S4695 A Real-Time Defocus Deblurring Method for Semiconductor Manufacturing

Transcription

1 S4695 A Real-Time Defocus Deblurring Method for Semiconductor Manufacturing T. Sakuyama*, Y. Hikida*, H. Sano*, K. Taniguchi* T. Funatomi**, M. Iiyama**, M. Minoh** Dainippon Screen Mfg. Co., Ltd.* Kyoto University**

2 Today s talk CPU GPU New!! FPGA Semiconductor manufacturing equipment We have adopted the GPU as the equipment s processer, to realize next generation alignment with Advanced computation. CPU GPU faster FPGA GPU more flexible 1

3 Agenda Background Semiconductor Manufacturing equipment always demands more precise technology Computations Load in manufacturing process Next generation alignment with Advanced computation Requirements in the processing time and solution Summary I d like to talk about a case of adopting the GPU to semiconductor manufacturing equipment. 2

4 Semiconductor Manufacturing equipment A chain of equipment cooperates to produce semiconductor. The Semicondutor manufacturing equipment always demands more precise technology. Our products photoresist Cleaning form film Cleaning coating lithography process the film etching ion implantation Resist Stripping detection 3

5 Computations Load in manufacturing process In equipment, various processes are running. Machine control -alignment Logger Image processing Process control etc. Zheng L. (2008), System-on-Chip Applications, Lecture Notes Electronics, Computer and Software Systems. Royal Institute of Technology (KTH), Stockholm. According to Moore's law, these are becoming more precise and that makes computational cost increase. Today, I will focus on the Alignment. 4

6 1Setting a material The material is a silicon wafer with some patterns on it. Detail of the Alignment Material Semiconductor manufacturing equipment 5

7 Detail of the Alignment 1Setting a material 2Capturing Alignment Mark images camera camera Material 6

8 Detail of the Alignment 1Setting a material 2Capturing Alignment Mark images The camera will capture fixed region of the material. camera camera Material Alignment Mark Alignment mark image 7

9 Detail of the Alignment 1Setting a material 2Capturing Alignment Mark images 3Measuring positions of the Marks the equipment recognizes the position of a material. camera camera Material detected the mark Alignment Mark Alignment Mark image 8

10 Detail of the Alignment 1Setting a material 2Capturing Alignment Mark images 3Detecting positions of the Marks 4Adjusting the position of a material camera As a result, Equipment will be able to realize expected performance. camera Material Alignment Mark According to the positions of detected marks, the equipment corrects the position of the material. 9

11 Problem of the Alignment out of focus Incomplete auto-focus makes the detection fail. If such an image was taken, equipment will fail to detect the alignment mark. As a result, equipment will not be able to adjust the position of a material. camera camera???? Material Why sometimes the auto-focus fail? Alignment Mark Alignment Mark image 10

12 Why sometimes auto-focus fail? Because focus range is narrow in high magnification. The depth of field is determined by the magnification. In case of semiconductor, depth of field is very small due to the required magnification for the ever becoming smaller pattern. a few micro-meters alignment Camera Out of focus Focus image Out of focus This problem can be solved with optical components, but the solution has a limitation. 11

13 Agenda Background Semiconductor Manufacturing equipment always demands more precise technology Computations Load in manufacturing process Next generation alignment with Advanced computation Requirements in the processing time and solution Summary So, we have attempted to solve the problem of out-of-focus using the advanced computation. 12

14 Advanced computation in next generation alignment The advanced computation uses a defocus deblurring technique. This technique is generally known, but there has not been realized in industrial equipment.? Defocus Deblurring Defocus image Focus Image Can t detect a mark Alignment fails If it is possible to always correct the out-of-focus image, the problem does not occur. Detect the mark Alignment succeeds! 13

15 Basic Principle of the defocus deblurring Blurred process can be expressed by the following equation. Captured image f (blur image) Ground truth f0 Blur kernel k (focus image) (Point spread function:psf) Noise n f = f0 * k + n F= F0 K + N Fourier transform In frequency domain, focus image f0 can be estimated. However, this method is not robust to noise 14

