TW459269B - A performance measure forecast system for semiconductor manufacturing - Google Patents

Abstract

This invention is a two-tier queueing network model for rapid analysis of several performance measures for a semiconductor foundry. System analyzer module provides the analysis of arrival, service pattern and overall equipment effectiveness for each tool group. System predictor module provides the forecast of important performance measures. Based on this invention, there is no limitation on the number of tool groups and product families. A solution to the problem of tool group over-lapping.

Description

4 5926 9 五、發明說明（l) ： -----— 一、發明背景 代工產業在我國工業中扮演著極為重要的角色，每 ::有極高的產值，帶動國内工業的發展。因其作業流程 、複雜、耗時且機台數目眾多，故如何掌握現場機台狀 ，並準確估測產品製程時間（pr〇duc t Cyc 1 e Ti ffle)、在製 數目（WIP)、產出量（Move)等重要性能指 easures)便成了 一項極為重要且具挑戰性的課題。 —般而言，常用的方法有：直接估測法、電腦模擬法及 解析模型法。直接估測法乃利用生產歷史資料直接作估測 並無理論依據’通常不甚準確。電腦模擬法雖較為彈性， 但因其建構過程需極大的時間與人力，且執行時間亦相對 長故應用上有其困難。相較於直接估測法及電腦模擬法， 解析模型法有其理論上的依據且執行時間極短。本發明即 為一以排隊理論（Queueing Theory)為基礎之解析模型。 文獻上，雖然排隊理論的研究行之有年，但於半導體上 的應用仍相當有限，Chen等人於1988年將BCMP排隊網路應 用於一半導體廢並得到不錯的性能指標估測結果，但其未 考慮機台故障等問題。Connors等人於1996年提出一極為 精巧的排隊網路模型且考慮機台故障問題，但其系統無法 應用於具有機台耦合特性的複雜代工廠。 雖然台灣目前半導體代工產業發展蓬勃’但目前各廠幾 乎都採用簡易的直接估測法，因此估測準確度相當有限， 且缺乏一整合性軟體輔助。4 5926 9 V. Description of the invention (l): ------ 1. Background of the invention The foundry industry plays an extremely important role in China's industry. Each :: has a very high output value and drives the development of domestic industry. . Due to its complicated operation process, time consuming, and large number of machines, how to grasp the shape of the on-site machine and accurately estimate the product process time (pr0duc t Cyc 1 e Ti ffle), the number of products in process (WIP), the production Important performance such as Move (easures) has become an extremely important and challenging subject. -In general, the commonly used methods are: direct estimation method, computer simulation method and analytical model method. The direct estimation method is based on the use of historical production data for direct estimation. There is no theoretical basis, and it is usually not very accurate. Although the computer simulation method is more flexible, its construction process requires considerable time and manpower, and its execution time is relatively long, so it has difficulties in application. Compared with the direct estimation method and computer simulation method, the analytical model method has its theoretical basis and extremely short execution time. The present invention is an analytical model based on Queueing Theory. In the literature, although research on queuing theory has been going on for many years, its application to semiconductors is still quite limited. Chen et al. Applied the BCMP queuing network to a semiconductor waste in 1988 and got good performance index estimation results, but It does not consider issues such as machine failure. Connors et al. Proposed a very sophisticated queuing network model in 1996 and considered the problem of machine failure, but its system could not be applied to complex foundries with machine coupling characteristics. Although Taiwan ’s semiconductor foundry industry is currently booming, almost all factories currently use a simple direct estimation method, so the estimation accuracy is quite limited, and it lacks an integrated software aid.

