Workshop - Statistical methods applied in microelectronics 13. June 2011, Catholic University of Milan, Milan, Italy Approaches for Implementation of Virtual Metrology and Predictive Maintenance into Existing Fab Systems G. Roeder 1), M. Schellenberger 1), U. Schoepka 1), M. Pfeffer 1), S. Winzer 2), S. Jank 2), D. Gleispach 3), G. Hayderer 3), L. Pfitzner 1) 1) Fraunhofer Institute for Integrated Systems and Device Technology (IISB), Schottkystraße 10, 91058 Erlangen, Germany 2) Infineon Technologies Dresden GmbH, Königsbrücker Straße 180, 01099 Dresden, Germany 3) austriamicrosystems AG, Tobelbader Straße 30, 8141 Unterpremstaetten, Austria
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM VM and PdM application examples Structured approach for VM and PdM development Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 2
Motivation Objective European project IMPROVE : IMPROVE European semiconductor fab competitiveness and efficiency processes reproducibility and quality efficiency of production equipment shorten cycle times Approach: Development of novel methods and algorithms for virtual metrology (VM) and predictive maintenance (PdM) Challenges: Implementation of new control paradigms in existing fab systems Reusability of developed solutions amongst the nine IC manufacturers fabs gathered in IMPROVE 3
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM VM and PdM application examples Structured approach for VM and PdM development Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 4
VM objectives and benefits VM objectives Predict post process physical and electrical quality parameters of wafers and/or devices from information collected from the manufacturing tools including support from other available information sources in the fab VM benefits Support or replacement of stand-alone and in-line metrology operations Support of FDC, run-to-run control, and PdM Improved understanding of unit processes Place of execution of VM in a process flow 5
VM key requirements Key requirements of a VM system Capability for estimation of the equipment state or wafer quality parameter within predefined reaction time, typically at wafer-to-wafer level Inclusion of metrology to control and adjust VM prediction and models Capability for integration into a fab infrastructure and interaction with other APC modules, e.g. run-to-run control VM module components 6
Concept of PdM for IC-manufacturing Current situation of scheduled maintenance in semiconductor manufacturing Maintenance scheduled based on elapsed time or fixed unit count usage Maintenance frequency depends on: process engineer s experience known wear out cycles of certain parts of the tool PdM considerations based on worst case scenarios to avoid unscheduled downs Ideal maintenance strategy - Run to almost fail Predictive maintenance aims at replacing/repairing an equipment part when it has nearly reached its end of life PdM workflow utilizing Bayesian Networks 7
PdM objectives, benefits and key requirements PdM objectives Predict upcoming equipment failures or events, their root causes and corresponding maintenance tasks in advance PdM benefits Improved uptime and availability - by reducing or eliminating unplanned failures Reduced operational cost by enhanced consumable lifetimes and efficiency of service personnel Improved product quality by eliminating degraded operation and tightening process windows Reduced scrap by maintenance actions before a failure occurs Key requirements of a PdM system Capability for reliable prediction of upcoming equipment failures, root causes and corresponding maintenance tasks Capability for integration into a fab infrastructure 8
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM Structured approach for VM and PdM development VM and PdM application examples Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 9
Concept for a generic VM and PdM implementation Definition of VM and PdM as EE applications on a conceptual level Abstraction from existing fab infrastructures applying UML as project standard Adoption of architectures following SEMI and SEMATECH, including existing SEMI standards (interface A, B) Consideration of user requirements for fab-wide master framework Develop component- and service-based models for VM and PdM 10
Architecture for generic VM and PdM implementation UML description of the EE system and of a generic VM/PdM module With contributions from the University of Augsburg Mapping to existing infrastructures Consideration of specific user infrastructure Inclusion of configuration, data analysis, and filter modules as plug-ins First framework realization available 11
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM VM and PdM application examples Structured approach for VM and PdM development Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 12
Structured approach for VM and PdM development Phases in VM and PdM development as adapted from the Cross-Industry Standard Process for Data-Mining (CRISP-DM) 13
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM VM and PdM application examples Structured approach for VM and PdM development Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 14
