TW201003810A

TW201003810A - Method for pattern recognition of wafer bin map and computer program product therefor

Info

Publication number: TW201003810A
Application number: TW97125943A
Authority: TW
Inventors: Wei-Cheng Lin
Original assignee: Nupoint Technology Co Ltd
Priority date: 2008-07-09
Filing date: 2008-07-09
Publication date: 2010-01-16
Also published as: TWI431706B

Abstract

A method for pattern recognition of a wafer bin map and a computer program product therefor are disclosed. Reference patterns are first defined, wherein each reference pattern is corresponding to a historical wafer bin map, and has a bin test value (BTV). A grouping step and a feature-abstracting step are performed for classifying the reference patterns having the same BTV into one group to obtain at least one pattern group; and counting the number of a certain BTV appearing in each historical wafer bin map to obtain a feature vector of each historical wafer bin map in each pattern group. With the reference patterns and the feature vectors in each pattern group, a data-mining step is performed to obtain a classifier for each pattern group. Thereafter, a proper classifier is selected, and a feature vector of a wafer bin map to be recognized is inputted thereto, thereby determining its pattern.

Description

201003810 九、發明說明【發明所屬之技術領域】本發明係有關於一種辨識晶圓圖（Wafer Bin Map)樣式（Pattern)的方法與電腦程式產品，特別是有關於一種利用Η料探勘方式之辨識晶圓圖樣式的方法與電腦程式產品。【先前技術】 ί 在積體電路（Integrated Circuits ; ic)的製作過程中，於其製程的不同階段皆會進行產品的測試步驟，並同時利用精密的分析儀器在整個製程中作有關於品質管制的各項檢驗，藉以確保製程良率及晶圓品質能夠達到最佳水準，並檢測積體電路在製造過程中所發生的瑕疵。然後，找出造成產品產生瑕疵的原因，以進一步地確保產品品質符合標準，達到提升製程良率的目的。因此，在積體電路的製造過程中’測試實為提升積體電路 ϋ元件之良率，並建立有效之資料以供工程分析使用的重要步驟。以測試之進行時機來區分，積體電路產品之測試主要可分成晶片針測（Chip Probe)與成品測試兩階段。其 ’ ’晶片針測步驟係在晶圓形式時執行，藉以在封装前先區分晶片的良赛，以避免不必要的浪費。晶片針測步驟係針對晶片作電性功能上的測試，使積體電路在進入 ,構裝前可先行過渡出電性功能不良的晶片，以避免對不 5 201003810 良品增加製造成本。晶圓上之晶片經晶片針測後，會被賦予一分類值（Bin Value) ’藉由分類值來表示晶片狀態，而形成晶圓圖（Wafer Bin Map)。請參照第1圖’其繪示晶圓圖的示意圖，其中晶圓圖1 〇被分為複數個晶片區2 〇 (如第1圖中的小格）。一晶片區20皆標示有被針測的結果，並用不同的剖面線標不其分類值，此些分類值係分別以一或多個分類代表值來表不。例如：分類代表值有BIN1、BIN2和BIN3，其中BIN3代表可用的良好晶片；BIN1及BIN2則分别代表不同異常狀態的缺陷晶片。一般由特定分類值（可由或^個BIN值的組合）的分佈情況，可判斷出晶圓的缺否存在特定的樣式。經由多片存在相同特定樣式的曰曰圓再加上製程的紀錄，便可推導出造成此結果之製程原因。如第1岡 _ 示圖所不，分類代表值BIN2的分佈狀況含有一特定樣式。 a曰圓薇對於晶圓圖的分析，仍需仰賴工程師以人工目視^^ 、式來判斷晶圓圖的缺陷樣式。然而，此201003810 IX. Description of the Invention [Technical Fields of the Invention] The present invention relates to a method for identifying a wafer pattern (Wafer Bin Map) and a computer program product, in particular, an identification method using a material exploration method. Wafer pattern style methods and computer program products. [Prior Art] ί In the manufacturing process of Integrated Circuits (IC), the test steps of the product are carried out at different stages of the process, and at the same time, the precision analysis instrument is used to make quality control throughout the process. The inspections are carried out to ensure that the process yield and wafer quality are at an optimal level and to detect flaws in the manufacturing process. Then, find out the cause of the flaw in the product, so as to further ensure that the product quality meets the standards and achieve the goal of improving the process yield. Therefore, testing in the manufacturing process of integrated circuits is an important step in improving the yield of integrated circuit components and establishing effective data for engineering analysis. Differentiated by the timing of testing, the testing of integrated circuit products can be divided into two stages: chip probe and finished product testing. The 'apos' wafer pinning step is performed in the form of a wafer to prioritize the wafer prior to packaging to avoid unnecessary waste. The wafer needle testing step is to test the electrical function of the wafer, so that the integrated circuit can be transferred to the wafer with poor electrical function before entering and constructing, so as to avoid increasing the manufacturing cost of the product. After the wafer on the wafer is subjected to wafer pinning, it is given a Bin Value. The wafer state is represented by the classification value to form a Wafer Bin Map. Referring to FIG. 1 , a schematic diagram of a wafer diagram is shown in which the wafer pattern 1 is divided into a plurality of wafer regions 2 〇 (such as the small cells in FIG. 1 ). A wafer area 20 is marked with the results of the needle test, and the different section lines are not labeled with the classification values, and the classification values are respectively represented by one or more classification representative values. For example, the classification representative values are BIN1, BIN2, and BIN3, where BIN3 represents a good wafer available; BIN1 and BIN2 represent defective wafers of different abnormal states, respectively. Generally, the distribution of specific classification values (a combination of BIN values or BIN values) can be used to determine the absence of a specific pattern of wafer defects. The process of causing this result can be derived from a number of rounds with the same specific pattern plus a record of the process. As shown in the 1st _ _ diagram, the distribution status of the classification representative value BIN2 contains a specific pattern. For the analysis of the wafer map, a 曰薇 depends on the engineer to judge the defect pattern of the wafer map by manual visual ^^. However, this

種人工目視的古-V、A L ^ 、巧會因人為主觀因素及對圖形辨識能力的差距，造成主丨阪a m ή ^ 斷、果的不一致與故障原因分類的人為Artificial-visual Gu-V, A L ^, and the difference between people's subjective factors and the ability to recognize patterns, causing the main 丨 a a m ή ^ broken, fruit inconsistency and the cause of the fault classification

