Proeedings of the 2012 Winter Simulation Conferene C. Laroque, J. Himmelspah, R. Pasupathy, O. Rose, and A.M. Uhrmaher, eds INDUSTRIAL IMPLEMENTATION OF A DYNAMIC SAMPLING ALGORITHM IN SEMICONDUCTOR MANUFACTURING: APPROACH AND CHALLENGES Justin Nduhura Munga 1,2 Stéphane Dauzère-Pérès 1 Philippe Vialletelle 2 Claude Yugma 1 1 Department of Manufaturing Sienes and Logistis Eole Nationale des Mines de St Etienne CMP F-13541 Gardanne, Frane 2 STMiroeletronis Centre Commun de Miroéletronique de Crolles F-38926 Crolles, Frane ABSTRACT In a worldwide environment, sustaining high yield with a minimum number of quality ontrols is key for manufaturing plants to remain ompetitive. In high-mix semiondutor plants, where more than 200 produts are onurrently run, the omplexity of designing effiient ontrol plans omes from the larger amount of data and number of prodution parameters to handle. Several sampling algorithms were proposed in the literature, but most of them are seen impratiable when oming to an industrial implementation. In this paper, we present and disuss the industrial implementation of a dynami sampling algorithm in a high-mix semiondutor plant. We desribe how the sampling algorithm has been modified, and point out the set of questions that have been raised by the industrial program. Results indiate that more than 30% of ontrol operations on lots ould be avoided without inreasing the material at risk in prodution. 1 INTRODUCTION One of the strategies identified by semiondutor manufaturers to improve ompetiveness is to better manage ontrols throughout prodution (Mouli and Sott 2007). As ontrolling 100-perent is neither feasible nor interesting, dynamially identifying the right lots to ontrol is ritial to redue osts without impating produt quality or inreasing risks. Several sampling algorithms are proposed in the literature (Purdy 2007; Sun and Johnson 2008). However, the speifiity of eah semiondutor manufaturer is suh that most dynami sampling algorithms proposed are too omplex to industrialize. The IT infrastruture, the number of data to handle, the return on investment, or the ompany ulture are some of fators that often lead to impratiability of many sampling algorithms. In this paper, we present the industrial implementation of the dynami sampling algorithm proposed by (Dauzère-Pérès et al. 2010) within the 300mm site of STMiroeletronis in Crolles, Frane. We first desribe the sampling algorithm and finanial metris we developed to assess the potential gains. Then, we explain how the sampling algorithm has onretely been implemented using a mehanism introdued in (Nduhura Munga et al. 2011) for handling a large amount of data in short CPU times. At last, we disuss the various questions that have been raised by the industrialization program. We fous on defetivity 978-1-4673-4782-2/12/$31.00 2012 IEEE 2151
ontrols and espeially on the redution of the material at risk, i.e. the potential loss in ase a problem ours. Finanial metris indiate a potential gain up to $1,000,000, and the industrial deployment showed that more than 30% of ontrol operations on lots ould be avoided without inreasing the material at risk in prodution. The paper is strutured as follows. In setion 2, we desribe ontrols in semiondutor manufaturing, and larify the position of our work. Setion 3 presents the problem we address. Setion 4 and Setion 5 are devoted to the desription of the dynami sampling algorithm and its industrial implementation respetively. In setion 6, we disuss the questions that have been raised when industrializing the approah. Setion 7 onludes the paper and gives diretions for further researh. 2 CONTROLS IN SEMICONDUCTOR MANUFACTURING In semiondutor manufaturing, ontrols are neessary evils beause of the prohibitive amount of time required to manufature a funtional IC (Boussetta and Cross 2005). Different levels of ontrols exist depending on the point of view adopted (Bassetto and Siadat 2009). For eah level of ontrol, one or several types of ontrols an be defined. More generally, six main levels of ontrols an be defined (Wright 2001): Failities or tehnial installations. The goal is to ensure the best possible environment for the fabriation of wafers: Clean room ambient harateristis (temperature, humidity, pressure, et.), fluids, liquids, gases, energy (load, intensity, voltage), et. Equipment sensors. To ensure effiient proessing operations, all variations have to be deteted and analyzed. For that, several types of sensors need to be plaed on different prodution tools to trigger alarms and ations in manufaturing systems. Fab or In-Line measurements. This level of ontrol groups measurements done on silion wafers with a large variety of tehniques: