TWI820399B - Wafer processing method and wafer processing system - Google Patents
Wafer processing method and wafer processing system Download PDFInfo
- Publication number
- TWI820399B TWI820399B TW110107111A TW110107111A TWI820399B TW I820399 B TWI820399 B TW I820399B TW 110107111 A TW110107111 A TW 110107111A TW 110107111 A TW110107111 A TW 110107111A TW I820399 B TWI820399 B TW I820399B
- Authority
- TW
- Taiwan
- Prior art keywords
- polishing
- wafer
- torque signal
- wafer processing
- interval
- Prior art date
Links
- 238000012545 processing Methods 0.000 title claims abstract description 119
- 238000003672 processing method Methods 0.000 title claims abstract description 31
- 238000005498 polishing Methods 0.000 claims abstract description 306
- 230000007704 transition Effects 0.000 claims description 29
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 claims description 26
- 229910052802 copper Inorganic materials 0.000 claims description 26
- 239000010949 copper Substances 0.000 claims description 26
- 238000013145 classification model Methods 0.000 abstract 1
- 235000012431 wafers Nutrition 0.000 description 165
- 238000012549 training Methods 0.000 description 27
- 238000013527 convolutional neural network Methods 0.000 description 12
- 238000010586 diagram Methods 0.000 description 9
- 238000000034 method Methods 0.000 description 8
- 238000007781 pre-processing Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- 238000012795 verification Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000004519 manufacturing process Methods 0.000 description 4
- 230000000717 retained effect Effects 0.000 description 4
- 239000012634 fragment Substances 0.000 description 3
- 239000000126 substance Substances 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000005055 memory storage Effects 0.000 description 2
- 239000002184 metal Substances 0.000 description 2
- 229910052751 metal Inorganic materials 0.000 description 2
- 238000007517 polishing process Methods 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 239000004065 semiconductor Substances 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- 230000003321 amplification Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000011017 operating method Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000002310 reflectometry Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Landscapes
- Mechanical Treatment Of Semiconductor (AREA)
- Finish Polishing, Edge Sharpening, And Grinding By Specific Grinding Devices (AREA)
Abstract
Description
本發明是有關於一種方法及系統，且特別是有關於一種晶圓加工方法及晶圓加工系統。The present invention relates to a method and a system, and in particular, to a wafer processing method and a wafer processing system.
CMP（Chemical-Mechanical Planarization）製程為化學機械平坦化技術，亦可稱為化學機械拋光（Chemical-Mechanical Polishing）技術，由IBM（International Business Machines Co.）將此技術用於晶圓平坦化，可有效改善金屬層間之介電質層的平坦化，減少半導體元件內堆疊層數增加所產生曝光聚焦困難的影響，因此於目前半導體廠在積體電路的製作上廣為使用此一技術。The CMP (Chemical-Mechanical Planarization) process is a chemical mechanical planarization technology, also known as chemical mechanical polishing (Chemical-Mechanical Polishing) technology. This technology is used by IBM (International Business Machines Co.) for wafer planarization. It effectively improves the planarization of the dielectric layer between metal layers and reduces the impact of exposure focusing difficulties caused by the increase in the number of stacked layers in semiconductor components. Therefore, this technology is widely used in the production of integrated circuits in current semiconductor factories.
在不同拋光機的廠商中，化學機械拋光的製程終點偵測的訊號源以及準則也不同，市上的系統常以渦電流（eddy current）系統偵測金屬層的厚度、光學系統量測晶圓表面反射率判斷不同材料的裸露來找出製程終點，兩者系統必須改裝拋光機的盤面所花費成本極高，且拋光盤轉動一圈，感測器才量測一次，會有時間上的延遲。Different polishing machine manufacturers have different signal sources and criteria for end-of-process detection of chemical mechanical polishing. Systems on the market often use eddy current systems to detect the thickness of metal layers and optical systems to measure wafers. The surface reflectivity determines the exposure of different materials to find the end point of the process. Both systems must modify the polishing machine's disk, which is extremely costly, and the polishing disk rotates once before the sensor measures it once, which will cause a time delay. .
本創作希望透過擷取伺服馬達放大器的扭矩訊號，將連續性的訊號做訊號處理後，丟入卷積神經網路做訓練，以最佳模型做終點偵測及訊號預測，並可套用於不同系統之機台。This creation hopes to capture the torque signal of the servo motor amplifier, perform signal processing on the continuous signal, and then throw it into the convolutional neural network for training, using the best model for end point detection and signal prediction, and can be applied to different applications. System machine.
本發明提供一種晶圓加工方法及晶圓加工系統，其可透過經訓練的拋光模型來依據拋光扭矩訊號準確地判斷出晶圓的區間資訊及/或拋光的終點時間，進而改善先前技術的缺點。The present invention provides a wafer processing method and a wafer processing system, which can accurately determine the interval information of the wafer and/or the end time of polishing based on the polishing torque signal through a trained polishing model, thereby improving the shortcomings of the prior art. .
本發明的一種晶圓加工方法適於控制晶圓加工台對晶圓進行的拋光。晶圓加工方法包括：取得晶圓加工台對晶圓拋光的拋光扭矩訊號；並將拋光扭矩訊號輸入CNN產生之拋光辨識模型以定義晶圓的區間資訊。A wafer processing method of the present invention is suitable for controlling the polishing of wafers by a wafer processing table. The wafer processing method includes: obtaining the polishing torque signal of the wafer polishing by the wafer processing table; and inputting the polishing torque signal into the polishing identification model generated by CNN to define the interval information of the wafer.