16 Basic Principle of the defocus deblurring Blurred process can be expressed by the following equation. Captured image f (blur image) Ground truth f0 Blur kernel k (focus image) (Point spread function:psf) Noise n F 0 = K F 2 K + C 2 K : the complex conjugate of K : noise-to-signal ratios C Wiener filter This method is robust to noise. In frequency domain, focus image f0 can be estimated. 15

17 Captured image f PSF k Detail of the process Defocus Deblurring By wiener filter Estimated focus image f0 Here, the important point is that the PSF varies depending on the degree of blur We must determined the PSF in some way. 16

18 Captured image f PSF k Detail of the process Defocus Deblurring By wiener filter Estimated focus image f0 1PSFs for several depths are captured. 17

19 Captured image f PSF k Detail of the process Defocus Deblurring By wiener filter Estimated focus image f0 1PSFs for several depths are captured. 2focus images are estimated with each PSF and captured image. Estimated Results f0 18

20 Captured image f PSF k Detail of the process Defocus Deblurring By wiener filter Estimated focus image f0 1PSFs for several depth are captured. 2focus images are estimated with each PSF and captured image. Estimated Results f0 3An image with the highest quality is selected for the equipment among the set of the results 19

21 Detail of the process Preprocess Main process Our process -PreProcess -MainProcess PSFs Input PSFs Fourier transformation Make Wiener filters Input an image Fourier transformation Wiener deconvolution Captured Image Inverse Fourier transformation Only the main processes run in equipment Select an image with the highest quality The main process computes a set of deblurring results calculated from each Wiener filter and the captured image in frequency domain. In equipment Result Image 20

22 Agenda Background Semiconductor Manufacturing equipment always demands more precise technology Computations Load in manufacturing process Next generation alignment with Advanced computation Requirements in the processing time and solution Summary 21

23 Requirements in the processing time Equipment has requirement. Real-time processing(100msec/image) The captured image is VGA, 8bit grayscale. Distance to the focus point is unknown. Intel Core i7 X980 : Computation time : 2.0sec/image 22

24 Solution Our process -PreProcess -MainProcess FFT Deconvolution PSFs Preprocess Input PSFs Fourier transformation Make Wiener filters These calculations are expected to speed up by parallelization,which suit GPU to achieve the processing time to the target. Main process Input an image Fourier transformation Wiener deconvolution Inverse Fourier transformation Select an image with the highest quality In equipment Defocus Image Captured Image Result image 23

25 experimental device & Result CPU Clock Speed 3.3GHz Max Frequency 3.6GHz Cache 12MB # of Cores 6 Price(approx.) $ 300 Computation time 2.0sec/image GPU Memory Clock 2.6GHz Memory Size 5GB CUDA Cores 2496 Core Clock 706MHz Price(approx.) $ 3,000 Computation time 95msec/image Intel Core i7 X980 NVIDIA TESLA K20 24

26 Computational time on CPU or GPU The requirement of 100msec has satisfied. Computational time[msec] Fourier transformation Deconvolution MKL:CPU cufft+cuda:gpu others Total time CPU:2.0sec GPU:95msec As originally planned, calculations parallelly performed on the GPU such as deconvolution or Fourier transform runs significantly faster than the other items 25

27 I think The GPU has good balance between computation speed and flexibility for the alignment It is possible that GPU works effectively in other process 26

28 Agenda Background Semiconductor Manufacturing equipment always demands more precise technology Computations Load in manufacturing process Next generation alignment with Advanced computation Requirements in the processing time and solution Summary 27

29 Summary We have adopted the GPU as equipment processer High Performance Computation is required to next generation alignment. -CPU(Intel Core i7 X980) : 2.0sec GPU(NVIDIA TESLA K20) : 95msec? Defocus Deblurring By Wiener filter Defocus image Deblurred Image It is possible that GPU works effectively in other process 28

30 CPU GPU New!! FPGA For another case that adopted the GPU on the equipment -poster session: P4158 please come to see details 29

31 Acknowledgment Mr. H. Sano Mr.Y. Hikida Dr. K. Taniguchi Ap. T. Funatomi Ap. M. Iiyama Prof. M. Minoh I would like to give heartful thanks to these members whose comments and suggestions were very important for me. And, I would like to express my gratitude to JST for their financial support. This work is supported by Adaptable and Seamless Technology Transfer Program through target-driven R&D, JST

32 Thank you for your kind attention 31

33

34