-45926 9 五、發明說明（2) 二、發明概述 在此背景之下，本人提出本發明並實作出一以此構想為 基礎之軟體系統（稱為QFAB，Queue i ng FAB)。此發明所欲 達到的目的如下所述： -能迅速預測半導體代工廠重要性能指標，如：產品製 程時間、在製品數量及產出量等，並達到一定程度的準讀 度。 —提供一分析機台群（Tool Group)之晶圓到達模式 (Arrival Pattern)及加工模式（Service Pattern) 的方法。 ~結合廠内的資料庫系統》 —能適用於具有數百部機台及複雜產品混合（pr〇duct Mix)的半導體代工廠。 本發明有以下數個特性： 一為一以機台群為基礎之分解式佇列線模型，即每一 個機台群被視為一個jj / G / c或G / G / c佇列線模型。此 一分解式概念使大型半導體廠的解析模型建構變為 可行。 —將半導體廠中常見的機台分為六類，依據其特性套 用適合之佇列線模型。 —可依需要套用數種佇列線模型，含G/G/c/Pri〇rity 及G/G[k]/c模型。 一考慮機台故障及預防維護等事件，此類事件稱為「機-45926 9 V. Description of the invention (2) 2. Summary of the invention In this context, I propose the invention and actually make a software system based on this idea (called QFAB, Queue i ng FAB). The purpose of this invention is as follows:-Can quickly predict important performance indicators of semiconductor foundries, such as: product process time, number of work in progress, output, etc., and achieve a certain degree of accuracy. —Provide a method for analyzing the Arrival Pattern and Service Pattern of the Tool Group. ~ Combined database system in the factory "— It can be applied to semiconductor foundries with hundreds of machines and complex product mix. The invention has the following characteristics: One is a decomposed line model based on machine groups, that is, each machine group is regarded as a jj / G / c or G / G / c line model. . This decomposed concept makes the construction of analytical models for large semiconductor factories feasible. -Divide the common machines in semiconductor factories into six categories, and apply the appropriate line model based on their characteristics. —Several types of line models can be applied as required, including G / G / c / Priority and G / G [k] / c models. Considering incidents such as machine failure and preventive maintenance, such incidents are called "machine

9 五、發明說明（3) 台不可用事件 提供一解決機台群耦合的方案 三、發明的詳細說明 本發明的計算流程如第一圖所示。 「資料收集」、「前處理 「诚给\〇群輪入模型由 估測」及「Goodness〜of_fit測試」.構斤」 參數 型包含「選擇適當佇列模型」、「計算成機。群佇列模 「計算性能指標」。每個機台群的模型建立：二；」J 整廠級性能指標》 〈後即可求知 第一圖中各項動作說明如下： (1)資料收集： 本發明需要的原始資料主要為晶圓開始加工 =時間點、機台與機台群的對應關係和機台晶圓：i時狀 態。 (2)前處理： 原始資料在此動作中前處理成適當形式，作為 模型的輸入值。此適當形式為：晶圓到達率（Arrivj v Rate)及到達間隔時間的標準差、「機台不可用事件」的 平均恢復時間（MTTR)及平均破壞間隔時間（MTBF)、等效機 台數目。9 V. Description of the invention (3) Station unavailable event Provide a solution to the coupling of the machine group. 3. Detailed description of the invention The calculation process of the present invention is shown in the first figure. "Data collection", "Pre-processing" "Come to the group turn model by estimation" and "Goodness ~ of_fit test". Constructing parameters Parameter types include "Choose an appropriate queue model", "Calculate into a machine. Group" Model "calculate performance indicators". Model establishment of each machine group: two; "J whole plant-level performance index" (you can find out the actions in the first figure as follows: (1) data collection: the original data required by the present invention is mainly wafers Start processing = time point, the correspondence between the machine and the machine group, and the machine wafer: i hour state. (2) Pre-processing: The raw data is pre-processed into an appropriate form in this action as the input value of the model. This appropriate form is: the standard deviation of the arrival rate of wafers (Arrivj v Rate) and the time between arrivals, the mean recovery time (MTTR) and mean time between failures (MTBF) of the "machine unavailable event", the number of equivalent machines .