一種晶圓加工方法適於控制晶圓加工台對晶圓進行的拋光。晶圓加工方法包括：取得晶圓加工台對晶圓拋光的拋光扭矩訊號；並將拋光扭矩訊號輸入CNN產生之拋光預測模型產生接續於拋光扭矩訊號之後的預測拋光扭矩訊號。A wafer processing method is adapted to control polishing of wafers by a wafer processing station. The wafer processing method includes: obtaining the polishing torque signal of the wafer polishing by the wafer processing station; and inputting the polishing torque signal into the polishing prediction model generated by CNN to generate a predicted polishing torque signal subsequent to the polishing torque signal.
本發明的一種晶圓加工系統適於晶圓進行拋光。晶圓加工系統包括晶圓加工台及電子裝置。晶圓加工台對晶圓進行拋光。電子裝置耦接晶圓加工台。電子裝置包括記憶體及處理器。記憶體儲存拋光辨識模型。處理器耦接記憶體及晶圓加工台。處理器取得晶圓加工台對晶圓拋光的拋光扭矩訊號。處理器藉由拋光辨識模型以依據拋光扭矩訊號產生晶圓的區間資訊，處理器依據區間資訊控制晶圓加工台的拋光。A wafer processing system of the present invention is suitable for polishing wafers. The wafer processing system includes a wafer processing table and electronic devices. A wafer processing station polishes the wafers. The electronic device is coupled to the wafer processing stage. Electronic devices include memory and processors. Memory storage of polished identification models. The processor is coupled to the memory and the wafer processing table. The processor obtains the polishing torque signal of the wafer polishing by the wafer processing table. The processor uses the polishing recognition model to generate interval information of the wafer based on the polishing torque signal, and the processor controls the polishing of the wafer processing table based on the interval information.
本發明的一種晶圓加工系統適於晶圓進行拋光。晶圓加工系統包括晶圓加工台及電子裝置。晶圓加工台對晶圓進行拋光。電子裝置耦接晶圓加工台。電子裝置包括記憶體及處理器。記憶體儲存拋光模型。處理器耦接記憶體及晶圓加工台。處理器取得晶圓加工台對晶圓拋光的拋光扭矩訊號。處理器藉由拋光預測模型以依據拋光扭矩訊號產生接續於拋光扭矩訊號之後的預測拋光扭矩訊號，處理器依據預測拋光扭矩訊號控制晶圓加工台的拋光。A wafer processing system of the present invention is suitable for polishing wafers. The wafer processing system includes a wafer processing table and electronic devices. A wafer processing station polishes the wafers. The electronic device is coupled to the wafer processing stage. Electronic devices include memory and processors. Memory storage polish model. The processor is coupled to the memory and the wafer processing table. The processor obtains the polishing torque signal of the wafer polishing by the wafer processing table. The processor uses the polishing prediction model to generate a predicted polishing torque signal following the polishing torque signal based on the polishing torque signal, and the processor controls polishing of the wafer processing table based on the predicted polishing torque signal.
基於上述，晶圓加工系統及晶圓加工方法可準確地判斷被拋光的程度以及拋光時程，因而有效降低晶圓在進行拋光時的風險，在有效提升整體製造效率的同時改善所產晶圓的品質Based on the above, the wafer processing system and wafer processing method can accurately determine the degree of polishing and the polishing schedule, thereby effectively reducing the risk of wafer polishing, effectively improving the overall manufacturing efficiency while improving the wafers produced. quality
圖1為本發明實施例一電子系統1的示意圖。電子系統1包括晶圓加工台10、伺服裝置11、介面卡12及電子裝置13。晶圓加工台10可用來對晶圓（未繪示於圖1）進行拋光，並提供拋光時的拋光扭矩訊號。伺服裝置11可監控晶圓加工台10的操作，並接收晶圓加工台10所提供的拋光扭矩訊號。介面卡12耦接於伺服裝置11及電子裝置13之間進行訊號通訊，介面卡12可將拋光扭矩訊號以適合的訊號格式提供至電子裝置13。電子裝置13中儲存有拋光終點模型（例如為拋光辨識模型）及/或拋光預測模型（例如為拋光預測模型），電子裝置13即可接收拋光扭矩訊號來對拋光終點模型進行訓練。FIG. 1 is a schematic diagram of an
詳細而言，晶圓加工台10可例如但不僅限於晶圓精密拋光機台HAMAI HS-720C，晶圓加工機台10可用來對晶圓進行化學機械平坦化（chemical-mechanical planarization, CMP）或化學機械研磨（chemical-mechanical polishing），以平坦化晶圓的表面。更具體而言，晶圓可被安置在晶圓加工台10的一拋光頭（未繪示於圖1）上，且晶圓的下方可被晶圓加工台10的拋光盤拋光。進一步，晶圓加工台10可將拋光盤的扭矩訊號及拋光頭的扭矩訊號共同整合為拋光扭矩訊號，並將拋光扭矩訊號提供至伺服裝置11。In detail, the wafer processing table 10 can be, for example, but not limited to, the wafer precision polishing machine HAMAI HS-720C. The
伺服裝置11可例如但不僅限於伺服放大機Mitsubishi servo amplifier。伺服裝置11可監控晶圓加工台10的操作並取得晶圓加工台10的拋光扭矩訊號。在一實施例中，伺服裝置11可對取得的拋光扭矩訊號進行簡單的訊號預處理，例如為訊號放大、去雜訊等處理的任意排列組合。伺服裝置11可以各種方式來實現。在一實施例中，伺服裝置11可為控制器與晶圓加工台10整合設置在同一機體中，控制晶圓加工台10的操作並取得拋光扭矩訊號。在一實施例中，伺服裝置11可分離式地設置並外接於晶圓加工台10，介以取得晶圓加工台10的拋光扭矩訊號。The
介面卡12耦接於伺服裝置11及電子裝置13之間。介面卡12可例如為但不僅限於介面卡NI DAQ USB-6351。介面卡12可將拋光扭矩訊號以相容的訊號格式提供至電子裝置13。The
