the problem above, we have recently developed a hybrid intelligent model parameter extraction technique which bases on the genetic algorithm, the monotone iterative Levenberg-Marquardt method, and the neural network algorithm [1]. A prototype was successfully implemented according to the proposed methodology.…”

“…When the stage of GA is performed on a processor, chromosomes are simultaneously exchanged among those results that computed by its neighbor 4 processors. Optimization process is then going to the next step according to the system configuration of the hybrid intelligent model parameter extraction technique [1]. Extraction will be terminated when the specified stopping criterion is met.…”