Virtual overlay metrology for fault detection supported with integrated metrology and machine learning

Transcription

1 Virtual overlay metrology for fault detection supported with integrated metrology and machine learning Emil Schmitt-Weaver MATLAB Expo 2016 Benelux June 28 th

2 Slide 2

3 Slide 3

4 Outline Introduction Slide 4 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

5 How the function works Function input comes from TWINSCAN metrology & context; Slide 5 Input Output Wafer Alignment metrology for all colors (NIR, FIR, red, green) Residuals with respect to color & model used Wafer quality Wafer Leveling metrology TWINSCAN Context Chuck number Field position Target position Function f : 3 inputs 1 output Predicted (F2N) x: 9.0 nm y: 7.4 nm

6 Outline Introduction Slide 6 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

7 Slide 7

8 Slide 8

9 Slide 9

10 Outline Introduction Slide 10 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

11 Slide 11

12 Slide 12

13 Outline Introduction Slide 13 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

14 Slide 14

15 Slide 15

16 Slide 16

17 Slide 17 Slide 17

18 Outline Introduction Slide 18 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

19 Precision of Trained Function as a vector map Noise between the measured and predicted overlay is relatively consistent for both Training and Testing groups Training Group = Avg Measure-Predcited Error Training - Testing Group Testing Group = Avg Measure-Predcited Error Consider the error as a plus or minus contribution per wafer coordinate position of any prediction from the trained function Slide 19 Training Group = Avg Measure-Predcited Error Testing Group = Avg Measure-Predcited Error a b c Training - Testing Group 1 nm x: x: nm nm y: y: 1.6 nm 1 nm 1 nm x: 2.4 nm y: 1.9 nm 1 nm x: 2.4 nm y: 1.9 nm 1 nm x: x: nm y: y: 1.4 nm Point - Point delta between a) and b)

20 Outline Introduction Slide 20 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

21 Results Measured Data SK hynix provided the on product overlay data for our proof book analysis. Process for the 20nm DRAM layer was intentionally manipulated as it was prepared for high volume production by the integration team Slide 21 Black lines denote start and end to lot Green lines denote wafers testing group measured m3s (nm) measured m3s (nm) Lot 1 Lot 1 Lot 2 Lot 2 Lot 3 Lot 3 Lot 4 Lot 4 Lot 5 Lot 5 Lot 6 Lot 6 Lot 7 Lot 7 Lot 8 Lot 8 Lot 9 Lot 9 Lot 10 Lot 11 Lot 12 Lot 13 Lot 14 Lot 15 Lot 16 Lot 17 Lot 18 Lot 19 Lot 20 Lot 21 Lot 22 Lot 23 Lot 24 Lot 25 Lot 26 Lot 27 Lot 28 Lot 29 Lot 30 Lot 31 Lot 32 Lot 33 Lot 34 Lot 35 Lot 36 Lot 37 Lot 38 Lot 39 Lot 40 Lot 41 Lot 42 Lot 43 Lot 44 Lot 45 Lot 46 Lot 47 Lot Lot 10 Lot 11 Lot 12 Lot 13 Lot 14 Lot 15 Lot 16 Lot 17 Lot 18 Lot 19 Lot 20 Lot 21 Lot 22 Lot 23 Lot 24 Lot 25 Lot 26 Lot 27 Lot 28 Lot 29 Lot 30 Lot 31 Lot 32 Overlay X m3s (nm) Measured per Wafer Lot 33 Overlay Y m3s (nm) Measured per Wafer Lot 34 Lot 35 Lot 36 Lot 37 Lot 38 Lot 39 Lot 40 Lot 41 Lot 42 Lot 43 Lot 44 Lot 45 Lot 46 Lot 47 Lot 48 Lot 49 Lot 50 Lot 51 Lot 52 Lot 53 Lot 54 Lot Lot Lot 59 Lot 60 Lot 61 Lot 62 Lot 63 Lot 64 Lot 65 Lot 66 Lot 67 Lot 68 Lot 69 Lot 70 Lot 71 Lot 72 Lot 73 Lot 74 Lot 75 Lot 76 Lot 77 Lot 78 Lot 79 Lot 80 Lot 81 Lot 82 Lot 83 Lot 84 Lot 85 Lot 86 Ck1 7= 9.61 nm Ck2 7= 9.26 nm Lot 49 Lot 50 Lot 51 Wafer Sequence Overlay Y m3s (nm) Prediction per Wafer Lot 52 Lot 53 Lot 54 Lot Lot Lot 59 Lot 60 Lot 61 Lot 62 Lot 63 Lot 64 Lot 65 Lot 66 Lot 67 Lot 68 Lot 69 Lot 70 Lot 71 Lot 72 Ck1 7= 6.60 nm Ck2 7= 5.78 nm Lot 73 Lot 74 Lot 75 Lot 76 Lot 77 Lot 78 Lot 79 Lot 80 Lot 81 Lot 82 Lot 83 Lot 84 Lot 85 Lot 86 Lot 87 Lot 87 Lot 88 Mean per chuck of m3s from all wafers measured Lot 88