Ellipsometry, refletivity, sanning eletron mirosopy, visual inspetion, pixel to pixel omparison, resistivity, satterometry, et. The goal is to monitor both the proess and the tool drifts. Parametri testing addresses basi parameters of eletrial devies: Transistor voltage thresholds, leakage urrent, oxide breakdown voltage, et. Final or funtional tests aim at verifying that the semiondutor devies funtion properly. Physial haraterization and wafer level reliability evaluate omponent life-time under various stressing onditions (humidity, temperature, orrosion, et). Among these six levels of ontrols, we fous on Fab or In-Line measurements, and espeially on defetivity ontrols that aim at deteting defets on wafers. Different types of defets exist (orrosion, partiles, srathes, voids, extra-patterns, et.) and they are mainly generated by prodution tools (Pepper et al. 2005). All prodution tools are onerned, sine even the smallest tool has mehanial parts moving, and thus, a potential generation of defets on wafers. The omplexity is therefore in the number of tools to ontrol with a minimum number of ontrol operations on lots. Additional fators suh as the depth of ontrol (one ontrol operation may validate several tools) (Nduhura Munga et al. 2011) (Rodriguez-Verjan et al. 2011), the apture rate (all defets annot be deteted at any prodution stage) (Shantikumar 2007), the kill ratio (all defets do not have the same impats on wafers), the produt ritiality (some produts are more ritial than others), the ustomer requirements, et. ontribute in inreasing the omplexity of designing effiient defetivity ontrol plans. 3 PROBLEM DESCRIPTION A ontrol without real added value is waste of time and money. This is what has been highlighted within the framework of our works related to defetivity ontrols. Figure 1 shows an example that illustrates the drawbaks of stati sampling for defetivity ontrols. Lots L2, L4, and L6 are flagged at the start of the prodution for defetivity ontrols after proessing operations. The sampling plan is to ontrol one lot every two lots, i.e. 50% of lots. However, the variability and fatory dynamis are suh that TOOL1 may 2152
proess all flagged lots i.e. L2, L4, L6 whereas TOOL2 does not proess any. Consequenes are an overontrol for TOOL1 and lak-of-ontrol for TOOL2. Figure 1: Drawbaks of Stati Sampling To solve this problem, lots should be dynamially seleted in front of defetivity steps depending on the history of eah lot. In the ase illustrated in Figure 2, a dynami sampling will lead to seleting, for example, lots L2, L3, and L6 to ensure a sampling rate of 50%. Figure 2: Dynami Sampling The hallenge is in the dynami seletion of the best lots to ontrol. A ontrol operation on a lot (defetivity ontrol) may over several proessing tools leading to a signifiant number of parameters to analyze. However, all tools and proessing operations are not equivalent i.e. the ritiality or priority vary depending on the tool, the proessing operation, or the produt type. Dynamially seleting the best lot to ontrol implies the ability to ompute and analyze in real-time all the information linked to both the lot history and prodution onstraints. (Dauzère-Pérès et al. 2010) proposed an algorithm based on a formula that helps to give a weight to eah lot arriving in front of inspetion steps. The smaller the weight assoiated to a set of lots, the larger the priority of the lots in the set on inspetion tools. This weight is alled Global Sampling Indiator (GSI) and it is based on both the depth of ontrol and tools ritiality. The next setion presents this GSI, its evaluation through simulations, and, finanial metris we developed to assess the potential gains before the industrial implementation. 4 DYNAMIC SAMPLING: GLOBAL SAMPLING INDICATOR (GSI) This setion first realls the approah for dynamially sampling lots, and then explains the finanial metris that are used to evaluate the impat of a sampling strategy. For more information on the approah, see (Dauzère-Pérès et al. 2010). 2153
4.1 GSI Sampling Algorithm The dynami sampling algorithm proposed by (Dauzère-Pérès et al. 2010) is based on a sore or weight alled GSI for Global Sampling Indiator. Eah time a lot l is available for inspetion and a set S of lots is already waiting to be inspeted, a GSI sore (and other parameters linked to the prodution onstraints) is used to deide whether to inlude the lot l in the set S and skip another lot l in S, or to skip the available lot l. The GSI sore is linked to sets of lots. The aim is to define a weight that gives information on both the added value of an available lot l in term of risk redution, and the added value of the available lot l within the set S of lots waiting to be inspeted. For