張時中臺灣大學：工業工程學研究所廖柏鈞Liao, Bo-JiunBo-JiunLiao2007-11-262018-06-292007-11-262018-06-292006http://ntur.lib.ntu.edu.tw//handle/246246/51190半導體供應鏈系統是由許多提供生產製造服務之節點所構成，其提供多樣化的服務，有別於傳統的製造業，此新興的製造服務業需要新的作業管理方法。在半導體供應鏈系統，服務等級差異化也就是產品具有優先等級順序是一種常見的生產作業模式。本研究之挑戰是在於有服務等級區分之情況下，如何克服供應鏈系統之可擴縮性及可預測性，和隨著產品種類及製程變動快速所引起的系統變異增加之問題。服務等級差異化確保服務的品質是可區分、是有差別的，因而依據不同的服務品質來分配有限的產能及採取不同之定價策略；而變異則會影響此供應鏈系統及此系統中的各節點之行為表現。為了描繪具有服務等級區分的半導體供應鏈之生產管理特性，本研究需符合以下幾點：生管特性指標必須是可預測的、方法必須可適用於供應鏈之系統面及節點面、基本了解具有變異之供應鏈系統及各節點的生管特性。 本研究為了提升具有變異及服務等級區別的半導體供應鏈管理之可擴縮性與可預測性，致力於發展生管特性模型及模型建構之方法。我們以晶圓廠生管特性模型之建構為目標，並視其為半導體供應鏈管理之基石。晶圓廠生管特性模型描繪了優先等級順序、資源分配與各項變動如何影響晶圓廠之行為表現，例如：生產週期時間、在製品數量、產出量與機台利用率之平均值與變異數。晶圓廠可以同時製造處理多種產品、且有許多需要停機進行修護調整的機台設施，同時又具有迴流特性的固定製程步驟，而我們將其視為一個沒有機台停機、沒有批次生產、有多種服務等級產品的迴流式開放排隊網路模型，並設計開發一種混合分解式近似法來解決具有服務等級區分之排隊網路問題。混合分解式近似法是藉由排隊網路分析法將晶圓廠網路模型分解成多個獨立的服務節點；並採用次序分解近似法依序建置各種服務等級產品在單一服務節點之生管特性，並研究網路中各節點的關係，將各節點串接起來進而了解晶圓廠網路系統的行為表現。混合分解式近似法可克服供應鏈系統之可擴縮性，能將系統面之行為表現分解至各節點面，亦能將節點面之行為表現聚集組成系統面。由混合分解式近似法所建構而成的網路模型也能提升供應鏈系統之可預測性，能因應不同的系統輸入而迅速估算各節點面及系統面行為表現指標之平均值與變異數。此外，我們還進一步調整生管特性模型，使其結果能更符合經驗上的數據。 用具有服務等級區分之晶圓廠模型的模擬結果和我們設計的混合分解式近似法作比較，發現混合分解式近似法在正確性及計算效率上皆有潛力可應用在真實具有服務等級區分的晶圓廠。生產週期時間的相對誤差大部分都在10%以內；只有服務等級較高之產品其標準差相對誤差高達50%，但其絕對誤差卻也很小(大約1.2~1.6)。而應用混合分解式近似法去計算具有60個製程之晶圓廠模型， CPU 2.8 GHz之個人電腦在4秒內即可計算完畢，其計算速度大約是模擬的百倍至千倍間。我們也可將混合分解式近似法運用在供應鏈最佳化方面，將各式各樣的輸入選擇，像是產能分配、不同服務等級之產品比例…等進行迅速評估來決定最佳的晶圓廠輸入以達到我們所希望看到的行為表現。晶圓廠生管特性模型可提供晶圓廠或半導體供應鏈管理者一個可進行設若分析之工具。A semiconductor supply chain is a system of nodes that provide manufacturing services—in fact, a variety of services. The new paradigm of manufacturing services requires new methods of operation management. Service differentiation, namely, prioritization, is common in operations of semiconductor supply chain (SSC). The grand challenges of this research will be to overcome the scalability and predictability problems with respect to service differentiation, and variability that are exacerbated by rapidly increasing product varieties and process variations in the chains. The service differentiation ensures that the quality of service (QoS) is differentiable, and quality of service affects how to charge prices and allocate resources. And the variability affects the performance of both network nodes and chains of nodes. In order to model the behavior of the SSC, our research is needed in follow aspects: performance metrics that are predictable, scalable with respect to chain structure, and fundamental understanding of the behavior of nodes and chains under variability. This study is to develop behavior models and modeling methods that enhance the scalability and predictability of the semiconductor supply chain management with respect to varieties, and service differentiation. We aim at the behavior modeling of fabs that provides a cornerstone for supply chain management. The fab behavior models describe how priority, resource allocation and sources of variations affect fab performance metrics such as mean and variability of cycle time, wafer-in-process, throughputs, and machine utilizations. Considering a fab with multiple part types, failure prone machines, and re-entrant process flows as a failure-free, batch-free and re-entrant OQN. Then we design and develop a hybrid decomposition approximation-based approach for network modeling with a focus on capturing operation priority and variations in fabs. The hybrid decomposition approximation-based approach (hybrid SDA+QNA) is to decompose the fab network model into many independent service nodes by QNA and model single service node behavior by sequential decomposition approximation (SDA) among priorities in one node, then study networking relationship among service nodes to approximate network performances. The hybrid decomposition approximation can handle scalability that allows chain metrics to be decomposed into node metrics. And the priority network models constructed by the hybrid decomposition approximation can handle predictability that allows very quick evaluation of mean and variability of both node and