^ 459 26 9 五、發明說明（4) (3 )經驗機率分佈： 如第二圖所示’將某段時間内之晶圓到達間隔時間 (Interarrival Times)及加工時間製作一經驗機率分佈 圖’在此圖表中繪出歷史值及假設其為某種機率分佈之期 望值。例如：若欲比較歷史.值是否類似一指 數（Exponential)分佈，則於該圖表中同時顯示出歷史值 與指數曲線進行比較β (4 )參數估測： 動作（3)決定各機台群的晶圓到達及加工模式之後， 晶圓到達間隔時間及加工時間機率分佈的母體 (Population)參數由樣本（Sample)參數代表之。 (5)Goodness-of-fit 測試： 計算指數（Exponential)、常態（Normal)及Erlang分佈 的Chi-Square值。由結果判定每一機台群的晶圓到達模式 與加工模式分屬何種分佈。 (6 )選擇適當之佇列線模型： 根據動作（5)的結果為每一機台群選擇一最適當的佇列 線模型’如：M/G/c、G/G/c或G/G⑴/c模盤》 (7)計算等效機台數目： 因為機台群有柄合現象’故機台群表定（n〇minal)之機459 26 9 V. Explanation of the invention (4) (3) Empirical probability distribution: As shown in the second figure, 'make an empirical probability distribution chart of the wafer arrival interval (Interarrival Times) and processing time within a certain period of time' Plot historical values and expected values assuming a certain probability distribution in this chart. For example, if you want to compare the historical value. If it is similar to an exponential distribution, then the chart will show the historical value and the exponential curve for comparison. Β (4) Parameter estimation: Action (3) determines the After the wafer arrival and processing modes, the population parameters of the wafer arrival interval and processing time probability distribution are represented by the sample parameters. (5) Goodness-of-fit test: Calculate the Chi-Square values of Exponential, Normal and Erlang distributions. Based on the results, determine the distribution of the wafer arrival mode and processing mode of each machine group. (6) Select the appropriate line model: Select the most appropriate line model for each machine group according to the result of action (5), such as: M / G / c, G / G / c, or G / "G⑴ / c die plate" (7) Calculate the number of equivalent machines: Because the machine group has a handle closing phenomenon, the machine group is determined (n〇minal).

45926 9 五、發明說明（5) 台數目無法代表該機台群實運作 一 顏，者坦υ π貝k逐仰义機台數目。為克服此 所示。 1Τ异等效機台數目之演算法，如下式 min Σ ,c)~ 2".〇 π 給定機台群g的平均等候時間公式 及η個觀察到的實際奪铉拄_ 晶圓到達率與加工，；c、及必分別為晶 圓達間隔時間及加工時間的scv(Squared CGeffic of Variance ) ° (8)計算性能指標： 由f述諸動作求得各機台群所屬之機率分佈、計算出等 t機台數目’接著即可套用適合的佇列線公式計算性能指 -、。第二圖顯不某個機台群的計算結果，「系統時間45926 9 V. Description of the invention (5) The number of units cannot represent the actual operation of the machine group. To overcome this is shown. The algorithm of the number of 1T different equivalent machines is as follows: min Σ, c) ~ 2 " .π The formula for the average waiting time for a given machine group g and η observed actual wins _ wafer arrival rate And processing, c, and scv (Squared CGeffic of Variance) which must be the wafer interval time and processing time, respectively. (8) Calculation of performance indicators: The probability distribution of each machine group by the actions described in f, Calculate the number of machines at equal t ', and then apply the appropriate line formula to calculate the performance index-,. The second figure shows the calculation results of a certain machine group, "System time

SyStem Tlme)」指晶圓在此機台群的總時間（含等候與加 夺門）’專候時間（W a i t i n g T i m e)」指晶圓在此機台 群的等候時間’ 「系統長度（System Length)」指在此機 。群内（含等候）之平均晶圓數目；「佇列長度（que ue Length)」指在此機台群前等候之晶圓數目。「AU Priority」欄位為針對所有晶圓之計算值；「pri〇i^ty 1」至rPri〇rity 3」欄位為針對個別優先度晶圓之計算"SyStem Tlme)" refers to the total time of wafers in this machine group (including waiting and adding doors) "Waiting T ime" refers to the waiting time of wafers in this machine group "" System length ( "System Length)" means here. The average number of wafers in the group (including waiting); "queue length" refers to the number of wafers waiting in front of this machine group. The “AU Priority” field is calculated for all wafers; the “pri〇i ^ ty 1” to rPri〇rity 3 ”fields are calculated for individual priority wafers

第9頁 45926 9______ 五、發明說明（6) 值；「Incapaci tation」欄位為機台不可用事件之計算 值。 (9 )整廠級性能指標預測： 根據分解式網路概念，可求得整廠級性能指標，如：產 品製程時間（Product Cycle Time)、晶圓剩餘製程時間 (Lot Remaining Cycle Time)等。計算結果如第四圖所 示’橫軸為晶圓加工步驟（Opera t ion Number )，縱轴為每 步驟之加工時間及等候時間。 軟體架構如第五圖所示，QFAB軟體透過區域網路與廠内 資料庫溝通、獲得所需之原始資料，經前處理模組（DataPage 9 45926 9______ 5. Description of the invention (6) Values; the “Incapaci tation” field is the calculated value of the machine unavailable event. (9) Forecast of whole plant-level performance indicators: According to the decomposed network concept, whole plant-level performance indicators can be obtained, such as: Product Cycle Time, Lot Remaining Cycle Time, and so on. The calculation result is shown in the fourth figure. The horizontal axis is the wafer processing steps (Operation Number), and the vertical axis is the processing time and waiting time for each step. The software architecture is shown in Figure 5. QFAB software communicates with the in-plant database through the local network to obtain the required original data.