that, the sampling algorithm ompares the GSI sore of different sets obtained by removing l and adding l, i.e. S l ' l l' S, and the original set of lots without inluding the available lot l, i.e. S. The set with the minimum GSI sore is seleted, hene the deision of sampling or not the available lot l. The GSI sore is given by: Where: GSI( S) R r1 NRV r ( S) ILr (1/ ) NRV r ( S) ILr R = Number of risk types. NRV(S) = New risk value if lots in set S are inspeted. IL r = Inhibit Limit for risk type r, i.e. the maximum number of wafers that an be run between two inspetions for risk r. α and β (α 1 and β 1) are parameters that are used to put more (for α) or less(for β) emphasis on getting as far as possible from IL r. Before being industrialized, the GSI sampling algorithm had to be simulated, validated, and potential gains assessed. For that, a sampling simulator has been developed (Yugma et al. 2011) and Table 1 presents some of the statistis (provided by the simulator) that have been used to quantify the potential gains. Simulations on historial data were run over a period of three months, and the results are provided for the GSI sampling algorithm and atual fab sampling. In our experiments, the risk value is the W@R (Wafer at Risk), whih is the number of wafers proessed on prodution tools between two defetivity ontrol operations, i.e. the material at risk. Reduing this number without inreasing the number of ontrols has a diret impat on the insurane osts for the ompany, hene the gain in the produt osts. Table 1 shows the number of sampled lots, measured lots and skipped lots, together with number of wafers above the Inhibit Limits, the average medium W@R and average maximum W@R for the prodution tools. Table 1: Evaluating the GSI sampling algorithm. GSI sampling Fab Sampling Number of sampled lots 3.8*A A Number of measured lots 0.9*A A Number of skipped lots 2.9*A 0 Wafers above IL 7,759,743 9,517,277 Medium W@R (Average) 0.28*B B Maximum W@R (Average) 0.39*C C Note that, with a redued number of measurements (0.9*A) ompared to Fab sampling, the GSI sampling algorithm redues the average medium W@R and average maximum W@R by 72% and 61% respetively. Among these six statistis, the number of wafers proessed on prodution tools above the Inhibit Limits has been seen to be most pratial for the assessment of finanial gains as desribed in the next setion. (1) 2154
4.2 Evaluating the GSI Sampling Algorithm: Finanial Metris To evaluate the gains of the GSI sampling algorithm and therefore quantify the return on investment, we defined three different metris. Among these three metris, one has been seen to be more pratial for our works related to defetivity ontrols. However, depending on the fab or the type of the risk addressed, one metri may be more suitable than another. The three metris are: 1. The number of metrology tools that an be saved by using the GSI sampling algorithm. The idea is to ompute the number of metrology tools required with the GSI sampling algorithm to obtain performane indiators that are as good as in urrent Fab sampling. However, as the goal is to not to redue the number of measured lots but to redue the number of measurements without added value, this first metri is not the most suitable. 2. The downtime osts inurred when Inhibit Limits are exeeded. The idea is to onsider that a finanial ost is inurred when a prodution tool is stopped beause an Inhibit Limit is exeeded. The resulting downtime is a non produtive time whih osts money. We assume that the prodution tool stays down until the W@R beomes smaller than its orresponding Inhibit Limit. However, the problem of this seond metri is that, in pratie, a prodution tool will not always be stopped when its Inhibit Limit is exeeded. Therefore, this seond metri is not the most relevant. 3. The risk related osts inurred when Inhibit Limits are exeeded. The idea is to onsider that the risk of losing wafers inreases when the W@R of a prodution tool is above its Inhibit Limit. Based on a probability of failure, it is possible to alulate how muh money ould be lost. This last metri seems to be the most pratial and was used to quantify the potential gains. Using results reported in Table 1, the potential gains were omputed using the following formula: Gain ( 9,517,277 7,759,743)* P * P. where P LOSS represents the probability of losing wafers when a prodution tool is above its Inhibit Limit, and P $ is the average wafer ost. P LOSS is linked to both the prodution environment and simulation period. Its value may thus be disussed depending on the proessing operations and produt types. In our works related to defetivity ontrols in the 300mm fab of STMiroeletronis, we onsidered a probability of 1/2000 and a wafer ost of $1,500, whih gives: LOSS Gain ( 9,517,277 7,759,743)*(1/ 2000)*1500 $1,318,150. This signifiant potential gain, in addition to the performane indiators introdued in (Yugma et al. 2011), has been one the main motivations that led to the industrialization of the GSI sampling algorithm. 