電子裝置13包含處理器130及記憶體131。電子裝置13可例如但不僅限於移動台、高級移動台（advanced mobile station, AMS）、伺服器、客戶端、桌上型電腦、筆記型電腦、網路型電腦、工作站、個人數位助理（personal digital assistant, PDA）、個人電腦（personal computer, PC）、平板電腦等。只要電子裝置13可具有處理器130及記憶體131即可。The
處理器130可例如但非僅限於中央處理單元（Central Processing Unit，CPU），或是其他可程式化之一般用途或特殊用途的微控制單元（Micro Control Unit，MCU）、微處理器（Microprocessor）、數位信號處理器（Digital Signal Processor，DSP）、可程式化控制器、特殊應用積體電路（Application Specific Integrated Circuit，ASIC）、圖形處理器（Graphics Processing Unit，GPU）、算數邏輯單元（Arithmetic Logic Unit，ALU）、複雜可程式邏輯裝置（Complex Programmable Logic Device，CPLD）、現場可程式化邏輯閘陣列（Field Programmable Gate Array，FPGA）或其他類似元件或上述元件的組合。The
記憶體131可例如但非僅限於任何型態的固定式或可移動式的隨機存取記憶體（Random Access Memory，RAM）、唯讀記憶體（Read-Only Memory，ROM）、快閃記憶體（Flash Memory）、硬碟（Hard Disk Drive，HDD）、固態硬碟（Solid State Drive，SSD）或類似元件或上述元件的組合。The
更具體而言，處理器130耦接於記憶體131，記憶體131中可儲存有拋光終點模型。處理器130可接收拋光扭矩訊號來訓練拋光終點模型。More specifically, the
圖2A～2J為本發明實施例處理器130處理拋光扭矩訊號的流程示意圖。更具體而言，拋光扭矩訊號經過處理器130如圖2A～2J所繪示的處理，可用來訓練拋光終點模型。2A to 2J are schematic flow charts of the
詳細而言，在圖2A中，處理器130可獲得晶圓加工台10所提供原始（或僅經過伺服裝置11預處理）的拋光扭矩訊號。舉例而言，拋光扭矩訊號可在100Hz的取樣頻率下所取得的。Specifically, in FIG. 2A , the
在圖2B中，拋光扭矩訊號可經過濾波器的濾波。舉例而言，拋光扭矩訊號可經過巴特沃斯帶通濾波器（Butterworth band pass filter）的濾波，以濾除低頻及高頻訊號。In Figure 2B, the polishing torque signal can be filtered by the filter. For example, the polishing torque signal can be filtered through a Butterworth band pass filter to filter out low-frequency and high-frequency signals.
在圖2C中，可針對濾波後的拋光扭矩訊號計算其能量。舉例而言，可針對濾波後的拋光扭矩訊號進行方均根運算，以取得拋光扭矩訊號的能量振幅。In Figure 2C, the energy can be calculated for the filtered polishing torque signal. For example, a root mean square operation can be performed on the filtered polishing torque signal to obtain the energy amplitude of the polishing torque signal.
在圖2D中，拋光扭矩訊號可進行平滑化處理，以利後續的模型訓練過程。舉例而言，拋光扭矩訊號可經中值濾波器（median filter）的處理而平滑化。In Figure 2D, the polishing torque signal can be smoothed to facilitate the subsequent model training process. For example, the polished torque signal can be smoothed by processing with a median filter.
因此，圖2A～2D可為拋光扭矩訊號進行前置處理的流程示意圖。不過如前述所說明的，拋光扭矩訊號包含有拋光盤的扭矩訊號及拋光頭的扭矩訊號，圖2A～2D為了說明方便僅繪示有拋光盤的扭矩訊號及拋光頭的扭矩訊號的其中一者，但本領域具通常知識者可知，圖2A～2D對拋光扭矩訊號所進行的前置處理是對拋光盤的扭矩訊號及拋光頭的扭矩訊號兩者共同進行的。Therefore, FIGS. 2A to 2D can be schematic flow diagrams of pre-processing the polishing torque signal. However, as explained above, the polishing torque signal includes the torque signal of the polishing disk and the torque signal of the polishing head. Figures 2A to 2D only show one of the torque signal of the polishing disk and the torque signal of the polishing head for the convenience of explanation. , but those with ordinary knowledge in the art will know that the pre-processing of the polishing torque signal in Figures 2A to 2D is performed on both the torque signal of the polishing disk and the torque signal of the polishing head.
接著，在圖2E中，可針對拋光扭矩訊號進行裁切，以移除拋光扭矩訊號中起始時間之前，晶圓加工台10尚未對晶圓進行拋光的部分，以及移除拋光扭矩訊終點時間之後（即後段時間）。在一實施例中，拋光扭矩訊號的裁切可將拋光扭矩訊終點時間之後一段時間之後的訊號移除，也就是保留終點時間之後一段的過拋時間，並將過拋時間之後的拋光扭矩訊號移除。Next, in FIG. 2E , the polishing torque signal can be trimmed to remove the portion of the polishing torque signal that has not yet been polished by the
更具體而言，晶圓中可具有銅膜區間及過渡區間，晶圓加工台10在對晶圓的拋光過程是起始於晶圓的銅膜區間開始被拋光，且拋光過程是在晶圓的過渡區間被拋光完之後而結束。晶圓加工台10對晶圓的拋光扭矩訊號可預先經過分析及標記，進而取得拋光扭矩訊號中的起始時間及終點時間，用來於圖2E所繪示的流程中進行拋光扭矩訊號的裁切。More specifically, the wafer may have a copper film section and a transition section. The polishing process of the wafer by the
在圖2F中，處理器130可區分出晶圓加工台10對晶圓的拋光是進行到銅膜區間或過渡區間，並判斷出兩區間交界處的時間（即前段時間）。In FIG. 2F , the
在判斷出兩區間交界處的時間時，可在終點時間往前一過渡區間的粗估時間長度TL。並在過渡區間的粗估時間長度TL中，取對過渡區間的粗估時間長度TL中拋光扭矩訊號具有最大值的時間做為銅膜區間與過渡區間交界處的時間（即前段時間）。When determining the time at the junction of two intervals, the end time can be used to roughly estimate the time length TL of the previous transition interval. And in the rough estimated time length TL of the transition interval, the time when the polishing torque signal has the maximum value in the rough estimated time length TL of the transition interval is taken as the time at the junction of the copper film interval and the transition interval (ie, the previous period).