system level output performance metrics with various input options. Besides, we consider model tuning to fit the empirical data. Comparisons with simulation results over two priorities fab models demonstrate the accuracy and computing efficiency of hybrid decomposition approximation-based approach and its potential for applications of real fab with service differentiation. The relative errors of cycle time performances are mostly within 10%; only for the cycle time standard deviation of hot lots, the relative error is high to 50% but the absolute error is very small (about 1.2 ~1.6). Applications of hybrid SDA+QNA to fab models with 60 processing steps only require less than 4 seconds of CPU time on a 2.8 GHz personal computer, which is about 2 to 3 orders faster than simulation. Quick evaluation of various input options in terms of capacity allocation, priority mix, etc. may be combined with supply chain optimization to determine the best fab input option that leads to a desirable performance. The fab behavior model therefore provides a tool for what-if analysis to fab or supply chain planners/ managers.Chapter 1 Introduction 1 1.1 Needs for Priority Fab Model in Semiconductor Supply Chain 1 1.2 Literature Review 3 1.2.1 Common Needs in Semiconductor Supply Chain 3 1.2.2 Response Surface Modeling 4 1.2.3 Approximation Analysis of Priority Queues 5 1.2.4 Queueing Network Model for Semiconductor Manufacturing 6 1.3 Scope of Research 7 1.4 Thesis Organization 11 Chapter 2 Behavior Modeling Problem for Fab with Priority Queues 13 2.1 Modeling for Supply Chain Configuration 13 2.1.1 Configuration of Semiconductor Supply Chain 14 2.1.2 Response Surface Methodology for Behavior Modeling 16 2.1.3 Needs and Issues for Fab Behavior Modeling 17 2.2 Open Queueing Network-Based Fab Behavior Modeling 18 2.2.1 Fab Modeling 19 2.2.2 Open Queueing Network Model 20 2.3 Decomposition-Based OQN Analysis 22 2.3.1 Introduction of Decomposition Approximation 22 2.3.2 Queueing Network Analysis 23 2.4 Challenges of Fab Behavior Modeling under Service Differentiation 29 2.4.1 Single Node Behavior Modeling with Priority 29 2.4.2 Fab Network Modeling with Priority 30 2.5 Summary 32 Chapter 3 Sequential Decomposition Approximation Modeling of Single Service Node 33 3.1 Sequential Decomposition Approximation among Priorities 33 3.1.1 Concept of SDA 34 3.1.2 Equivalent Service Time 35 3.1.3 Individual Priority Queue Performance Measures 42 3.2 Algorithm of Sequential Decomposition Approximation 46 3.3 Implementation and Verification of SDA 49 3.3.1 Verification with Respect to M/M/1:PR 49 3.3.2 Verification with Respect to GI/G/1:PR 52 3.4 Model Tuning of Single Service Node 54 3.5 Summary 57 Chapter 4 Hybrid Decomposition Approximation-Based Network Modeling 59 4.1 Priority Network Modeling for Fab 59 4.1.1 Priority Fab Queueing Network Model 60 4.1.2 Hybrid Combination of SDA and QNA 61 4.1.3 Interactions among Nodes of Re-entrant Flows 62 4.2 Algorithm of Hybrid Decomposition Approximation 64 4.3 Network Level Performance Measures 69 4.4 Verification of Hybrid Decomposition Approximation 70 4.4.1 Multi-Nodes Tandem Queue 70 4.4.2 Multi-Nodes with Re-entrant Line 73 4.5 Summary 75 Chapter 5 Numerical Experiments 77 5.1 Fab Data Description 77 5.1.1 FAB1: Single Product Model 77 5.1.2 FAB2: Two Priorities Model 79 5.1.3 Empirical Data of Fab Models 81 5.2 Numerical Results of Fab Models 82 5.2.1 Numerical Results of FAB1 82 5.2.2 Numerical Results of Special FAB2 85 5.2.3 Numerical Results of General FAB2 88 5.3 Summary 92 Chapter 6 Conclusions and Future Research 93 6.1 Conclusions 93 6.2 Directions of Future Research 95 Appendix A Calculations of Remaining Service Time 97 Appendix B Moments of Queue Length 99 Appendix C Data of Node Level Performances in Cycle Time 106 Bibliography 109744044 bytesapplication/pdfen-US生管特性模型服務等級模擬經驗模型Behavior ModelingService DifferentiationPriorityQueueing Networkthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/51190/1/ntu-95-R93546017-1.pdf