5 INDUSTRIAL IMPLEMENTATION A survey of the literature (Nduhura Munga et al. 2012) showed that several sampling algorithms were impratiable for industrial implementation. This was also one of the main hallenges for the GSI sampling algorithm. If simulations showed that signifiant gains ould be ahieved, additional types of evaluations had to be performed regarding the IT infrastruture, the data management and availability, the training of operators, the ompany ulture, or the management of resoures. The risk was to see potential savings ompletely eradiated by additional types of investments (IT, resoures, trainings, et.). Another hallenge that had to be faed was the type of risk to be mastered. Indeed, in the first evaluation of the GSI sampling algorithm, only the risk at the tool level (W@R, Wafer at Risk) was evaluated. However, in a high-mix environment as in the 300mm site of STMiroeletronis in Crolles, Frane, several types of risks related to defetivity ontrols need to be mastered: the number of wafers proessed in the same hamber between two ontrols, the number of wafers at the same operation, with the same tehnology, reipe, resin, et. The dynami seletion of lots should therefore ensure an optimal overage of all of the types of risk. For that, the IPC (for the Frenh Indie Permanent par Contexte, whih means Permanent Index per Context ) mehanism proposed in (Nduhura Munga et al. 2011) was industrialized to support the industrial implementation of the GSI sampling algorithm. The IPC is a ounter linked to $ 2155
several types of ontexts (reipe, resin, tehnology, et.). It has been introdued to simplify the risk omputations. Eah time a lot is proessed on a prodution tool and verifies a given ontext, an IPC ( IPC ) is attahed to the lot for the onsidered ontext as desribed in (Nduhura Munga et al. 2011). Using this IPC information, the GSI formula for a set S of lots was adapted as follows: where: GSI( S) N 1 Min ls (1/ ) IPC IPC Min IPC IPC l IL N : Number of ontexts, IPC : IPC of lot l for ontext, l IPC LLM ) LLM ( ) ls l IL LLM ( ) ( : IPC of the lot Last Lot Measured (LLM) for ontext, i.e. the last lot that has redued the risk for ontext, IL : Inhibit Limit for ontext, α and β (α 1 and β 1) are parameters that are used to put more (for α) or less(for β) emphasis on getting as far as possible from the Inhibit Limit. Using the IPC information, the industrial implementation of the GSI sampling algorithm was onsidered as the only way to optimize the use of metrology tools by dynamially seleting the best lots to measure. Eah ontext (tool, hamber, reipe, resin, tehnology, et.) is taken into aount in the GSI formula, and the dynami seletion of lots aims at minimizing all the types of risks (ontexts) in prodution. The industrialization phase has been divided into three parts devoted to the sampling, skipping and sheduling of lots on metrology tools. The first part of the projet foused on the skipping of lots in metrology and a first evaluation showed that more than 30% of ontrol operations on lots ould be released without inreasing the risk in prodution. However, several questions have been raised by this transition from stati to dynami sampling. In the next setion, we present some of these questions that onstitute new diretions for further researh. 6 DISCUSSIONS AND PERSPECTIVES The industrial implementation of the GSI sampling algorithm showed that a signifiant number of ontrol operations on lots ould be released without inreasing the risk in prodution, hene a better management of metrology tools. However, modifying the sampling poliy, i.e. replaing stati sampling by dynami sampling, hanged the way engineers were working, leading to new problems that need to be solved. Here are some these new problems that are perspetives for further work: 1. Defetivity tool qualifiations. Using start or stati sampling, lots were flagged at the start of the prodution, i.e. the perentage of lots to be measured was defined at the start of prodution. Therefore, information on the produts of the lots to be measured was known, and the qualifiations of defetivity tools was taken into aount to balane the measurement workload on the different types of defetivity tools. By dynamially seleting lots in front of metrology steps, there is no prior information on the set of lots to be measured. The onsequene is that some defetivity tools may not be used enough if the right lots are not seleted, and other defetivity tools may be too muh loaded, leading to long yle times for the lots waiting to be measured. 2. Exursion management, i.e. when a proess or tool falls out of speifiation. Using stati sampling, lots were flagged for regular measurements at some predefined steps. This means that, eah time a problem was deteted on a lot, it was easy to quikly identify the soure of the problem by quantifying the added defets of eah proessing step. Using a dynami sampling approah, the omplexity of isolating the soure of defets is inreased, sine lots are seleted based l (2) 2156
