1 / 25

Scalable Hierarchical Yield Control System For Semiconductor Manufacturing A Feasibility Study

Scalable Hierarchical Yield Control System For Semiconductor Manufacturing A Feasibility Study. Bill Martin, Jill Card, Wai Chan, Joyce Hyde, Yi-Min Lai IBEX Process Technology A Division of Neumath, Inc., Haverhill, MA John Doxsey, Paul Fearon National Semiconductor, S. Portland ME.

mercer
Télécharger la présentation

Scalable Hierarchical Yield Control System For Semiconductor Manufacturing A Feasibility Study

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Scalable Hierarchical Yield Control System For Semiconductor ManufacturingA Feasibility Study Bill Martin, Jill Card, Wai Chan, Joyce Hyde, Yi-Min Lai IBEX Process Technology A Division of Neumath, Inc., Haverhill, MA John Doxsey, Paul Fearon National Semiconductor, S. Portland ME

  2. Outline • Overview • Design Approach • Feasibility Study Data Collection • Feasibility Study Results • Conclusions/Next Steps

  3. Overview • Adaptive Hierarchical Design Works in conjunction with local tool controllers such as Neumath's Dynamic Neural Controller (DNC) product. • Optimises overall yield and end-of-line performance characteristics • Works with partial data to permit adaptive adjustment of downstream operation quality targets to minimize scrap

  4. Yield Controller Hierarchical Design Y I E L D CONTROLLER Optimize Metrology Targets across Products and Process Steps X X DNC Deposition DNC CMP DNC Photo DNC Etch Dep Quality Metrics Photo Quality Metrics Etch Quality Metrics CMP Quality Metrics

  5. Comprehensive Software Tool Development • Yield Control Layer • Accurate prediction of End Of Line (EOL) metrics • Automatic metrology target and spec adjustment • Tool Control Layer • Accurate prediction of post-process metrology and action advisory (DNCs, DNCe) • Automated recipe parameter target and spec adjustment

  6. Additional Benefits • Full use of all in-situ and ex-situ sensors across products. • Automatic optimization of recipe parameters and metrology specifications. No setup required. • Determination of quantified sensor importance • Pinpointing troubled tools • Quantified impact on EOL metrics and Yield • W2W Detection, Diagnosis, and Fix.

  7. Feasibility Study • Goal: demonstrate that accurate predictive models can be built • Using step-wise quality measurements as input • Predict end-of-line electrical parameters plus final product yield • Uses Neural Network-based predictive engine • Adaptive • Flexible • Accurate

  8. Feasibility Study Data Collection • National Semiconductor, South Portland, ME • 0.18μm CMOS technology • Covering 2 different product designs • Data from 381 wafers, 33 lots • Data Collected: • Quality measurements from CMP, Photo, Etch • All metal layers (31 operations in total) • 15 End of Line electrical parameters • Final Yield (% good die)

  9. End of Line Parameters Modelled • Metal 1,2,3 Bridging • N-Type Silicide Bridging • P-Type Silicide Bridging • Via 1,2,3 Contact Resistance • Metal 1,2,3,4 Continuity Resistance • Poly Continuity Resistance • Product Yield

  10. Measurements Used As Model Inputs • Metal layers • Photoresist top and bottom CD • Post-etch top and bottom CD • Defect density • Dielectric layers • Pre and post CMP thickness and non-uniformity • Pre and post CMP ILD thickness • Via Layers • Photoresist bottom CD (dense and isolated structures) • Post-etch bottom CD (dense and isolated structures)

  11. Data Preparation I • Quality metrics measured on a sample basis after each processing step. • Not always on the same wafers within the lot • Results is a sparse data set, insufficient for model training. • Algorithm to supply estimates for the missing measurements: • Use lot-based average if available. • Use time-based moving average otherwise.

  12. Data Preparation II • Merge quality metric data from all processing steps with the end-of-line electrical parameters and final yield • Using lot number and wafer identifier • Divide data into two subsets • Training data (70% of total) • Divided into Train and Test • Validation data (remaining 30%) • Never used during Neural Network training • Helps avoid over-fitting

  13. Neural Model Training • Use one Neural Network each for: • End of Line electrical parameters • Final Yield • Minimal Network acceptance criteria: • Root Mean Square Error (RMSE) must be less than standard deviation of the observed data • Better than “guessing the mean” • Model must generalize: • RMSE (validation data) < (1+α) RMSE (training data) • Acceptable accuracy measure.

  14. NeuMath Accuracy Measure • Each EOL parameter assigned a target and limits in accordance with product specification. • Divide limits range into 7 regions • Accuracy defined as the fraction of time the observed and predicted fall into the same sub-region. • Since accuracy is tied to the spec limits: • Remains consistent with current decision-making criteria • The fraction of time the decision would be the same using the prediction as it would be using the observed value.

  15. Results Overview • Models for 14 of the 16 End Of Line measurements converged : • Including final yield. • Provide an average accuracy of 90%!! • Comparison of RMS Errors of validation and training sets shows very good model generalization. • Critical for model-based decision-making.

  16. Predictive Model PerformanceMetal 1 Bridging

  17. Prediction Performance for Metal 1 Continuity Resistance

  18. Prediction Performance for Via 1 Contact Resistance

  19. Prediction Performance for Poly Continuity Resistance

  20. Prediction Performance for N-Type Silicide Bridging

  21. Prediction Performance for P-Type Silicide Bridging

  22. Predictive Model Performance for Product Yield

  23. Overall Yield

  24. Optimization Results Summary

  25. Conclusions / Next Steps • Feasibility study a success • Excellent Model performance (90% accuracy) • Optimization shows possible EOL improvements of up to 22% • Next Steps • Repeat on additional data sets from our partner and additional partners • Full beta trial early 2005 • Software product release Q2 2005

More Related