智慧型模擬演算法在奈米金屬氧化物半導體場效電晶體特性最佳化設計之研究

Abstract

參數萃取在半導體元件模式發展中是很重要的一項步驟。目前有許多等效模式已被用於模擬金屬氧化物半導體電晶體(MOSFET)的特性，而準確的模式參數便是用來描述該元件各項特性的重要依據。參數萃取是一個找出各個參數使其電路特性符合預期結果的最佳化問題。而隨著元件尺寸不斷的縮小，等效模式及其所需的參數也變的愈來愈複雜，況且參數萃取必須同時對多種不同尺寸的元件做最佳化的動作，一些傳統方法如數值及統計方法已變得不敷使用。本論文提出一個智慧型參數萃取系統用以萃取奈米級BSIM4 MOSFET模式參數並輔以平行計算技術以增加效能。此系統由基因演算法、類神經網路、李文伯格-馬奎特法(Levenberg-Marquardt method, LM)及BSIM4模式萃取經驗等主要部分所構成。基於對模式分析的結果，此方法一開始便會對目標曲線作資料簡化的前處理動作；前處理完成後，基因演算法便會搜尋最佳參數組以降低誤差，同時LM法會在基因演算法找出一個大略的解之後開始啟動並快速的找出附近的最佳解；而類神經網路則是經由辨析曲線之間的形狀關係及物理參數來引導基因演算法作更有效的搜尋。在求出一組符合停止條件的參數組後，後處理會檢測此參數是否符合物理現象與電特性，以確保解的正確與可用性。在幾項90與130奈米NMOSFET元件的實驗證明下，其結果與量測資料比較後均能展現極高的準確度，同時與一般方法比較，我們的方法有很高的萃取效能，也因此其可有效降低參數萃取所需的人力與時間。此系統架構將來更可運用在其他等效模式的參數萃取、系統晶片設計及高頻積體電路設計等領域。

Semiconductor model parameters extraction plays an important role between device foundries and integrated circuit (IC) design companies yet a bottleneck in microelectronics industry. Various compact models have been of great interest and studied for nanoscale metal-oxide-semiconductor field effect transistor (MOSFET) device simulation. The model parameters extraction intrinsically characterizes properties of designed and fabricated devices. It leads to a multidimensional optimization problem to be solved and extracted efficiently for the applications to very large scaled integrated (VLSI) circuit and system-on-a-chip (SoC) design. Different approaches, such as empirical, numerical, and statistical methods have been proposed to solve this problem. However, these approaches encounter serious problems in nanoscale MOSFET era, such as poor accuracy, time-consuming, ineffective extraction process, and lack the predict capacity in practical applications. In this work we present an intelligent hybrid system for BSIM4 model in nanoscale MOSFET model parameters extraction. This approach combines with the genetic algorithm, the numerical optimization, the neural network, and the empirical experiences has been proposed and successfully developed on a Linux based personal computer (PC) cluster system. First of all, the preprocess including empirical constrains and data reduction is performed to reduce massive computation depending on the continuity of the model. The genetic algorithm is then applied to extract a rough solution and the numerical optimization is functioned to obtain an further improved solution. The neural network is adopted for curves and physical quantities inspection, this procedure adjusts the evolutionary direction of genetic algorithm to enhance the extraction performance. After a set of parameters is found and a sensible search path is detected, the genetic algorithm keeps evolving the next generation until a set of optimal parameters reaches certain stopping criteria. Excellent accuracy has been obtained in several experiments for both the 90 nm and 130 nm NMOSFET devices in terms of IDS-VDS and IDS-VGS curves, and related physical quantities. Comparing to conventional approaches, the proposed methodology contained the advantage of both numerical and soft computing methods solves the extracting problem cost efficiently. The proposed extraction architecture significantly reduces the demand of time and enhances the working performance in the practice use, which is useful in VLSI circuit and SoC design.