在圖2G中，處理器130可依據前段時間來將拋光扭矩訊號區分為兩部分，其包含有銅膜區間扭矩訊號及過渡區間扭矩訊號。In FIG. 2G , the
因此，在圖2E～2G中，處理器130可依據晶圓的區間來對拋光扭矩訊號進行切割或分類，以適於進行拋光終點模型的訓練。Therefore, in FIGS. 2E to 2G , the
在圖2H中，拋光扭矩訊號可再進一步的進行切割以產生時間長度相等的拋光扭矩訊號片段。舉例而言，拋光扭矩訊號可以移動窗口（sliding window）的方式來進行切割。In FIG. 2H , the polishing torque signal can be further cut to generate polishing torque signal segments with equal time lengths. For example, the polishing torque signal can be cut in a sliding window manner.
在圖2I中，切割完的拋光扭矩訊號片段可依據晶圓的區間來進行標示，以標示為對應於銅膜區間或過渡區間。舉例而言，對應於銅膜區間的拋光扭矩訊號片段可加註上數位值0來進行標記，對應於過渡區間的拋光扭矩訊號片段可加註上數位值1來進行標記。In Figure 2I, the polishing torque signal segments after cutting can be marked according to the section of the wafer, so as to correspond to the copper film section or the transition section. For example, the polishing torque signal segment corresponding to the copper film interval can be marked with a digital value of 0, and the polishing torque signal segment corresponding to the transition interval can be marked with a digital value of 1.
在圖2J中，切割完且標示完的拋光扭矩訊號片段可進行亂數排列（shuffle），使拋光扭矩訊號依據亂數排列，降低拋光扭矩訊號片段之間在排列上與時間的相關性，以利於拋光終點模型的訓練。In Figure 2J, the cut and marked polishing torque signal segments can be randomly arranged (shuffled), so that the polishing torque signals are arranged according to random numbers, thereby reducing the correlation between the polishing torque signal segments in arrangement and time, so as to Conducive to the training of polished endpoint models.
因此，在圖2H～2J中，處理器130可對拋光扭矩訊號進行切割、標記以及重新洗牌，以提供適合的拋光扭矩訊號片段來訓練拋光終點模型。Therefore, in Figures 2H-2J, the
更具體而言，訓練拋光終點模型時，可將圖2A～2J所產生的拋光扭矩訊號片段以及拋光扭矩訊號片段的標記可隨機批量地丟入卷積神經網路（convolutional neural network）來進行訓練，卷積神經網路可具有多層卷積層，每層卷積層內有多個過濾器，過濾器內部儲存有多個權重值，每個權重會隨機初始化產生。進一步，拋光扭矩訊號片段可經過多層過濾器的乘積、池化去除雜訊、以及激活函數放大或縮小特徵值後，進到全連接層返回這兩類的機率值。接著以機率模型的損失函數計算該值與拋光扭矩訊號片段的標記的損失，加總該批訓練資料的損失，進而更新權重值，持續此動作直到所有資料皆丟入訓練，即為完成一次完整的訓練，並於完成一次訓練後返回這次所訓練出的模型的準確度、召回率及特異率，最後進行多種參數多次完整訓練，找出最佳模型，以訓練出拋光終點模型。More specifically, when training the polishing end point model, the polishing torque signal segments generated in Figures 2A to 2J and the markers of the polishing torque signal segments can be randomly thrown into the convolutional neural network (convolutional neural network) in batches for training. , a convolutional neural network can have multiple convolutional layers. Each convolutional layer has multiple filters. Multiple weight values are stored inside the filter, and each weight is randomly initialized. Furthermore, the polished torque signal fragment can be passed through the product of multi-layer filters, pooling to remove noise, and the activation function to amplify or reduce the feature value, and then enter the fully connected layer to return the probability values of these two types. Then use the loss function of the probabilistic model to calculate the loss of this value and the mark of the polished torque signal segment, add up the loss of this batch of training data, and then update the weight value. Continue this action until all data are thrown into the training, which is a complete training, and after completing one training, the accuracy, recall rate and specificity rate of the trained model are returned. Finally, multiple complete trainings with multiple parameters are performed to find the best model to train a polished end-point model.
在一實施例中，上述的卷積神經網路電路可以透過處理器130內部的電路來實現，也就是說處理器130可以在其內部進行拋光終點模型的訓練，並於訓練完成後將模型儲存在記憶體131中。在另一實施例中，上述的卷積神經網路電路可以透過分離於處理器130，且特別設計以應用於卷積神經網路計算的電路來實現。處理器130可透過將拋光扭矩訊號片段輸入至拋光終點模型來進行訓練，並於訓練完成後將模型儲存在記憶體131中。In one embodiment, the above-mentioned convolutional neural network circuit can be implemented through the circuit inside the
因此，經過上述訓練過程的拋光終點模型可依據拋光扭矩訊號片段以及拋光扭矩訊號片段的標記被訓練，如此一來，經過訓練的拋光終點模型即可接收拋光扭矩訊號來判斷晶圓加工台對晶圓在進行拋光的區間。Therefore, the polishing end point model that has gone through the above training process can be trained based on the polishing torque signal fragments and the marks of the polishing torque signal fragments. In this way, the trained polishing end point model can receive the polishing torque signal to determine the wafer processing table alignment. The circle is in the polishing range.