偏差’因而鼓、本H …、去與逮地排除製程的問題。【發明内容】 ^⑥要發展出一種辨識晶圓圖樣式的方法與電腦私式產品，蕪曰有政地判斷出晶圓圖所示之缺陷分佈 6 201003810 樣式。明一 -4-—.. 方面為提供一種辨識晶圓圖樣式的與電腦程式產品，# 去 v y, ^ , + 藉M有效地判斷出晶圓圖所示之缺陷分佈樣式，來避免人 ^ 误判的情形，並節省人力與時間。根據本發明之眘# /1 玄卞^ +貫知例，棱供—種辨識晶圓圖樣式的 a 't ^ ^ ^ 先，提供複數個歷史晶圓圖，並定義複數個參考樣+妖，，疋 .^ ^ 式於例如一樣式庫（Pattern Library 中，其中母一個歷史曰 y) 區，此此第一晶片广曰曰®圖係被分為複數個第—晶片八el神Γ女… 經一測試步驟（例如：針測步驟）後为別“示有複數個第一八 > 支 6 - V ^ 刀類值’母一個第一分類值俜潠自稷數個分類代表值中之至係選代表不同的測試結果·而一 ’此些分類代表值係果’而母一個參考一個歷史晶圓圖，| y 可僳式係對應至至少 •… 個參考樣式且有一分類測峙扯 (Bin Test Value ; Βτ 、有一刀頰冽试值對應至歷史晶圓圖上不同此白、刀類測試值係由參考樣式所後，進行分群步驟和第_的第—一分類值組合而成。然 Ο 將具有分別與至少—個第—徵萃取步驟，其中分群步驟測試值的參考樣式歸預設分類測試值相同之分類群組，其中每—個樣雜5 群’而獲得至少一個樣式第-參考樣式’每一個第―：二考樣式中之至少-個圖之至少-個第-晶圓圖；第一 7式係對應至歷史晶圓一個第一晶圓圖之筮—曰特徵萃取步驟係計算每曰曰片區中鱼笛相同之第一分類值的數 /、弟—預設分類測試值The deviation 'and thus the drum, the H ..., go to the ground to eliminate the problem of the process. SUMMARY OF THE INVENTION ^6 To develop a method for identifying wafer pattern patterns and computer private products, 芜曰 politicize the defect distribution shown in the wafer diagram 6 201003810 style. Ming -4--.. In order to provide a way to identify the wafer pattern and computer program products, #去vy, ^, + borrow M to effectively determine the defect distribution pattern shown in the wafer map to avoid people ^ Misjudgment and save manpower and time. According to the invention, the ###1 卞卞^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ , , 疋.^ ^ is, for example, in a style library (in the Pattern Library, where the mother is a history 曰 y), this first wafer 曰曰曰曰图图 is divided into a plurality of first-wafer eight el goddess ... after a test step (for example: a needle test step), it is "displayed with a plurality of first eight" and a branch 6 - V ^ knife value 'mother one first classification value 稷 from a plurality of classification representative values The selection of the representative represents different test results. And a 'these classifications represent the value of the fruit' and the parent refers to a historical wafer map, | y can be compared to at least... reference styles and have a classification test (Bin Test Value; Βτ, a knife cheek test value corresponds to the history of the wafer map is different, the knife test value is based on the reference pattern, the grouping step and the _th categorical value combination Then, there will be a separate and at least one-thum extracting step, wherein the grouping step The reference pattern of the test value belongs to the same classification group as the preset classification test value, wherein each of the five groups is obtained by at least one style first-reference pattern 'each one': at least one of the second test styles At least one of the first wafer maps; the first 7 equation corresponds to a first wafer map of the historical wafer - the feature extraction step is to calculate the number of the first classification value of the same fish sling in each slab area /, brother - preset classification test value

私W妖曰，而霜^ A -第-特徵向量。接著，根據每：：~個第一晶圓圖的 m^ 個樣式群組所對應之 7 201003810 第一參考樣式、向量，進行一第組之一第一分類一演算法，第-Network)演算 9 支向機（Support 群。在完成每一 V' 識之一第二晶圓片區所組成’第數個第二分類值代表值中之至少值，組合出一第之第二預設分類徵萃取步驟，以值相同之第二分 -一特徵向ϊ。接：選取出其分類測分類器，其中第二參考樣式。然以判斷出第二晶在又一實施 . 一個第一晶圓圖數算出每一個第以及每一個樣式群組所對應之第—特徵賀料k勘步驟，以獲得每一個樣式群器’其中第一資料探勘步驟係根據_第 -演算法選自由一類神經網路（Neuw 决琅樹（D e c i s i ο η T r e e)演算法或_ vector machines)演算法所組成之—族 S 其中第二晶圓圖係由複數個第二晶二晶片區經該測試步驟後分別標示有複每個第二分類值係選自前述之分類一者。接著，由第二晶圓圖之第二分類二預設分類測試值，以搜尋第三晶圓圖測試值的分佈樣式。然後，進行第二特計算第二晶片區中與第二預設分類測試類值的數目’而獲得第二晶圓圖的 I自每-個樣式群組之第—分類器中， 4值與第二預設分類測試值相同之第二一分類器具有參考樣式中之至後，輸入第二特徵向量 !第—分_ ，圓圖係屬於第二參考樣式中之一者為例中’在第一特徵萃將的第-晶片區分為複數個第一巴：將：一區塊之第一晶片區中鬼，再中分別與第—預設 8 201003810 相同之第了分類值的數目後除以此第-區塊測1&的數目’ q得針對每—個第—預設分類罐之每一個第一晶圓圖的第一特徵萃取步驟，先將第二晶圓圖的第 =寺第-“鬼’其中第二區塊的數目等於】再數算出每一個第二區塊4的數目’ _制1佶士 — v弟一曰日片區中與第二預設分類測试值相同之第二分類值笛—日H ^ 的数目後除以此第二區塊之弟一曰日片區的數目，而獲得第_曰 — 于弟—日日®圖的第二特徵向量。在又-貫施例中，第二參考樣式的數目是至少三個’而辨識晶圓圖樣式的方法更至少包括：、興”應之第—特徵向量、以及第-夂考樣式之其餘者的聯集與其一冬哲次丄，丁愿之弟—特徵向量，進行第一貝料探勘步驟，以獲得第_ 于第—分類器之複數個子分類盗其中第二資料探勘步驟係根據第二々算法選自由類神經網路演算法、決二一展演算法所組成之-族群。 m法和支向機 Ο 在又—實施例中，第二演曾、车及t 栌摅士 & τ去係與第一演算法相同。根據本發明之貫施例，提 — g, ^ ^ ^ , ’、種内儲用於辨識晶圓圖樣式淨王式之電腦程式產品，卷 σ廿劫广1 田電月甸載入此電腦程式產口口並執仃後，可完成上述之、压 s ^ 辨識日日圓圖樣式的方法。因此’應用本發明之實施例，所示之缺陷分佈樣式，因而避免：有效地判斷出晶圓圖幅地節省人力與時間。避免人工誤判的情形’並大 9 201003810 f實施方式j 本發明主要是利用資料探勘的分類分群方式，並以芦前人判斷確認無誤的歷史晶圓圖資料做為樣式庫的訓練基礎，來產生自動分類分群的分類器。透過此些分類器'”，可自動並有效率的判斷待辨識晶圓圖中是否存在特^、樣式。，疋的Private W enchanting, and frost ^ A - the first - feature vector. Then, according to the 7 201003810 first reference pattern and vector corresponding to the m^ pattern groups of each of the first wafer maps, one of the first group one algorithm and the first network are calculated. 9 branching machine (Support group. At least one of the second sub-category value representative values formed in one of the second wafer areas of each V' is combined, and a second preset classification sign is combined. The extraction step is performed by the second sub-characteristic of the same value. The following is selected: the second reference pattern is selected, and the second crystal is in another implementation. A first wafer map is selected. Counting the first feature and each of the style groups corresponding to the first feature highlighting step to obtain each style grouper' wherein the first data exploration step is selected according to a type of neural network according to the _th algorithm (Neuw D 琅 ( ( T ree algorithm or _ vector machines) algorithm consists of - family S where the second wafer map is composed of a plurality of second crystal two wafer regions after the test step Marked with each second classification value The method is selected from the foregoing classification. Next, the test value is preset by the second classification 2 of the second wafer map to search for the distribution pattern of the third wafer map test value. Then, the second special calculation is performed. The number of I in the wafer area and the number of the second preset classification test class' is obtained. The I of the second wafer map is the same as the second predetermined classification test value in the first classifier of each of the pattern groups. After the second classifier has the reference pattern, the second feature vector is input, the first-minute _, and the circular image belongs to one of the second reference patterns, for example, the first-chip in the first feature extraction Divided into a plurality of first bars: will: the number of the first classification values in the first chip area of a block, and then the same as the first - preset 8 201003810, respectively, divided by the first block test 1 & The number of 'q is the first feature extraction step for each first wafer map of each of the first-predetermined classification cans, first the second wafer map of the first = "ghost" where the second region The number of blocks is equal to 】 Calculate the number of each second block 4 again _ system 1 gentleman - v brother In the next day, the number of the second categorical value flute-day H ^ which is the same as the second preset classification test value is obtained by dividing the number of the second squad of the second block by the number of the second block. The second feature vector of the Di-Day® map. In the further embodiment, the number of the second reference patterns is at least three ', and the method for recognizing the wafer pattern style includes at least: - the feature vector, and the union of the rest of the first-test styles, and a winter singer, the singer-feature vector, perform the first beech exploration step to obtain a plurality of sub-classifiers The second data exploration step is based on the second algorithm and is selected from the group consisting of a neural network algorithm and a second algorithm. m method and branching machine Ο In the further embodiment, the second acting, the car and the t gentleman & τ are the same as the first algorithm. According to the embodiment of the present invention, the g - ^ ^ ^ , ', the internal storage computer program product for identifying the wafer pattern style net king type, the volume σ 廿广 Guang 1 Tiandian Yuedian loading this computer program production After the mouth is executed, you can complete the above method of s ^ to identify the day and day chart style. Therefore, the embodiment of the present invention is applied to the defect distribution pattern shown, thereby avoiding the labor and time saving of effectively determining the wafer layout. Avoiding the situation of artificial misjudgment's large 9 201003810 f Implementation mode j The present invention mainly utilizes the classification and grouping method of data exploration, and uses the historical wafer map data confirmed by the former Lu people as the training base of the style library to generate A classifier that automatically classifies groups. Through these classifiers, it is possible to automatically and efficiently determine whether there are special patterns and patterns in the wafer map to be identified.