Preprocessing Module)、作列計算模組（Queueing Manipulation Module)、系統分析模組（System Analyzer Module)及預測模組（Prediction Module)處理後可計算得 各項性能指標。第五圖中之各項目說明如下： —「半導體代工廠（Real Foundry)」：實際半導體廠 —「區域網路（LAN)」：廠内之區域網路 一「資料庫伺服器（Database Sever)」：廠内之大型 資料庫伺服器 —「原始資料庫（Raw Data)」：晶圓加工原始資料 一「QFAB」：依本發明概念發展之軟體系統 — MS SQL Sever」：本端資料庫伺服器Preprocessing Module), Queueing Manipulation Module, System Analyzer Module and Prediction Module can calculate various performance indicators after processing. The items in the fifth figure are explained as follows: — "Semiconductor Foundry (Real Foundry)": The actual semiconductor plant-"Local Area Network (LAN)": the local area network within the plant-"Database Server" ": Large database server in the factory-" Raw Data ": Raw data for wafer processing-" QFAB ": Software system developed according to the concept of the invention-MS SQL Sever": Local database server Device

第10頁 45926 9Page 10 45926 9

Claims (1)

459269 六、申請專利範圍 1. 一種半導體晶圓廠之重要性能指標估測暨機台分析系 統’其係以排隊理論為依據’其中包含一個系統分析模組 (System Analyzer Module)和一個性能指標預測模組 (Predictor Module)，此系統可分析廠内各機台群之晶圓 到達模式（Arrival Pattern)與加工模式（Service Pattern)，並準確預測含產品製程時間在内之各項性能指 標’此系統特徵是將機台依操作特性分成六類，分別對^ 類機台以適當之排隊理論模型建模、考慮機台不可用事 及晶圓優先等級不同之特性，此種方法不受實際機台、 台群及產品種類繁多之限制β 风 2_如申請專利範圍第1項所述之分析系統，其系統分 組包括下面步驟：〇)原始資料收集，（2)原始資料前' 理，（3)針對各機台群進行經驗分佈（Empirical Distribution)i 析 ’（4)參數估測，（5)G〇〇dness~0f、f. 測試，測試範圍至少包含指數（Exp〇nential)、常態 (Normal)及Erlang機率分佈。 利範圍第1項所述之分析系統，其性能指襟 (ϋ 面算步等驟效J針對各機台群選擇適當之佇珂 十算等效機台數目，（3)計算各機台群 ，（4)及計算整& & ........— 各項 3.如申請 測模組包 線模型， 性能指標459269 6. Scope of patent application 1. An important performance index estimation and machine analysis system for semiconductor wafer fabs 'based on queuing theory', which includes a System Analyzer Module and a performance index prediction Module (Predictor Module), this system can analyze the wafer arrival mode (Arrival Pattern) and processing mode (Service Pattern) of each machine group in the factory, and accurately predict various performance indicators including product process time The system feature is to divide the machine into six categories according to the operating characteristics, and model the ^ machine with an appropriate queuing theory model, considering the unavailability of the machine and the different characteristics of the wafer priority. This method is not affected by the actual machine Restrictions on various types of products, groups, and products β Wind 2_ The analysis system described in item 1 of the scope of patent application, the system grouping includes the following steps: 0) collection of raw data, (2) pre-processing of raw data, (3 ) Empirical distribution for each machine group. (4) Parameter estimation, (5) G〇〇dness ~ 0f, f. Test, the test range includes at least Includes Exponential, Normal, and Erlang probability distributions. The performance of the analysis system described in item 1 of the profit scope is as follows: (Simultaneous effects such as face counting steps, etc.) J. Select the appropriate number of equivalent machines for each machine group. (3) Calculate each machine group. (4) and calculations & & ........ — Items 3. If applying for test module enveloping model, performance index