圖3A～3D為本發明實施例處理器130處理拋光扭矩訊號的流程示意圖。更具體而言，拋光扭矩訊號經過處理器130如圖3A～3D所繪示的處理，可用來訓練拋光預測模型。3A to 3D are schematic flow charts of the
雖然未繪示，不過圖3A可接續於圖2A～2D的處理之後。也就是說，在圖2A～2D對拋光扭矩訊號所進行的前置處理之後，可再接續進行圖3A的處理。另外，為了方便說明，雖然圖3A～3D中所繪示的拋光扭矩訊號僅有拋光頭扭矩訊號，但本發明不限於此。處理器130所接收的拋光扭矩訊號可如圖2A～2J中所繪示的包含有拋光頭扭矩訊號及拋光盤扭矩訊號。Although not shown, FIG. 3A may continue after the processing of FIGS. 2A-2D. That is to say, after the pre-processing of the polishing torque signal in Figures 2A to 2D, the processing of Figure 3A can be continued. In addition, for convenience of explanation, although the polishing torque signals shown in FIGS. 3A to 3D are only polishing head torque signals, the present invention is not limited thereto. The polishing torque signal received by the
詳細而言，在圖3A中，處理器130可針對拋光扭矩訊號進行裁切，以移除拋光扭矩訊號中起始時間之前，晶圓加工台10尚未對晶圓進行拋光的部分，以及移除拋光扭矩訊終點時間之後（即後段時間）。在一實施例中，拋光扭矩訊號的裁切可將拋光扭矩訊終點時間之後一段時間之後的訊號移除，也就是保留終點時間之後一段的過拋時間，並將過拋時間之後的拋光扭矩訊號移除。Specifically, in FIG. 3A , the
在圖3B中，處理器130可依據時序對拋光扭矩訊號進行切割，以產生多個互相對應的前段時間長度的拋光扭矩訊號片段及後段時間長度的拋光扭矩訊號片段。In FIG. 3B , the
在圖3C中，切割完的拋光扭矩訊號片段可進行洗牌，使拋光扭矩訊號依據亂數排列，降低拋光扭矩訊號片段之間在排列上與時間的相關性，以利於拋光預測模型的訓練。In Figure 3C, the cut polishing torque signal segments can be shuffled so that the polishing torque signals are arranged according to random numbers, thereby reducing the correlation between the arrangement and time of the polishing torque signal segments, thereby facilitating the training of the polishing prediction model.
在圖3D中，切割完的多個前段長度拋光扭矩訊號片段及相對應的後段長度拋光片段即可用來訓練拋光預測模型。In Figure 3D, the multiple polishing torque signal segments of the front length after cutting and the corresponding polishing segments of the rear length can be used to train the polishing prediction model.
更具體而言，拋光扭矩訊號可被區分為訓練群組及驗證群組，訓練群組的拋光扭矩訊號可被用來進行拋光預測模型的訓練，而驗證群組的拋光扭矩訊號可被用來進行拋光預測模型的驗證。更具體而言，訓練群組中的前段長度拋光扭矩訊號片段及後段長度拋光扭矩訊號片段可被隨機批量地丟入卷積神經網路進行訓練，卷積神經網路可具有多層卷積層，每層卷積層有多個過濾器(內部有多個權重值)，每個權重會隨機初始化產生，訓練資料經過多層過濾器的乘積，池化去除雜訊，以及激活函數放大或縮小特徵值後，進到全連接層返回預測出來的時序資料，接著以回歸模型的損失函數計算預測值與相對應後段時間長度的拋光扭矩訊號片段的損失，加總該批訓練資料的損失，進而更新權重值，持續此動作直到所有資料皆丟入訓練，即為完成一次完整的訓練，並於完成一次訓練後返回這次所訓練出的模型的對訓練資料的損失。接著，驗證群組的拋光扭矩訊號可被丟入卷積神經網路進行驗證，驗證群組中的前段長度拋光扭矩訊號片段可被丟入拋光預測模型，使拋光預測模型產生預測的後段長度拋光扭矩訊號片段，透過返回預測的後段長度拋光扭矩訊號片段以檢驗後段長度拋光扭矩訊號片段的損失，最後進行多種參數多次完整訓練，找出最佳模型，以訓練出拋光預測模型。More specifically, the polishing torque signal can be divided into a training group and a verification group. The polishing torque signal of the training group can be used to train the polishing prediction model, and the polishing torque signal of the verification group can be used. Validation of the polishing prediction model was performed. More specifically, the front-length polished torque signal segments and the rear-length polished torque signal segments in the training group can be randomly thrown into the convolutional neural network in batches for training. The convolutional neural network can have multiple convolutional layers, each of which The convolutional layer has multiple filters (with multiple weight values inside), and each weight is randomly initialized. After the training data is multiplied by multiple layers of filters, pooling removes noise, and the activation function amplifies or shrinks the feature values, Entering the fully connected layer returns the predicted time series data, and then uses the loss function of the regression model to calculate the loss of the predicted value and the polished torque signal segment corresponding to the later period of time, sum up the loss of the batch of training data, and then update the weight value. Continue this action until all data is thrown into training, which means a complete training is completed. After completing a training, the loss of the training data for the trained model is returned. Then, the polishing torque signal of the verification group can be thrown into the convolutional neural network for verification. The front-length polishing torque signal segments in the verification group can be thrown into the polishing prediction model, so that the polishing prediction model generates the predicted back-segment length polishing. For the torque signal segment, the loss of the back-segment length polishing torque signal segment is tested by returning the predicted back-segment length polishing torque signal segment. Finally, multiple complete trainings with multiple parameters are performed to find the best model to train the polishing prediction model.