睛參照第2園，其係繪示根據本發明之實施例之辨識晶圓圖樣之方法的流程示意圖。在本實施例中，首先提供複數個歷史晶圓圖（步驟100)，並定義複數個參考樣式（步驟11 0)於例如一樣式庫40中。類似如第丨圖所示之晶圓圖10，每一張歷史晶圓圖係被分為複數個曰^ 區’此些晶片區經測試步驟(例如：㈣步驟)後分：不有複數個分類值，每―張歷史晶圓圖之晶片: =值係由分類代表值(例如刪、…Bm:;) 錄卜者組合而成。在步驟110巾，使用者定義各The eye is referred to the second garden, which is a schematic flow chart showing a method of identifying a wafer pattern according to an embodiment of the present invention. In this embodiment, a plurality of historical wafer maps are first provided (step 100), and a plurality of reference patterns (steps 110) are defined, for example, in a pattern library 40. Similar to the wafer pattern 10 shown in Figure ,, each historical wafer pattern is divided into a plurality of '^ areas' such wafer areas after the test steps (for example: (4) steps): no multiple The classification value, the wafer of each historical wafer map: = value is composed of the classification representative values (for example, delete, ... Bm:;). In step 110, the user defines each

類代表值分佈樣式為參考樣式。如表-所示，本實施例之-參考樣式的本貫特定的分類值所_^ 樣式係、由哪些 . Γ v 成（一或多個特定的分類代表值；即所 S月的「分類測钴姑 1 ^ t 4 # 、」）、此參考樣式的名稱、以及具有此參考樣式的歷史曰同固， -令此圖）等，以建立樣气座㈣應此參考樣式的歷史晶圓，式庫40，來供後續建立分類器使用。The class representative value distribution style is a reference style. As shown in Table--, the local specific classification value of the reference pattern of this embodiment is _^ style system, which is Γ v (one or more specific classification representative values; that is, the classification of the S month Cobalt test 1 ^ t 4 # , "), the name of this reference style, and the history of this reference style, and so on, to create a sample gas seat (4) the historical wafer should be referenced , library 40, for subsequent use of the classifier.

10 201003810 P5 BIN2 w5 p6 BIN2+BIN3 w6 p7 BIN2 W7 表一然後，進行分群步驟1 20和特徵萃取步驟1 3 0，其中可先進行分群步驟1 20、再進行特徵萃取步驟1 3 0 ;先進行特徵萃取步驟1 3 0、再進行分群步驟1 2 0 ;或同時進行分群步驟1 2 0和特徵萃取步驟1 3 0。如表二所示，其中稱分群後之參考樣式為「第一參考樣式」；分群後之歷史晶圓圖為「第一晶圓圖」。分群步驟1 20係將具有分別與預設分類測試值ΒΙΝ1、ΒΙΝ2和ΒΙΝ2 + ΒΙΝ3相同之分類測試值的參考樣式歸類為同一群，而獲得至少一個樣式群組G,、G2和G3，其中樣式群組G,具有第一參考樣式P,和P4，樣式群組G2具有第一參考樣式P2、P5和P7，樣式群組G3具有第一參考樣式P3和P6 ;第一參考樣式 P,和P4係分別對應至第一晶圓圖Wi、W2、W3和W4、 W5、W6 ;第一參考樣式P2、P5和P7係分別對應至第一晶圓圖w2、W5和W7 ;第一參考樣式P3和P6係分別對應至第一晶圓圖W3和W6。當然，每一個樣式群組G,、 G2和G3中之第一參考樣式均具有相同的分類測試值 (BTV)，即預設分類測試值之一者。樣式群組第一參考樣式名稱分類測試值 BTV 第一晶圓圖 Gi Pi、P4 BIN1 Wi、W2、W3、W4、 w5、w6 g2 P 2、P 5、P 7 BIN2 W2 ' w5 ' w7 g3 p3 ' P6 BIN2+BIN3 w3、w6 11 201003810 請參照第3圖，其係繪示用以說明本發明之杂p 特徵萃取步驟130的示意圖。在特徵萃取%例之 4, 乂鄉13〇中，貫先將母一張第一晶圓圖的晶片區分為複數個區土 5〇。接者，數算出每一個區塊5〇中之晶片區分：設分類測試值BIN1、BIN2和BIN2 + BIN3相3 ^預 :的數目後除以此區塊之晶片區的數目，而獲3得針刀類— :個預設分類測試值之每-張第—晶圓圖的、母 :。舉例而言’先將一張第一晶圓圖分個。定預設分類測試值為_，以決定要取晶二：：設組合的特徵向量。接著，計算區塊Α之，示之分類值蛊BIN i相n的鲂 θ A所標數），再卜相同的數目N(即⑽1出現的次再除以區塊A之晶片區的數目忆區塊A之特徵值為輝χκ)。在丄？，因而獲得10 201003810 P5 BIN2 w5 p6 BIN2+BIN3 w6 p7 BIN2 W7 Table 1 Then, the grouping step 1 20 and the feature extraction step 1 3 0 are performed, wherein the grouping step 1 20 can be performed first, and then the characteristic extraction step 1 3 0 is performed; Feature extraction step 130, re-grouping step 1 2 0; or simultaneous grouping step 120 and feature extraction step 130. As shown in Table 2, the reference pattern after grouping is “first reference pattern”; the historical wafer map after grouping is “first wafer map”. The grouping step 1 20 classifies the reference patterns having the same classification test values as the preset classification test values ΒΙΝ1, ΒΙΝ2, and ΒΙΝ2 + 分别3 into the same group, and obtain at least one pattern group G, G2, and G3, wherein a style group G having a first reference pattern P, and a P4 having a first reference pattern P2, P5 and P7, the pattern group G3 having a first reference pattern P3 and P6; a first reference pattern P, and P4 corresponds to the first wafer patterns Wi, W2, W3 and W4, W5, W6, respectively; the first reference patterns P2, P5 and P7 correspond to the first wafer patterns w2, W5 and W7, respectively; the first reference pattern The P3 and P6 systems correspond to the first wafer maps W3 and W6, respectively. Of course, the first reference pattern of each of the style groups G, G2, and G3 has the same classification test value (BTV), that is, one of the preset classification test values. Style group first reference style name classification test value BTV first wafer map Gi Pi, P4 BIN1 Wi, W2, W3, W4, w5, w6 g2 P 2, P 5, P 7 BIN2 W2 ' w5 ' w7 g3 p3 'P6 BIN2+BIN3 w3, w6 11 201003810 Please refer to Fig. 3, which is a schematic diagram showing the hybrid p feature extraction step 130 of the present invention. In the feature extraction example 4, 乂乡13〇, the wafer of the first wafer of the parent is first divided into a plurality of regions. The number of wafers in each block is calculated by the number: Set the number of classification test values BIN1, BIN2 and BIN2 + BIN3 to 3 ^ pre: after the number of wafer areas in this block, get 3 Needle cutters - : Each of the preset classification test values - the first - wafer map, mother:. For example, a first wafer map is first divided. The preset classification test value is _ to determine the eigenvector to be taken. Next, calculate the block ,, the classification value 蛊 BIN i phase n 鲂 θ A is marked), and then the same number N (ie, the number of (10) 1 occurrences divided by the number of wafer areas of block A The characteristic value of block A is χκκ). What? Obtained