22 measured m3s (nm) predicted m3s (nm) residual m3s (nm) Results Overlay X Measured, Predicted & Residual Integrated Metrology (IM) Lot 1 Lot 2 Lot 3 Lot 4 Lot 5 Lot 6 Lot 7 Lot 8 Lot 9 Lot 10 Lot 11 Lot 12 Lot 13 Lot 14 Lot 15 Lot 16 Lot 17 Lot 18 Lot 19 Lot 20 Lot 21 Lot 22 Lot 23 Lot 24 Lot 25 Lot 26 Lot 27 Lot 28 Lot 29 Lot 30 Lot 31 Lot 32 Lot 33 Overlay X m3s (nm) Measured per Wafer Lot 34 Lot 35 Lot 36 Select same wafer (705) from testing group Lot 37 Lot 38 Lot 39 Lot 40 Lot 41 Lot 42 Lot 43 Lot 44 Lot 45 Lot 46 Lot 47 Lot Overlay X m3s (nm) Prediction per Wafer Lot 49 Lot 50 Lot 51 Lot 52 Lot 53 Lot 54 Lot Lot Lot 59 Lot 60 Lot 61 With Predictions we identify jumps where process was intentionally manipulated a b Lot 62 Lot 63 Lot 64 Lot 65 Lot 66 Lot 67 Lot 68 Lot 69 Lot 70 Lot 71 Lot 72 Lot 73 Lot 74 Lot 75 Lot 76 Lot 77 Lot 78 Lot 79 Lot 80 Lot 81 Lot 82 Lot 83 Lot 84 Lot 85 Lot 86 Ck1 7= 9.61 nm Ck2 7= 9.26 nm Overlay X m3s (nm) Residual of Measured Wafers minus Prediction IM Residuals are used to flag changes to process not covered by training input c testing group R value Lot 87 Testing Wafers R= Ck1 7= 4.41 nm Ck2 7= Wafer Sequance Lot 88 a Measured Testing w705 Measured (F2N) x: 11.3 nm y: 8. b Predicted Testing w705 Predicted (F2N) x: 9.0 nm y: 7.4 nm c Residual Testing w705 Residual x: 6.6 nm y: 5.1 nm Slide 22