圖4A為本發明實施例一晶圓加工系統4的示意圖。晶圓加工系統4包括晶圓加工台40及電子裝置41。電子裝置41包括處理器410及記憶體411。圖4A所繪示的晶圓加工台40相似於圖1所繪示的晶圓加工台10，可對晶圓（未繪示於圖1）進行拋光，並提供拋光時的拋光扭矩訊號。電子裝置41可接收拋光扭矩訊號。電子裝置41的記憶體411可儲存有拋光終點模型及/或拋光預測模型。處理器410耦接記憶體411及晶圓加工台40，其中處理器410可取得晶圓加工台40對晶圓拋光的拋光扭矩訊號。處理器410可依據拋光終點模型及拋光扭矩訊號產生晶圓的區間資訊。處理器410則可依據產生的區間資訊來控制晶圓加工台40的拋光。FIG. 4A is a schematic diagram of a
雖然圖4A中未繪示，不過在相似於圖1的一替代實施例中，晶圓加工系統4可另外包含伺服裝置及介面卡串聯耦接於晶圓加工台40及電子裝置41之間，用來聯絡晶圓加工台40與電子裝置41。Although not shown in FIG. 4A , in an alternative embodiment similar to FIG. 1 , the
在一實施例中，當電子裝置41的記憶體411中儲存有拋光終點模型時，電子裝置4即可接收晶圓加工台40所提供的拋光扭矩訊號來判斷晶圓的區間資訊，並據此控制晶圓加工台40對晶圓的拋光。In one embodiment, when the
圖4B為本發明實施例一處理器410依據拋光終點模型及拋光扭矩訊號產生的操作波型示意圖。圖4B中繪示了曲線L41～L44，其中曲線L41為拋光頭扭矩訊號、曲線L42為拋光盤扭矩訊號，曲線L43為拋光終點模型所判斷的銅膜區間機率值，曲線L44為拋光終點模型所判斷的過渡區間機率值，其中曲線L41、L42可為晶圓加工台40所提供至電子裝置41的拋光扭矩訊號，而曲線L43、L44可為電子裝置41依據曲線L41、L42及拋光終點模型所產生的輸出波型，電子裝置41可依據曲線L43、L44判斷出晶圓的區間資訊。在一實施例中，曲線L41、L42可為經過訊號預處理的拋光扭矩訊號。或是在另一實施例中，處理器410也可直接接收未經訊號預處理的拋光扭矩訊號的曲線，以判斷出銅膜區間機率值及過渡區間機率值。FIG. 4B is a schematic diagram of the operation waveform generated by the
在一實施例中，處理器410可依據拋光終點模型及曲線L41、L42判斷晶圓加工台40對晶圓的拋光是進行到晶圓的銅膜區間或過渡區間，並產生曲線L43的銅膜區間機率值及曲線L44的過渡區間機率值。In one embodiment, the
詳細而言，在時間T0時，當晶圓加工台40開始對晶圓進行拋光時，處理器410依據拋光終點模型及曲線L41、L42所產生的曲線L43、L44中，曲線L43的銅膜區間機率值為相對高且曲線L44的過渡區間機率值為相對低。因此，處理器410可據此判斷晶圓的區間資訊，也就是在時間T0時，晶圓加工台40是在對晶圓的銅膜區間進行拋光。Specifically, at time T0, when the
在時間T1時，處理器410依據拋光終點模型及曲線L41、L42所產生的曲線L43、L44中產生了交錯。也就是說，在時間T1時，過渡區間機率值大於等於銅膜區間機率值。因此，處理器410可據此判斷出晶圓的區間資訊，也就是在時間T1時，晶圓加工台40是由晶圓的銅膜區間拋光至晶圓的過渡區間。At time T1, an intersection occurs in the curves L43 and L44 generated by the
在時間T2時，處理器410依據拋光終點模型及曲線L41、L42所產生的曲線L43、L44中即產生了再一次的交錯。也就是說，在時間T2時，銅膜區間機率值大於等於過渡區間機率值。因此，處理器410可據此判斷出晶圓的區間資訊，也就是在時間T2時，晶圓加工台40完成晶圓的過渡區間的拋光，因此，處理器410可據此判斷出時間T2為拋光終點。At time T2, another interleaving occurs in the curves L43 and L44 generated by the
如此一來，晶圓加工系統4即可依據預先訓練的模型來精準地判斷出晶圓加工台對晶圓進行拋光的區間資訊。電子裝置41可在判斷出達到拋光終點時停止晶圓加工台40的拋光操作，或是電子裝置41可在判斷出達到拋光終點後的一預設時間長度後停止晶圓加工台40的拋光操作，有效提升整體製造效率及所產晶圓的品質。In this way, the
另一方面，在一實施例中，當電子裝置41的記憶體411中儲存有拋光預測模型時，電子裝置41即可接收晶圓加工台40所提供的拋光扭矩訊號來產生接續於拋光扭矩訊號之後的預測拋光扭矩訊號。On the other hand, in one embodiment, when the
綜合比較拋光終點模型及拋光預測模型兩者，電子裝置41依據拋光終點模型所產生的區間資訊以終點判斷是介於拋光扭矩訊號的時間範圍中。也就是說，區間資訊可用來在拋光扭矩訊號的時間範圍中指出晶圓是被加工至銅膜區間或過渡區間，終點判斷則可在拋光扭矩訊號的時間範圍中判斷晶圓的拋光是否已達到終點，以及當達到終點時的時間點。By comprehensively comparing the polishing end point model and the polishing prediction model, the
另一方面，電子裝置41依據拋光預測模型所產生的預測拋光扭矩訊號是接續於拋光扭矩訊號。也就是說，預測拋光扭矩訊號可用來預測接下來晶圓加工台接著對晶圓加工的操作所產生的拋光扭矩訊號。On the other hand, the predicted polishing torque signal generated by the
在一實施例中，當電子裝置41的記憶體411同時儲存有拋光終點模型及拋光預測模型時，電子裝置41可同時依據晶圓加工台40所提供的拋光扭矩訊號進行預測及終點判斷。In one embodiment, when the
更具體而言，電子裝置41可依據拋光扭矩訊號及拋光預測模型來產生接續於拋光扭矩訊號之後的預測拋光扭矩訊號。接著，電子裝置41可依據預測拋光扭矩訊號及拋光終點模型來產生出區間資訊及終點判斷。如此一來，電子裝置41即可針對接收的拋光扭矩訊號來進行區間資訊及終點判斷的預測。More specifically, the
在另一實施例中，電子裝置41的記憶體411同時儲存有拋光終點模型及拋光預測模型，而電子裝置41除了可針對接收的拋光扭矩訊號來進行區間資訊及終點判斷的預測，電子裝置41還可同時針對接收的拋光扭矩訊號來產生區間資訊及終點判斷。In another embodiment, the