L …。的特徵值後，便可獲得二':二圓圖，所有 ΒΙΝ1之~晶圓圖的特徵向量(長Ν又刀1測4值所示： 1食度為Ν X Μ)，如表三 3:1 5:1 5；2 10., 21ί2 23:1 26:1 32：i *?；i S3;l 56:1 68:1 ?ΰ:1 33:1 ’其BIN 1的組合所 !、G2和G3而言，和 BIN2+BIN3 ，故其中例如：3 ·· 1代表在區塊3 佔的比例為i。對每一個樣式其分别具有分類測試值β_、βίΝ2 12 201003810 可以分類測試值BINl、BIN2和BIN2 + BIN3為預設分類測試值，分別對樣式群組Gi、G2和G3内之弟·一晶圓圖萃取特徵向量。對一樣式群組的特徵向量格式係如表四所示：L .... After the eigenvalues, you can get the two ': two-circle graph, all ΒΙΝ1~ wafer map eigenvectors (long Ν 刀 knife 1 measured 4 values: 1 food Ν X Μ), as shown in Table 3 :1 5:1 5;2 10., 21ί2 23:1 26:1 32:i *?;i S3;l 56:1 68:1 ?ΰ:1 33:1 'The combination of its BIN 1!, For G2 and G3, and BIN2+BIN3, for example: 3 ·· 1 represents the proportion of block 3 in i. For each style, it has a classification test value β_, βίΝ2 12 201003810. The test values BIN1, BIN2, and BIN2 + BIN3 can be classified into preset classification test values, respectively for the pattern group Gi, G2, and G3. The graph extracts the feature vector. The eigenvector format for a style group is shown in Table 4:

p 厂：二 SG:: 3S:：I MU 35:2 45:: Θ:：! 5七：€2:1 €1:: SS:: €5乂 ~D:1 ”2:1 U 2 S:1 5t. 9:：：0:：：4:： 2：:： 23:1 26;I 3S:： S4:l SS:： ΐ"7:： SSr： 56:； 6：:： ^D:i 2S:1 =5:： «0:： §2;： ?4；l Q£:l ：0：:! ；：:：£:；：：.:! ：£t： 24:1 33;： 34:1 3S:1 40:： S2:： =9:1 €6:1 6Ί： "0:： c-0:： i5:： SD;： 3：:： 96:：：D4:： 106 3 S:1 5:：：2:1 ：E:： 16:； .5:1 H：- 25il 3D;： 3：S;1 3€:1 3S：1 ^Ol： i3:： ^6:1 5：;： 57il 60;： €1:1 65:： ¢5:1 "I:： 2 1:1 6;1 ：i：l 1€ΐΙ ：§:： 2?:： 23:： 34:： 37ϊ； 4^ii 4S:1 if:： 55：： 60·：： €3:1 ^:； "5i： 75:： ?7il §^；： £?：： 9D:： 53·： 1 :2:1 :4:: :::: 23:1 2S:i 2;:: 32:: 3S:: 3，：： G:: SO:: £2:1 Ξ9:: ¢3:: 1-:: Ή:: "S-:i 30:: E3i: SS:: S ::1 :.3: :S:: 2C:.二 22:: 2€:: 3:: 3老：：4D:: 45:: 45:: 53:: €2【二 64:: β:: ”：：： 52::拉：：S2:: W:: 2 i:: :5:: 2:: 2::: 2S:: 3::: 4，：：三2:: S£:: 6B:: Ή:: Π:: 5S:: 6S.:: Sh: 5k二：〔：：10H ::0:::: 3 0:: 5乂：2:: H::广：：23::. 5€:: 3b: 4S:: G:: 45:: H:: SS:: iS:: id 65:: "2::二：：Η:: 3:: £3:二p Factory: Two SG:: 3S::I MU 35:2 45:: Θ::! 5: €2:1 €1:: SS:: €5乂~D:1 ”2:1 U 2 S :1 5t. 9:: :0:: :4:: 2::: 23:1 26; I 3S:: S4:l SS:: ΐ"7:: SSr: 56:; 6::: ^D :i 2S:1 =5:: «0:: §2;: ?4;l Q£:l :0::! ;:::£:; ::.:! :£t: 24:1 33 ;: 34:1 3S:1 40:: S2:: =9:1 €6:1 6Ί: "0:: c-0:: i5:: SD;: 3::: 96:: :D4: : 106 3 S:1 5:: :2:1 :E:: 16:; .5:1 H:- 25il 3D;: 3:S;1 3€:1 3S:1 ^Ol: i3:: ^ 6:1 5:;: 57il 60;: €1:1 65:: ¢5:1 "I:: 2 1:1 6;1 :i:l 1€ΐΙ :§:: 2?:: 23 :: 34:: 37ϊ; 4^ii 4S:1 if:: 55:: 60·:: €3:1 ^:; "5i: 75:: ?7il §^;: £?:: 9D:: 53·: 1 :2:1 :4:: :::: 23:1 2S:i 2;:: 32:: 3S:: 3,:: G:: SO:: £2:1 Ξ9:: ¢ 3:: 1-:: Ή:: "S-:i 30:: E3i: SS:: S ::1 :.3: :S:: 2C:.22:: 2€:: 3:: 3老::4D:: 45:: 45:: 53:: €2[two 64:: β:: ”::: 52::la::S2:: W:: 2 i:: :5:: 2:: 2::: 2S:: 3::: 4,::3 2:: S£:: 6B:: Ή:: Π:: 5S:: 6S.:: Sh: 5k 2: [::10H :: 0:::: 3 0:: 5乂:2:: H::广::23::. 5€:: 3b: 4S:: G:: 45:: H:: SS:: iS:: id 65:: "2::2::Η:: 3:: £3:2