Introduction to the etch process Trench etch process The IT etch defines the active regions The process is carried out in four steps: 1. Etching of the organic ARC and nitride layer (mask open) 2. Conditioning step 3. Conditioning step 4. Etching of the poly silicon (IT etch) status first process step Strip of resist and of anti-reflective coating (ARC) by etching in a plasma Steps 4 and step 1 are expected to primarily define the etched depth status fourth process step final process result 15
Data understanding and preparation Data analysis and understanding Step and summary data collected over three months for two slightly different etch recipes performed on four chambers Data reduction step Derivation of 8 subsets of data with 130 predictor/target Predictor selection by rule based elimination of correlated variables Prioritization of data sources, e.g. logistic, equipment, and sensor data 16
Modeling approach - overview Modeling approaches Criteria for model selection Stepwise linear regressionalgorithm Identification of a small set of predictor variables Inclusion of model on FDC system Time-series neural network Model development for variables selected in stepwise linear regression Test of models on new data collected on the FDC system 17
Predictor selection using stepwise linear regression Modeling Predictor selection and model development using bagging, and repeated stepwise linear regression Result Prediction of etch-depth is possible Predictor selection is unique and independent from selection of sample subset Prediction capability for sub-sets identical as for training on complete data set Prediction of etch depth by stepwise linear regression 18
Prediction capability of different models Prediction capability of stepwise regression and time-series neural network Parameter Stepwise regression TSNN Std. dev. abs. 4.0 nm 3.8 nm Std. dev. rel. 0.8 % 0.75% MAE 3.2 nm 3.0 nm Prediction capability of TSNN slightly better than for stepwise regression Modeling techniques provide comparable results 19
Assessment of model adaptation Capability of prediction after model adaptation on additional FDC test data Regression model TSNN model Due to modifications in database, errors occur in VM for test set Prediction errors are lower for TSNN Capability for model adaptation can be tested: Comparable adaptation for both models (MAE: 12 nm); models rebuilding from full predictor set necessary 20
Outline Motivation Virtual metrology (VM) and predictive maintenance (PdM) Concept of VM and PdM for IC-manufacturing Framework for implementation of VM and PdM Structured approach for VM and PdM development VM and PdM application examples Prediction of etch depth by VM PdM for prediction of filament break-down in ion implantation Conclusions and outlook 21
Ion implantation overview Implantation process Different ions (B, BF 2, P, As, Sb, ) for doping of certain chip regions Ion source: 1. Electron generation from heated cathode 2. Creation of ions in process gas through collisions with accelerated electrons 3. Extraction and acceleration of ion beam Degeneration of cathode/heating filament through sputtering End of lifetime: breakdown of filament Problem: measurement of filament degradation not possible => PdM! source: http://en.wikipedia.org/wiki/ion_implantation 22
PdM modeling Predictor parameters Different currents and voltages related to ion source, power Gas flow rates, source pressure, time Modeling Method: Bayesian Networks with soft discretization (discretization required for non-gaussian data) Model learning with real production data (noisy, different recipes/ions) Soft discretization used for broadening of data basis and reduction of quantization error p(nextbreakdown=state1 Fil-I,Ext-I,Arc-I,GAS) 23
PdM Results Data 7 maintenance cycles for training 2 maintenance cycles for test Simple model with 4 predictors Prognosis Calculation of probability not to fail within the next 50h based on actual data Can be directly used as filament health factor Further investigations Improved pre-processing for more robust prediction (considering influence of recipe changes for outlier prevention) 24
Conclusions and outlook Achievements Common architecture to integrate VM and PdM into the different existing fab systems developed Software for implementation of VM and PdM modules in fab environments available VM and PdM modules for important fabrication steps demonstrated Development may follow a structured approach Data quality and preparation is of key importance Prediction quality but also other properties (e.g. model adaption, automation) are key to model selection Next steps in IMPROVE Refinement of VM and PdM algorithms Continued testing of VM algorithm for etch-depth prediction 25
Acknowledgment This research is funded by the German Federal Ministry of Education and Research (BMBF) and the European Nanoelectronics Initiative Advisory Council (ENIAC) The work is carried out in the ENIAC project IMPROVE (Implementing Manufacturing science solutions to increase equipment PROductiVity and fab performance) 26