on the information they bring and not expliitly at all the metrology steps. An additional analysis must be performed before isolating the soure of the exursion knowing that no prodution tool is stopped before this soure is identified. These problems show the types of omplexity that need to be faed when trying to deploy a new strategy or approah within a high-mix semiondutor plant. As different types of organizations are alled to work and ollaborate (qualifiations, sampling, sheduling, exursion management, prodution, et.), eah time a novel solution is proposed, the impat on the other ativity types need to be learly understood and assessed. Although this may vary from one ompany to another, signifiant efforts are required to onvine and highlight the added value of the new solution. This is what has been experiened before industrially implementing the GSI sampling algorithm, explaining why many sampling algorithm proposed in the literature were impratiable. 7 CONCLUSION In this paper, we presented and disussed the industrial implementation of a dynami sampling algorithm within the 300mm site of STMiroeletronis in Crolles, Frane. We foused on defetivity ontrols and espeially on the redution of the material at risk, i.e. the potential loss in ase a problems ours. The dynami sampling algorithm is based on a Global Sampling Indiator that evaluates the impat of measuring a set of lots and not only one lot. It has been industrialized using a Permanent Index per Context mehanism that help handling a very large amount of data in short CPU times. Results show a quik return on investment and a signifiant redution of the number of ontrol operations on lots without added value. The industrial program has highlighted new problems that are interesting perspetives for further researh. These new problems onern defetivity tool qualifiations, the design of ontrol plans and exursion management. ACKNOWLEDGMENTS This work has been done within the framework of a joint ollaboration between STMiroeletronis in Crolles, Frane, and the Center for Miroeletronis in Provene of the Eole des Mines de Saint-Etienne in Gardanne, Frane. It has also been written as a part of the European projet IMPROVE (Implementing Manufaturing siene solutions to inrease equipment produtivity anf fab performane). REFERENCES Bassetto, S., and A. Siadat. 2009. Operational Methods for Improving Manufaturing Control Plans: Case Study in a Semiondutor Manufaturing. Journal of Intelligent Manufaturing 20: 55 65. Boussetta, A., and A. J. Cross. 2005. Adaptive Sampling Methodology for In-Line Defet Inspetion. In Proeedings of the 2005 IEEE/SEMI Advaned Semiondutor Manufaturing Conferene and Workshop, Munih, Germany, 25 31. Dauzère-Pérès, S., J.-L. Rouveyrol, C. Yugma, and P. Vialletelle. 2010. A Smart Sampling Algorithm to Minimize Risk Dynamially. In Proeedings of the 2010 IEEE/SEMI Advaned Semiondutor Manufaturing Conferene, San Franiso, California, USA, 307 310. Mouli, C., and M. J. Sott. 2007. Adaptive Metrology Sampling Tehniques Enabling Higher Preision in Variability Detetion and Control. In Proeedings of the 2007 IEEE/SEMI Advaned Semiondutor Manufaturing Conferene, Stresa, Italy, 12 17. Nduhura Munga, J., S. Dauzère-Pérès, P. Vialletelle, and C. Yugma. 2011. Dynami Management of Controls in Semiondutor Manufaturing. In Proeedings of the 2011 IEEE/SEMI Advaned Semiondutor Manufaturing Conferene, Saratoga Springs, New York, USA, 18 23. 2157