23 measured m3s (nm) predicted m3s (nm) residual m3s (nm) Results Overlay Y Measured, Predicted & Residual Integrated Metrology (IM) Lot 1 Lot 2 Lot 3 Lot 4 Lot 5 Lot 6 Lot 7 Lot 8 Lot 9 Lot 10 Lot 11 Lot 12 Lot 13 Lot 14 Lot 15 Lot 16 Lot 17 Lot 18 Lot 19 Lot 20 Lot 21 Lot 22 Lot 23 Lot 24 Lot 25 Lot 26 Lot 27 Lot 28 Lot 29 Lot 30 Lot 31 Lot 32 Lot 33 Overlay Y m3s (nm) Measured per Wafer Lot 34 Lot 35 Lot 36 Select same wafer (1278) from testing group Lot 37 Lot 38 Lot 39 Lot 40 Lot 41 Lot 42 Lot 43 Lot 44 Lot 45 Lot 46 Lot 47 Lot Overlay Y m3s (nm) Prediction per Wafer Lot 49 Lot 50 Lot 51 Lot 52 Lot 53 Lot 54 Lot Lot Lot 59 Lot 60 Lot 61 With Predictions we identify jumps where process was intentionally manipulated Lot 62 Lot 63 Lot 64 Lot 65 Lot 66 Lot 67 Lot 68 Lot 69 Lot 70 Lot 71 Lot 72 Ck1 7= 6.60 nm Ck2 7= 5.78 nm Overlay Y m3s (nm) Residual of Measured Wafers minus Prediction Residual are used to flag changes in process conditions d e f Lot 73 Lot 74 Lot 75 Lot 76 Lot 77 Lot 78 testing group R value Lot 79 Lot 80 Lot 81 Ck1 7= 4.36 nm Ck2 7= 4.77 nm Lot 82 Lot 83 Lot 84 Lot 85 Lot 86 Lot 87 Testing Wafers R= Lot 88 d Measured Testing w1278 Measured (F2N) x: 11.4 nm y: 8.9 nm e f Predicted Testing w1278 Predicted (F2N) x: 9.3 nm y: 4.9 nm Residual Testing w1278 Residual Slide Wafer Sequance x: 4.4 nm y: 5.7 nm

24 Outline Introduction Slide 24 How the function works Data separation into Training and Testing groups Training with Bayesian Automated Regularization Prediction Vs. Measured Overlay as regression plots Precision of Trained Function as a vector map Results Conclusion

25 Conclusion 1 With the Predictions we identify jumps in the overlay data where process was intentionally manipulated by the integration team Slide 25 a b c d Predicted Wafer 151 (F2N) Predicted Wafer 215 (F2N) Predicted Wafer 1493 (F2N) Predicted Wafer 1583 (F2N) x: 11. y: 8.2 nm x: 10.0 nm y: 4.6 nm x: 14.2 nm y: 13.0 nm x: 12. y: 7.6 nm predicted m3s (nm) Overlay X m3s (nm) Prediction per Wafer Testing Wafers R= a b c d

26 Conclusion 2 With Residuals we flag wafers from IM. Something other then inputs we trained with is effecting the overlay signature This can be used to remove a wafer from APC or to trigger an investigation Slide 26 a b c d Residual Wafer 146 Residual Wafer 811 Residual Wafer 1112 Residual Wafer 1567 x: 7. y: 6.9 nm x: 7.7 nm y: 5.6 nm x: 8.4 nm y: 7.9 nm x: 7.2 nm y: 6.8 nm residual m3s (nm) Overlay X m3s (nm) Residual of Measured Wafers minus Prediction a b c d Ck1 7= 4.41 nm Ck2 7= Wafer Sequance

27 Moving forward Work on this subject is open to users with interest in exploring the application, both in production and development environments Slide 27 Topics of interest include exploring effect; Fab context from outside the lithocluster has on the overlay prediction Increasing the number of parallel works and neurons has toward improving the R value (correlation coefficient) between the predicted output and target values in the testing dataset

28 : Hong-Goo Lee, Min-Suk Kim, Sang-Jun Han, Myoung-Soo Kim, Won-Taik Kwon & Sung-Ki Park : Kevin Ryan, Thomas Theeuwes, Kyu-Tae Sun, Young-Wan Lim, Daan Slotboom, Michael Kubis & Jens Staecker : The authors would like to thank ASML colleagues Jan Mulkens, Marcel Beems, Wolfgang Henke, Henry Megens, Peter ten Berge, Paul Luehrmann, Dick Verkleij, Frank van de Mast, Christophe Fouquet & coworkers in the Holistic Applications Development group for there assistance & feedback