；ί:：：3:： 2：:： Γ:： 30:： 2：：： 34:： 44:： SO:： S5:； S;;： €：:： r：： 'ΐ:； :0:： £a：： SC:： S3:： r':：：OD 2 e:: Ξ;: :::: IS:: 2::: :2:: ah: M:: 3-k: H;: is:: 55:: 65:: G:: Si::ϊ::: H:: fry: 3 1:； 5;： 1：： 12a '.111 20:1 22:： 3C:： 22:1 3;:： 4^:： ΐ":： S4:l till 5£：： Γ：； ^0:1 ^S：： *：:： H:： i ' 3 ：:： 5:1 ：：;：：$:： lt：l till It:- 2tti 32:1 36:： ί：:： *5ϊ： 4":： 50:. ;3:： S€;l 6^;： €：:： 6i:： €t：l "：：： '4:1' 3 2:: S。£:: U:: i3:: ih: 3二 2€.:；l 2i:! 3:乂 3$::号2乂 *T:: 54“ 5S:: 5Si: e:.;: ¢6:: f:: li: ，5:: 1 2:2 :2:: 13::二S:: 23:: 26:: 3W: 3::: 36:: 42:: ·Τ:1 SC:: 5Ϊ:: €3:: Γ:: ~3·:: Stu: S2:: r:: ?C:: 3 S乂 U:: U:: 2::: 22:: 2(:: 25:: i::: SC:: Sk: K:: Sh: €6:1 ~6:: 50:: S.?:: Γ:: 9i:: ；e：： 9i：：：:： 20;； ZZil 2r:： 2^:： 3S;： 3S:： 4y：： 5D:： S3：： ii:： 60：： 6：：； €£；： 6?:： "S:： Ξ2:： 55;： fO;： 3 2:： £：:：；2：1 ：S:1 ：e：l 2：ii 22:： 2€;： 30:； 32:： 3S:； SS:I i：:： H:I 4€:： 50:； S5:： €0;： 62;： €S:： ~D:： "3:： 3 6;2 ：：:： 13：：：6;I 2i:I ：；t： 3D:： 2-2:1 56:： 3ϊ：； iS:：；：:： 5*:： 55:： i6：： 6·：： * = :： SD:： =;:： §6:： 5^:；：DD:：：0：：2:： ®:；：：;1 13;： 2D:： 32;Z 39:： *0;： 4i;； €C-：I 73:： -;i -J:： £3:： S5:： ?0:： 9?:1 ：Di:：：Q3:1 ：Dr：；：：£:：；；y：l 2 2;： 5:： E:1 ：!:：；.3:： 1B:1 12;： 2^:1 2S:1 3：;1 35;： i：:： i3;： C:： SC:： £3:： =5:1 62;； €=:： ~i：： ~':1 i〇：： :3:：: 3 3:1 ίϊΐ S：I 10:1 13:： 27:1 1?:： 2^:1 2€：1 2S:1 35；： 3^:1 40:1 45；： ^€v： 53:： 55i： 62:* €9:； 7：：l *"3:： ^6:2 ! 2 2:1 4:1 ":： 10:1 15:1 2Q：1 25:1 4G:1 5^:1 53:1 €5:1 T3:： "€;： £5:1 90:1 53:1 111;： 115:1 123:1 124:1 12^:： *S3:： 2 2:1 ?;： t4:l 5D;： 53:1 5€:1 59:： £2;： 64:1 7：；： *3:1 ?S:： S3:： 55:1 iC:： S3:： ??:；：D0:： iD3:：：D?;1 l：';；；：£；： 3 1:1 i;：：Q:：：3:1 I9：I Z2:： 2E:1 3：:： 57;： 4Λ；： 49:1 52:： =5:1 6：:： €4;： 6":： 52:1 55:1 = = :： 9：:： ; 表四其中左邊第一行代表第一參考樣式的編號，每一列〇代表一張第一晶圓圖的特徵向量。在本實施例中，前述之預設分類測試值可由分類代表值（例如BIN1、BIN2和BIN3等）的組合中選定，亦可由出現在參考樣式之分類測試值的組合中選定。請繼續參照第2圖。接著，進行資料探勘步驟1 40，以根據每一個樣式群組所對應之第一參考樣式、以及每 . 一個樣式群組所對應之第一晶圓圖的特徵向量，來獲得每一個樣式群組之一第一分類器，其中資料探勘步驟1 40 13 201003810 係根據一演算法’此演算法可為例如：類神經網路演算法、決策樹演算法或支向機演算法等，但本發明並不在此限。以下以支向機演算法來說明本實施例： r Ο 請參照第4圖，其係繪示本發明之實施例之使用支向機演算之法資料探勘步驟140的示意圖。如第4圖所示，對於一樣式群組中之多個參考樣式Ρι、ρ4所對應之晶圓圖的特徵向量，可將其看成散佈於空間中的資料點。支向機演算法即希望可以在這空間中找出一個超平面（Hyper-plan)，且此超平面要可以將資料點分成兩群，將分類超平面（Classification Hype卜plan ;如實線所示）定義為WX -B = 0。根據幾何知識，』却遏W向量垂直於 /刀類超平面’加入位移b的目的是要辦 + j疋要增加間隔。如果沒有b的話、則超平面將—定通 <你點會限制了這個方法的靈活性。由於本實施例要求道支持向量以及與最佳超平二 =:，因此需要知超平面’這些平行超平面可由最近的田下面的方程式來表示： (1) 透過上述步驟’可找出區面。然後，針對每一個樣式群級，樣式ρι、h的超平找出每一個樣式群組的分類器’重複上述步驟，即可器」），用以區隔此樣式群組（二下稱為「第―分類 (第―）參考樣式。 14 201003810 从工尸；τ遂之文說明。至於類神經網路演算法、決策；舉例内嫵〜瞀土楚a 束樹，，、异法或其他支向機次开法4的使用，係本發明所屬領域中具識者所熟知’故不在此贅述。 ^ 吊知 ,請繼續參照第2圖。在完成每—個樣式群 — 分類器後’便可將這些第—分類器整合至—模型中，、準備提供辨識晶圓圖樣式的服務，1 ^ /、T却個第一分类g; ί:: :3:: 2::: Γ:: 30:: 2::: 34:: 44:: SO:: S5:; S;;: €::: r:: 'ΐ:; : 0:: £a:: SC:: S3:: r':: :OD 2 e:: Ξ;: :::: IS:: 2::: :2:: ah: M:: 3-k: H;: is:: 55:: 65:: G:: Si::ϊ::: H:: fry: 3 1:; 5;: 1:: 12a '.111 20:1 22:: 3C:: 22:1 3;:: 4^:: ΐ":: S4:l till 5£:: Γ:; ^0:1 ^S:: *::: H:: i ' 3 ::: 5:1 ::;: :$:: lt:l till It:- 2tti 32:1 36:: ί::: *5ϊ: 4":: 50:. ;3:: S€;l 6^;: €: :: 6i:: €t:l "::: '4:1' 3 2:: S. £:: U:: i3:: ih: 3 2 2€.:;l 2i:! 3:乂3$::2乂*T:: 54" 5S:: 5Si: e:.;: ¢6 :: f:: li: ,5:: 1 2:2 :2:: 13::2 S:: 23:: 26:: 3W: 3::: 36:: 42:: ·Τ:1 SC: : 5Ϊ:: €3:: Γ:: ~3·:: Stu: S2:: r:: ?C:: 3 S乂U:: U:: 2::: 22:: 2(:: 25: : i::: SC:: Sk: K:: Sh: €6:1 ~6:: 50:: S.?:: Γ:: 9i:: ;e:: 9i: :::: 20;; ZZil 2r:: 2^:: 3S;: 3S:: 4y:: 5D:: S3:: ii:: 60:: 6::; €£;: 6?:: "S:: Ξ2:: 55 ;: fO;: 3 2:: £::: ;2:1 :S:1 :e:l 2:ii 22:: 2€;: 30:; 32:: 3S:; SS:I i:: : H:I 4€:: 50:; S5:: €0;: 62;: €S:: ~D:: "3:: 3 6;2 :::: 13:: :6;I 2i :I :;t: 3D:: 2-2:1 56:: 3ϊ:; iS:: ;::: 5*:: 55:: i6:: 6·:: * = :: SD:: =; :: §6:: 5^:; :DD:: :0: :2:: ®:; ::;1 13;: 2D:: 32;Z 39:: *0;: 4i;; €C- :I 73:: -;i -J:: £3:: S5:: ?0:: 9?:1 :Di:: :Q3:1 :Dr:; ::£:: ; y:l 2 2;: 5:: E:1 :!:: ;3:: 1B:1 12;: 2^:1 2S:1 3:;1 35;: i::: i3;: C :: SC:: £3:: =5:1 62;; €=:: ~i:: ~':1 i〇:: :3::: 3 3:1 ίϊΐ S:I 10:1 13: : 27:1 1?:: 2^:1 2€:1 2S:1 35;: 3^:1 40:1 45;: ^€v: 53:: 55i: 62:* €9:; 7: :l *"3:: ^6:2 ! 