圖5A為本發明實施例一晶圓加工方法的流程圖，上述晶圓加工系統4中關於拋光終點模型的操作，可歸納為圖5A中所繪示的晶圓加工方法。在圖5A的步驟S50中，可取得晶圓加工台40對晶圓拋光的拋光扭矩訊號。接著，在步驟S51中，可藉由拋光終點模型以依據拋光扭矩訊號產生晶圓的區間資訊。關於步驟S50、S51的細節可參考上方相關段落，於此不另贅述。FIG. 5A is a flow chart of a wafer processing method according to an embodiment of the present invention. The operations of the polishing end point model in the above-mentioned
簡言之，圖4A的晶圓加工系統4依據圖5A所繪示的晶圓加工方法可產生晶圓的區間資訊及終點判斷。晶圓加工系統4依據區間資訊及終點判斷可準確且有效率地進行晶圓的加工，避免人工操作的誤判以及冗長操作流程。In short, the
圖5B為本發明實施例一晶圓加工方法的流程圖，上述晶圓加工系統4中關於拋光預測模型的操作，可歸納為圖5B中所繪示的晶圓加工方法。在圖5B的步驟S52中，可取得晶圓加工台40對晶圓拋光的拋光扭矩訊號。接著，在步驟S53中，可藉由拋光預測模型以依據拋光扭矩訊號產生接續於拋光扭矩訊號之後的預測拋光扭矩訊號。關於步驟S52、S53的細節可參考上方相關段落，於此不另贅述。FIG. 5B is a flow chart of a wafer processing method according to an embodiment of the present invention. The operations of the polishing prediction model in the
簡言之，圖4A的晶圓加工系統4依據圖5B所繪示的晶圓加工方法可對晶圓加工台40接續的加工操作進行預測而產生預測拋光扭矩訊號。晶圓加工系統4可依據預測拋光扭矩訊號來針對尚未發生的時間範圍進行預測，使晶圓加工系統4可依據預測結果獲得更充足的判斷時間，進而改善加工晶圓中的未知風險。In short, the
綜上所述，本發明的晶圓加工系統及晶圓加工方法可依據晶圓加工台的拋光扭矩訊號來進行拋光的判斷。經訓練的拋光模型可依據拋光扭矩訊號來判斷出晶圓的區間資訊及/或終點時間，或者經訓練的拋光模型可依據拋光扭矩訊號來預測接下來的拋光扭矩訊號。如此一來，晶圓加工系統及晶圓加工方法可準確地判斷被拋光的程度以及拋光時程，因而有效降低晶圓在進行拋光時的風險，在有效提升整體製造效率的同時改善所產晶圓的品質。In summary, the wafer processing system and wafer processing method of the present invention can make polishing judgments based on the polishing torque signal of the wafer processing table. The trained polishing model can determine the interval information and/or end time of the wafer based on the polishing torque signal, or the trained polishing model can predict the next polishing torque signal based on the polishing torque signal. In this way, the wafer processing system and wafer processing method can accurately determine the degree of polishing and the polishing schedule, thereby effectively reducing the risk of wafer polishing, effectively improving the overall manufacturing efficiency while improving the quality of the produced wafers. round quality.
1:電子系統
10、40:晶圓加工台
11:伺服裝置
12:介面卡
13、41:電子裝置
130、410:處理器
131、411:記憶體
4:晶圓加工系統
L41、L42、L43、L44:曲線
S50、S51、S52、S53:步驟
T0、T1、T2:時間
TL:粗估時間長度
1:
圖1為本發明實施例一電子系統的示意圖。 圖2A～2J為本發明實施例處理器處理拋光扭矩訊號的流程示意圖。 圖3A～3D為本發明實施例處理器處理拋光扭矩訊號的流程示意圖。 圖4A為本發明實施例一晶圓加工系統的示意圖。 圖4B為本發明實施例一處理器依據拋光終點模型及拋光扭矩訊號產生的操作波型示意圖。 圖5A為本發明實施例一晶圓加工方法的流程圖。 圖5B為本發明實施例一晶圓加工方法的流程圖。 FIG. 1 is a schematic diagram of an electronic system according to an embodiment of the present invention. 2A to 2J are schematic flow diagrams of a processor processing polishing torque signals according to an embodiment of the present invention. 3A to 3D are schematic flow diagrams of a processor processing polishing torque signals according to an embodiment of the present invention. FIG. 4A is a schematic diagram of a wafer processing system according to an embodiment of the present invention. 4B is a schematic diagram of an operation waveform generated by a processor based on the polishing end point model and the polishing torque signal according to an embodiment of the present invention. FIG. 5A is a flow chart of a wafer processing method according to an embodiment of the present invention. FIG. 5B is a flow chart of a wafer processing method according to an embodiment of the present invention.