Nduhura Munga, J., G. Rodriguez-Verjan, S. Dauzère-Pérès, C. Yugma, P. Vialletelle, and J. Pinaton. 2012. A Literature Review on Sampling Tehniques in Semiondutor Manufaturing. Working Paper, submitted. Pepper, D., O. Moreau, and G. Hennion. 2005. Inline Automated Defet Classifiation: a Novel Approah to Defet Management. In Proeedings of the 2005 IEEE/SEMI Advaned Semiondutor Manufaturing Conferene and Workshop, Munih, Germany, 43 48. Purdy, M. 2007. Dynami, Weight-Based Sampling Algorithm. In Proeedings of the 2007 IEEE International Symposium on Semiondutor Manufaturing, Santa Clara, USA, 1 4. Rodriguez-Verjan, G., S. Dauzère-Pérès, and J. Pinaton. 2011. Impat of Control Plan Design on Tool Risk Management: A Simulation Study in Semiondutor Manufaturing. MASM 2011 (7 th International Conferene on Modeling and Analysis of Semiondutor Manufaturing), In Proeedings of the 2011 Winter Simulation Conferene, Edited by S. Jain, R. R. Creasey, J. Himmelspah, K. P. White, and M. Fu, 1918 1925. Pisataway, New Jersey: Institute of Eletrial and Eletronis Engineers, In. Shantikumar, J. G. 2007. Effets of Capture Rate and Its Repeatability on Optimal Sampling Requirements in Semiondutor Manufaturing. In Proeedings of the 2007 IEEE International Symposium on Semiondutor Manufaturing, Santa Clara, USA, 1 6. Sun, S., and K. Johnson. 2008. Method and System for Determining Optimal Wafer Sampling in Real- Time Inline Monitoring and Experimental Design. In Proeedings of the 2008 IEEE International Symposium on Semiondutor Manufaturing, Tokyo, Japan, 44 47. Wright, P. K. 2001. 21 st Century Manufaturing. 1 st ed., Prentie-Hall, In. Yugma, C., S. Dauzère-Pérès, J.-L. Rouveyrol, and P. Vialletelle. 2011. A Smart Sampling Sheduling and Skipping Simulator and its Evaluation on Real Data Sets. MASM 2011 (7 th International Conferene on Modeling and Analysis of Semiondutor Manufaturing), In Proeedings of the 2011 Winter Simulation Conferene, Edited by S. Jain, R. R. Creasey, J. Himmelspah, K. P. White, and M. Fu, 1908 1917. Pisataway, New Jersey: Institute of Eletrial and Eletronis Engineers, In. AUTHOR BIOGRAPHIES JUSTIN NDUHURA MUNGA reeived his Engineering degree in Computer siene, Miroeletronis, and Automation from the Eole Polytehnique de Lille in Lille, Frane in 2009. He also reeived a Master of Siene in Miroeletronis from the University of Lille, Frane in 2009. Currently he is a Ph.D. student in Industrial Engineering at the Eole des Mines de Saint-Etienne in Gardanne, Frane, and works at STMiroeletronis in Crolles, Frane. His works mostly fous on implementing dynami ontrols in high-mix semiondutor plants. He reeived the best student paper award from the IEEE/SEMI Advaned Semiondutor Manufaturing Conferene, New-York, USA, 2011. His email address is justin.nduhuramunga@st.om. STEPHANE DAUZÈRE-PERÈS is Professor at the Provene Miroeletronis Center of the Eole des Mines de Saint-Etienne, where he is heading the Manufaturing Sienes and Logistis Department. He reeived the Ph.D. degree from the Paul Sabatier University in Toulouse, Frane, in 1992; and his Habilitation à Diriger des Reherhes from the Pierre and Marie Curie University, Paris, Frane, in 1998. He was a PostDo Fellow at the Massahusetts Institute of Tehnology, U.S.A., in 1992 and 1993, and Researh Sientist at Erasmus University Rotterdam, The Netherlands, in 1994. He has been Assoiate Professor and Professor from 1994 to 2004 at the Eole des Mines de Nantes in Frane. He was invited Professor at the Norwegian Shool of Eonomis and Business Administration, Bergen, Norway, in 1999. Sine Marh 2004, he is Professor at the Eole des Mines de Saint-Etienne. His researh mostly fouses on optimization in prodution and logistis, with appliations in planning, sheduling, distribution and transportation. He has published more than 40 papers in international journals and 100 ommuniations in onferenes. His email address is dauzere-peres@emse.fr. 2158
CLAUDE YUGMA is Assoiate Professor at EMSE. He earned a PhD degree in Computer Siene and Combinatorial Optimization at the Grenoble Institute of Tehnology, Frane. His researh was initially foused on sheduling problems and onsisteny of global and loal sheduling deisions in semiondutor manufaturing. He is now also working on the interations between Advaned Proess Control and sheduling and dispathing deisions in semiondutor manufaturing, with topis suh as: dynami sampling, preventive maintenane sheduling, et. He was involved in the in the regional projet Rousset 2003-2008 (with STMiroeletronis) and in the MEDEA+ European projet HYMNE. Currently, he is involved in the ENIAC European Projet IMPROVE. His email address is yugma@emse.fr. PHILIPPE VIALLETELLE is manager of the Operations and Methods System group at STMiroeletronis. After reeiving an Engineering degree in Physis, he entered the semiondutor industry working on ESD and physial haraterization. His next experienes were Metrology and Proess Control where he drove the deployment of methodologies and tools for a 200mm fab. He finally integrated Industrial Engineering and is now responsible for the development of advaned programs for the management of Crolles 300mm prodution line. At European level, Philippe is in harge of the definition and follow-up of ollaborative programs in the field of Manufaturing Sienes suh as HYMNE or IMPROVE. His email address is philippe.vialletelle@st.om. 2159