2 2:1 4:1 ":: 10:1 15:1 2Q:1 25:1 4G:1 5^:1 53:1 €5:1 T3:: "€;: £5:1 90:1 53:1 111;: 115:1 123:1 124:1 12^:: *S3:: 2 2:1 ?;: t4:l 5D; : 53:1 5€:1 59:: £2;: 64:1 7:;: *3:1 ?S:: S3:: 55:1 iC:: S3:: ??:; :D0:: iD3:: :D?;1 l:';; ;:£;: 3 1:1 i;: :Q:: :3:1 I9:I Z2:: 2E:1 3::: 57;: 4Λ ;: 49:1 52:: =5:1 6::: €4;: 6":: 52:1 55:1 == :: 9::: ; Table 4 where the first line on the left represents the first reference The number of the style, each column 〇 represents a feature vector of the first wafer map. In the present embodiment, the aforementioned preset classification test value may be selected from a combination of classification representative values (e.g., BIN1, BIN2, and BIN3, etc.), or may be selected from a combination of classification test values appearing in the reference pattern. Please continue to refer to Figure 2. Next, a data exploration step 144 is performed to obtain each style group according to the first reference pattern corresponding to each pattern group and the feature vector of the first wafer map corresponding to each pattern group. A first classifier, wherein the data exploration step 1 40 13 201003810 is based on an algorithm 'this algorithm can be, for example, a neural network algorithm, a decision tree algorithm or a branch machine algorithm, etc., but the present invention Not limited to this. The present embodiment is described below with a branching machine algorithm: r Ο Referring to Figure 4, there is shown a schematic diagram of a data mining step 140 using a governor calculation of an embodiment of the present invention. As shown in Fig. 4, the feature vector of the wafer map corresponding to the plurality of reference patterns Ρι, ρ4 in a pattern group can be regarded as data points scattered in the space. The branch machine algorithm hopes to find a hyper-plan in this space, and this hyperplane can divide the data points into two groups, and classify the hyperplane (Classification Hype plan; as shown by the solid line ) is defined as WX -B = 0. According to the geometric knowledge, the purpose of adding the displacement b to the W vector perpendicular to the / knife hyperplane is to do + j疋 to increase the interval. If there is no b, the hyperplane will be fixed. You will limit the flexibility of this method. Since this embodiment requires the track support vector and the best super flat ==, it is necessary to know the hyperplane 'the parallel hyperplanes can be represented by the nearest equation below the field: (1) Through the above steps, the area can be found. . Then, for each style group level, the super-flats of the styles ρι, h find the classifier of each style group 'repeating the above steps, the device"), to distinguish the style group (two is called "The first classification (the -) reference style. 14 201003810 From the work corpse; the description of the τ遂 text. As for the neural network algorithm, decision-making; example 妩瞀瞀瞀瞀 a a a , , , , , , , , , The use of the machine-opening method 4 is well known to those skilled in the art to which the present invention pertains. Therefore, it is not described here. ^ For the knowledge, please continue to refer to Figure 2. After completing each style group-classifier, Integrate these first-classifiers into the model, and prepare to provide the service to identify the wafer pattern style, 1 ^ /, T but the first classification g