S50、S51:步驟 S50, S51: steps
Claims (9)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW110107111A TWI820399B (en) | 2021-02-26 | 2021-02-26 | Wafer processing method and wafer processing system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW110107111A TWI820399B (en) | 2021-02-26 | 2021-02-26 | Wafer processing method and wafer processing system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| TW202234185A TW202234185A (en) | 2022-09-01 |
| TWI820399B true TWI820399B (en) | 2023-11-01 |
Family
ID=84957102
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| TW110107111A TWI820399B (en) | 2021-02-26 | 2021-02-26 | Wafer processing method and wafer processing system |
Country Status (1)
| Country | Link |
|---|---|
| TW (1) | TWI820399B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW200614408A (en) * | 2004-09-24 | 2006-05-01 | Applied Materials Inc | Endpoint compensation in electroprocessing |
| TW200908120A (en) * | 2007-04-27 | 2009-02-16 | Sharp Kk | Manufacturing method of semiconductor device |
| TW201323149A (en) * | 2011-10-21 | 2013-06-16 | Strasbaugh | Systems and methods of wafer grinding |
| TW201412457A (en) * | 2012-08-28 | 2014-04-01 | Ebara Corp | Method of monitoring a dressing process and polishing apparatus |
| TW201923881A (en) * | 2017-10-31 | 2019-06-16 | 日商荏原製作所股份有限公司 | Polishing apparatus, polishing system, substrate processing apparatus, polishing method, and program |
| WO2019198622A1 (en) * | 2018-04-11 | 2019-10-17 | 日揮触媒化成株式会社 | Polishing composition |
| TW201947401A (en) * | 2018-05-16 | 2019-12-16 | 台灣積體電路製造股份有限公司 | Condition monitoring method for semiconductor manufacturing tool and semiconductor manufacturing system |
| TW202007481A (en) * | 2018-07-31 | 2020-02-16 | 台灣積體電路製造股份有限公司 | Apparatus and methods for chemical mechanical polishing |
| US20200101579A1 (en) * | 2018-09-27 | 2020-04-02 | Ebara Corporation | Polishing apparatus, polishing method, and machine learning apparatus |
-
2021
- 2021-02-26 TW TW110107111A patent/TWI820399B/en active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW200614408A (en) * | 2004-09-24 | 2006-05-01 | Applied Materials Inc | Endpoint compensation in electroprocessing |
| TW200908120A (en) * | 2007-04-27 | 2009-02-16 | Sharp Kk | Manufacturing method of semiconductor device |
| TW201323149A (en) * | 2011-10-21 | 2013-06-16 | Strasbaugh | Systems and methods of wafer grinding |
| TW201412457A (en) * | 2012-08-28 | 2014-04-01 | Ebara Corp | Method of monitoring a dressing process and polishing apparatus |
| TW201923881A (en) * | 2017-10-31 | 2019-06-16 | 日商荏原製作所股份有限公司 | Polishing apparatus, polishing system, substrate processing apparatus, polishing method, and program |
| WO2019198622A1 (en) * | 2018-04-11 | 2019-10-17 | 日揮触媒化成株式会社 | Polishing composition |
| TW201947401A (en) * | 2018-05-16 | 2019-12-16 | 台灣積體電路製造股份有限公司 | Condition monitoring method for semiconductor manufacturing tool and semiconductor manufacturing system |
| TW202007481A (en) * | 2018-07-31 | 2020-02-16 | 台灣積體電路製造股份有限公司 | Apparatus and methods for chemical mechanical polishing |
| US20200101579A1 (en) * | 2018-09-27 | 2020-04-02 | Ebara Corporation | Polishing apparatus, polishing method, and machine learning apparatus |
Also Published As
| Publication number | Publication date |
|---|---|
| TW202234185A (en) | 2022-09-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Nasir et al. | Acoustic emission monitoring of sawing process: artificial intelligence approach for optimal sensory feature selection | |
| Arun et al. | Tool condition monitoring of cylindrical grinding process using acoustic emission sensor | |
| Moia et al. | Tool condition monitoring of aluminum oxide grinding wheel in dressing operation using acoustic emission and neural networks | |
| Mannan et al. | Application of image and sound analysis techniques to monitor the condition of cutting tools | |
| Wang et al. | Hidden Markov model-based tool wear monitoring in turning | |
| US10654148B2 (en) | Systems and methods for conditioning blades | |
| KR102218204B1 (en) | Polishing-amount simulation method for buffing process, and buffing device | |
| US8992286B2 (en) | Weighted regression of thickness maps from spectral data | |
| Miranda et al. | Monitoring single-point dressers using fuzzy models | |
| CN111127533B (en) | Multi-feature fusion grinding wheel grinding performance prediction method based on neural network | |
| Adibi et al. | Grinding wheel loading evaluation using digital image processing | |
| TWI820399B (en) | Wafer processing method and wafer processing system | |
| CN116551475B (en) | Grinding processing method and system for hardware tool | |
| JP2009004442A (en) | Polishing method for semiconductor wafer | |
| JPH1110535A (en) | Grinding wheel longevity judging device | |
| TW201249598A (en) | Varying coefficients and functions for polishing control | |
| JP6437608B1 (en) | Polishing apparatus, polishing method, and polishing control apparatus | |
| Pilný et al. | Acoustic emission-based in-process monitoring of surface generation in robot-assisted polishing | |
| US11921474B2 (en) | Virtual metrology method using convolutional neural network and computer program product thereof | |
| CN116306047B (en) | Aluminum alloy part surface polishing control method and system | |
| KR102374525B1 (en) | Keyword Spotting Apparatus, Method and Computer Readable Recording Medium Thereof | |
| EP3736648A1 (en) | Method for autonomous optimization of a grinding process | |
| JP2014086733A5 (en) | ||
| Ferreira et al. | Inferential measurement of the dresser width for the grinding process automation | |
| Safont et al. | Comparison of dimensionality reduction methods for road surface identification system |