器具有分類測試值（BTV)。在提供辨識晶圓圖樣式的時’ f先提供待辨識之晶圓圖（以下稱為「第二圖」K步驟15〇)，其中第二晶圓圖亦係由複數個晶片巴所組成，此些晶片區經測試步驟後分別標示有複數個分類值，每一個分類值係選自前述之分類代表值中之至少 -者。接著，由出現在第二晶圓圖的分類值種類，組合出一預設分類測試值（例如：BIN1)(步驟160)，以搜尋第二晶圓圖之預設分類測試值的分佈樣式。然後，進行特徵萃取步驟170,以計算第二晶圓圖的晶片區中與此預設分類測試值相同之分類值的數目，即第二晶圓圖的晶片區中出現此預設分類測試值的次數，而獲得第二晶圓圖的一特徵向量。與特徵萃取步驟130相類似，在特徵萃取步驟1 70中先將第二晶圓圖的晶片區分為複數個區塊，其t此些區塊的數目等於第一晶圓圖的區塊數目。接著，數算出第二晶圓圖之每一個區塊之晶片區中與此預設分類測3式值相同之分類值的數目後除以此區塊之晶片區的數目’而獲得第二晶圓圖的特徵向量。 201003810 接著’進行步驟丨8 〇，以自前述之每一個樣式群組之第一分類器中，選取出其分類測試值（BTV)與步驟160 之預設分類測試值相同的分類器（以下稱為「第二分類器」）’其中第二分類器具有至少一個參考樣式（以下稱為「第二參考樣式」）。然後’輸入第二晶圓圖的特徵向量至第二分類器（步驟1 9 〇 )，以判斷出針對步驟丨6 〇之預設分類測試值，第二晶圓圖係屬於第二參考樣式中之哪一者°在完成一種預設分類測試值後，可再回到步驟 ( 1 6 0 ’以組合出另—種預設分類測試值，再進行步驟1 7 〇至步驟1 9 0，以判斷出針對此另一種預設分類測試值，第二晶圓圖係屬於參考樣式中之哪一者。在又一實施例中，當第二分類器之第二參考樣式的數目是至少三個時，為了可以分辨出未被辨認過的樣式 (即不屬於參考樣式之任一種），本實施例提供一個複合型分類機制，以根據第二參考樣式令之一者與其對應之第一特徵向量、以及第二參考樣式之其餘者的聯集與其〇對應之第—特徵向量，進行另一資料探勘步驟，以獲得第·一为類器之複數個子分類器，其中此資料探勘步驟所使用的演算法可為例如：類神經網路演算法、決策樹演算法或支向機演算法，其中此演算法可與步驟1 40之演算法相同。以下以一例子來說明本機制：假設分類器Α有3種不同的的參考樣式，分別為 Pll、Pl2、Pi3。經由分類器A可得知某片晶圓之晶圓圖具有哪一種參考樣式的特質。此時，再額外建立三個子 16 201003810 分類器子分類器i、子分類^ 2和子分類器3,其中子分類器1歸類為由兩種參考樣式Ph以及Pn (即Pl2和P13 的如集)所建立；+分類器2歸類為由兩種參考樣式P12 以及Pm (pn和ρ!3的聯集）所建立；子分類器3歸類為由兩種參考樣式Pu以及Pk(p"和pa的聯集）所建立。绞由子刀類益1，可以得知此片晶圓之晶圓圖是否有 P u，、’二由子分類器2，可以得知此片晶圓之晶圓圖是否有Ρ〗2 ’經由子分類器3得知此片晶圚之晶圓圖是否有 Ρ〗3。絰由分類器Α所得知之晶圓圖所具有的參考樣式，應與、松由子分類器i、2和3所得知之參考樣式相同，否則可推，為此片晶圓之晶圓圖具有未被辨認過的樣式，應父、·’σ工％師進行人工辨識。此外，當第二晶圓圖被辨識成功後’第二晶圓圖與其樣式的資料可被加入至樣式庫中’而成為新的訓練基礎。、本戸'施例之辨識晶圓圖樣式的方法可使用例如電腦程式產品的型式來實施’當電腦載入此電腦程式產品並執行後，即可完成上述之辨識晶圓圖樣式的方法。由上述本實施例可知’本發明可有效地判斷出晶圓圖：示：缺陷分佈樣式’因而避免人工誤判的情形，並大巾w地I卩省人力與時間。 =本發明已以較佳實施例揭露如上，然其並非用以限疋本發明，任何熟習此技藝者’在不脫離本發明之圍内，當可作各種之更動與潤和靶視後附之申請專利範圍所界定者為準此。本發明之保護範圍當 17 201003810 【圖式簡單說明】為了更完整了解本發明及其優點，請參照上述敘述並配合下列之圖式，其中：第1圖係繪示晶圓圖的示意圖。第2圖係繪示係繪示根據本發明之實施例之辨識晶圓圖樣之方法的流程示意圖。第3圖係繪示係繪示用以說明本發明之實施例之特徵萃取步驟的示意圖。 (’ 第4圖係繪示本發明之實施例之使用支向機演算之法資料探勘步驟的示意圖。【主要元件符號說明】 10 晶圓圖 2 0 晶片區 40 樣式庫 50 區塊 U 1〇〇提供歷史晶圓圖 110 定義參考樣式 120 分群步驟 130 特徵萃取步驟 140 資料探勘步驟 150 提供待辨識之晶圓圖 . 160 組合出預設分類測試值 170 特徵萃取步驟 18 201003810 180 選取分類器 190 輸入特徵向量至分類器 A 區塊 BIN1、BIN2 、BIN3 分類代表值 N、Μ 個數 Ρ, ' Ρ4參考樣式The device has a classification test value (BTV). When the identification of the wafer pattern is provided, the wafer pattern to be recognized is first provided (hereinafter referred to as "second diagram" K step 15A), wherein the second wafer pattern is also composed of a plurality of wafers. The wafer areas are respectively labeled with a plurality of classification values after the test step, and each of the classification values is selected from at least one of the foregoing classification representative values. Next, a predetermined classification test value (for example, BIN1) is combined (step 160) from the classification value type appearing in the second wafer map to search for the distribution pattern of the preset classification test values of the second wafer map. Then, a feature extraction step 170 is performed to calculate the number of classification values in the wafer area of the second wafer map that are the same as the preset classification test value, that is, the preset classification test value appears in the wafer area of the second wafer map. The number of times to obtain a feature vector of the second wafer map. Similar to the feature extraction step 130, the wafer of the second wafer pattern is first divided into a plurality of blocks in the feature extraction step 170, wherein the number of such blocks is equal to the number of blocks of the first wafer map. Then, calculating the number of the classification values in the wafer area of each block of the second wafer map that is the same as the value of the preset classification method, and then obtaining the second crystal by dividing the number of the wafer regions of the block The feature vector of the circle. 201003810 Then, 'go step 丨8 〇, from the first classifier of each of the foregoing style groups, select the classifier whose classification test value (BTV) is the same as the preset classification test value of step 160 (hereinafter referred to as It is a "second classifier") where the second classifier has at least one reference pattern (hereinafter referred to as "second reference pattern"). Then 'input the feature vector of the second wafer map to the second classifier (step 1 9 〇) to determine the preset classification test value for step 丨6 ,, the second wafer map belongs to the second reference pattern Which one is after completing a preset classification test value, you can go back to the step (1 600 ' to combine another preset classification test value, and then go to step 1 7 to step 1 90 to Determining which of the reference patterns the second wafer map belongs to for the other preset classification test value. In still another embodiment, when the number of the second reference patterns of the second classifier is at least three In order to distinguish the unrecognized pattern (ie, not belonging to any of the reference patterns), the present embodiment provides a composite classification mechanism to correspond to the first eigenvector corresponding to one of the second reference patterns. And the first feature of the second reference pattern and the first feature vector corresponding to the ,, performing another data exploration step to obtain a plurality of sub-classifiers of the first class, wherein the data exploration step is used Calculus It can be, for example, a neural network algorithm, a decision tree algorithm, or a branch machine algorithm, where the algorithm can be the same as the algorithm of step 140. The following is an example to illustrate the mechanism: Assume that the classifier has 3 Different reference patterns are Pll, Pl2, and Pi3. The classifier A can be used to know which reference pattern the wafer map of a wafer has. In this case, three additional sub-16 201003810 classifiers are created. Sub-classifier i, sub-category ^ 2 and sub-classifier 3, wherein sub-classifier 1 is classified as being established by two reference patterns Ph and Pn (ie, sets of Pl2 and P13); + classifier 2 is classified as Two reference patterns P12 and Pm (the union of pn and ρ!3) are established; sub-classifier 3 is classified as being created by two reference patterns Pu and Pk (a combination of p" and pa). Class benefit 1, you can know whether the wafer map of this wafer has P u, and 'two sub-classifier 2, you can know whether the wafer map of this wafer has Ρ 2 'via sub-classifier 3 It is known whether the wafer pattern of the wafer is Ρ3. The wafer map known by the classifier is The reference pattern should be the same as the reference pattern known by the sub-classifiers i, 2, and 3. Otherwise, the wafer map of the wafer has an unrecognized pattern, and should be the father. % division performs manual identification. In addition, when the second wafer map is successfully identified, 'the second wafer map and its style data can be added to the pattern library' and become the new training basis. The method of recognizing the wafer pattern can be implemented by using, for example, a type of a computer program product. After the computer is loaded into the computer program product and executed, the method for recognizing the wafer pattern can be completed. The above embodiment can be known. The invention can effectively judge the wafer map: show: the defect distribution pattern' thus avoiding the situation of artificial misjudgment, and the manpower and time are saved. The present invention has been disclosed in the preferred embodiments as above, but it is not intended to limit the invention, and any person skilled in the art can make various changes and movements and target attachments without departing from the scope of the invention. The definition of the scope of patent application is the same. The scope of protection of the present invention is 17 201003810 [Simple Description of the Drawings] For a more complete understanding of the present invention and its advantages, reference is made to the above description and in conjunction with the following drawings, wherein: FIG. 1 is a schematic diagram showing a wafer diagram. Fig. 2 is a flow chart showing a method of recognizing a crystal pattern according to an embodiment of the present invention. Figure 3 is a schematic diagram showing the characteristic extraction steps for illustrating an embodiment of the present invention. (' Fig. 4 is a schematic diagram showing the steps of data exploration using the method of the branching machine calculation according to the embodiment of the present invention. [Explanation of main component symbols] 10 Wafer pattern 2 0 wafer area 40 style library 50 block U 1〇〇Providing a historical wafer map 110 Defining a reference pattern 120 Grouping step 130 Feature extraction step 140 Data exploration step 150 provides a wafer map to be identified. 160 Combining a preset classification test value 170 Feature extraction step 18 201003810 180 Selecting a classifier 190 input Feature vector to classifier A block BIN1, BIN2, BIN3 classification representative value N, Μ number Ρ, ' Ρ 4 reference pattern

1919

Claims

201003810 X. Patent Application Range 1. A method for identifying a wafer pattern pattern, comprising at least: providing a plurality of historical wafer maps (Wafer Bin Maps) and defining a plurality of reference patterns, wherein each of the historical wafer maps is Dividing into a plurality of first wafer regions, the first wafer regions are respectively labeled with a plurality of first classification values (Bin Values), and each of the first classification values is selected from a plurality of classification representative values. At least one of the classification representative values represents different test results, and each of the reference patterns corresponds to at least one of the historical wafer maps, each of the reference patterns having one a classification test value (BTV), the classification test value is a combination of different first classification values on the at least one of the historical wafer maps; performing a grouping step and a first feature extraction a step, wherein the grouping step classifies the reference patterns having the same classification test values as the at least one first preset classification test value into the same group, and obtain at least one style group, wherein each At least one first reference pattern having at least one first reference pattern, each of the at least one first reference pattern corresponding to at least one first wafer pattern of the historical wafer maps; a feature extraction step is to calculate a number of first classification values in the first wafer area of each of the at least one first wafer map that are the same as the at least one first predetermined classification test value, to obtain each of the at least one a first feature vector of the first wafer map; the at least one first reference 20 201003810 pattern corresponding to each of the pattern groups and the first wafer feature direction corresponding to each of the pattern groups And performing a first data exploration step to obtain a pattern group of the first classifier, wherein the first data query system is based on a first algorithm selected from a class of paths (Neural Netw〇) Rk) algorithm, a decision tree (Decisi〇) / support vector machines (calculated into a group); • ^ supply - wafer map, wherein the second wafer map is second Chip area, these The second chip area has no plurality of second classification values through the test step, and each of the second categories is from at least one of the representative representative values; and the second classifications of the second wafer map a value, combining a preset classification test value; performing a second feature extraction step 'to calculate the second second classification value of the second and the second predetermined classification test values to obtain the second wafer map a second feature vector; from the first classifier of each of the pattern groups, the selected class S value is the same as the second preset class full S value, wherein the second class Having at least a reference pattern of the reference patterns; and inputting the second feature vector to the second classifier to determine that the second wafer map belongs to one of the at least one second reference pattern. 2. The plurality of post-sequence scores of the first step of the identification of the wafer pattern diagram described in the first step are selected by the number of second chip regions, and the second and second breaks are taken out. Out of the person. The method of claim 21, wherein the first feature extraction step comprises: dividing the portion of each of the at least one first wafer map into a plurality of first blocks; and calculating each of the first blocks And dividing the number of the first class with the at least one first predetermined classification test value by the number of each of the first plurality of slice regions, and obtaining at least each of the at least one type of test value a first wafer pattern vector; the second feature extraction step includes: at least two second blocks of the second wafer pattern, wherein the number of the second blocks is one block And calculating a second score of the second second predetermined classification test value of each of the second blocks, dividing by the first number of each of the second blocks, and obtaining The method of identifying crystals according to the first aspect of the invention, wherein the number of the at least one second reference pattern is 4. As described in claim 3 The identification crystal method further includes at least: a first wafer wafer Di carved with the first block of a crystalline first predetermined points of the first region is divided into a plurality of equal Laid white spoon eigenvectors of the plurality of wafer region with a class number of values two wafer region. At least three of the circle styles. 22 201003810 according to a circular pattern style, according to one of the second reference patterns and a corresponding first feature vector, and a union of the remaining of the second reference patterns and a corresponding first feature vector, performing a second a data exploration step to obtain a plurality of sub-classifiers of the second classifier, wherein the second data exploration step is based on a second algorithm selected from a type of neural network algorithm and a decision tree algorithm Or a group of computer algorithms. (5) The method for identifying a wafer pattern as described in claim 4, wherein the second algorithm is the same as the first algorithm. 6. The identification wafer as described in claim 1 The method of the pattern pattern further includes: combining a third preset classification test value from the second classification values of the second wafer map; performing a third feature extraction step to calculate the second region a second eigenvector of the second wafer region of the second wafer region of the Q, and obtaining a third eigenvector of the second wafer map; And selecting, by the first classifier, a third classifier whose classification test value is the same as the third preset classification test value, wherein the third classifier has at least one third reference pattern of the reference patterns; And inputting the third feature vector to the third classifier to determine that the second wafer map belongs to one of the at least one third reference pattern. 23 201003810 7. As described in claim 6 a method of identifying a wafer map style, The third feature extraction step includes: dividing the second wafers of the second wafer into a plurality of second blocks, wherein the number of the second blocks is equal to the number of the first blocks; And calculating a number of second classification values in the second chip area of each of the second blocks that are the same as the third preset classification test value, and dividing by the each C*:: the second blocks The third feature vector of the second wafer map is obtained by the number of the second wafer regions. 8. The method for identifying a wafer pattern according to claim 1, wherein the at least one The number of the three reference patterns is at least three. 9. The method for recognizing the wafer pattern as described in claim 8 further includes: 〇 according to one of the third reference patterns and the corresponding one thereof Performing a third data exploration step to obtain a plurality of sub-classifiers of the third classifier, wherein the third feature search vector and the first feature vector corresponding to the other of the third reference patterns perform a third data search step, wherein the third Data exploration steps are based on a The third algorithm is selected from the group consisting of a neural network algorithm, a decision tree algorithm and a machine-oriented algorithm. 24 201003810 10. As described in the scope of claim 9 A method of identifying a wafer pattern, wherein the third algorithm is the same as the method of identifying a wafer pattern. 11. The method for identifying a wafer pattern as described in claim 1 further comprises: The classification representative value is combined with the first preset classification test value. 12. The method for identifying the wafer pattern as described in claim 1 further comprises: at least: The classification test value combines the at least one first preset classification test value. 1 3. The identification pattern of the wafer pattern as described in claim 1 is defined in a style library. (Pattern Library) 0 25 201003810 16. - Internal storage for identifying the wafer pattern product, when the computer loads the computer program product and applies for the computer program execution of the identification wafer pattern program described in the first paragraph of the patent scope , The method may be achieved as styles.

26