Five Deep Learning Recipes For The Mask Making Industry
Five Deep Learning Recipes For The Mask Making Industry
Five Deep Learning Recipes For The Mask Making Industry
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc.Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members.
DL can solve many mask industry problems
https://doi.org/10.1117/12.2282943
https://digital.hbs.edu/platform-digit/submission/applied-materials-when-machines-talk-to-each-othe/
https://semiengineering.com/searching-for-euv-mask-defects/
Update Observation
Update
Model & Reward
Model
Action
Evaluate Evaluate Agent
Input Model Input Model Environment
Error Error (Model)
Output
Label
4
Recommendation engines use deep learning
DL technique used: autoencoding (unsupervised DL)
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 5
Mask rule check can generate thousands of errors
Best to report similar errors together
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 6
Unsupervised deep autoencoder captures similarity
After training, encoding abstracts out similar features
5.8539410e+01,
2.1572763e+02,
1.4674566e+02,
6.1346753e+01,
…
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 7
Mask rule check groups similar errors together automatically
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 8
Recipe #2:
Fast mask design construction
9
Medical diagnosis uses deep learning
DL technique used: U-net architecture (supervised learning)
https://github.com/jantic/DeOldify
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 10
Faster but less accurate simulation is useful for evaluation
Quickly estimate downstream impact
TrueModel(R)
Simulation
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 11
U-Net architecture translates images using pixel values
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 12
DL can construct mask shapes from shots faster
Current implementation reduces run time by 2.5x
DL
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 13
Max EPE error is below 3 nm even with overlapping shots
Not to replace simulation but useful for evaluation function
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. D2S confidential and proprietary 14
Recipe #3:
PCB component pick-and-place (PnP) error classification
15
Amazon Go automatic checkout uses deep learning
DL technique used: object classification, others (supervised DL)
https://www.cnet.com/pictures/photos-inside-amazon-go-store-no-cashiers-seattle/7/
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 16
Incorrectly picked components need to be identified
Monitored using images of picked components
https://www.alibaba.com/product-detail/pick-and-place-robot-arm_62008663790.html
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. Mycronic confidential and proprietary 17
This problem is difficult
Classical computer vision is not sufficient
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. Mycronic confidential and proprietary 18
Object classification uses deep CNN
https://neurohive.io/en/popular-networks/vgg16/
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 19
CNN based classifier was used to identify “Bad” PnP
Accuracy 99.76%, even on different machines
Bad-”Billboarded” Bad-”Tombstone”
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. Mycronic confidential and proprietary 20
Recipe #4:
Servo log anomaly detection for Flat Panel Displays (FPD)
21
Credit fraud detection uses deep learning
DL technique used: Deep anomaly detection (unsupervised learning)
https://contentpond.com/how-ai-can-analyse-the-fraud-detection-in-credit-card/
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 22
“Muras*” are unwanted defects in FPD mask writing
Surface log
Servo parameter logs record values for every x,y position of plate
Surface log is used to identify “Muras”
*Mura means irregularity in Japanese
https://webobjects2.cdw.com/is/image/CDW/5036714?wid=1142&hei=818&resMode=bilin&fit=fit,1
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. Mycronic confidential and proprietary 23
Anomaly detection use Generative Adversarial Networks (GAN)
https://arxiv.org/pdf/1901.08954.pdf
© Copyright 2019 Center for Deep Learning in Electronics Manufacturing, Inc. Mycronic confidential and proprietary 25
Recipe #5:
Digital twins creation
26
Digital twins can benefit from deep learning
DL technique used: mixture models, GANs (semi-supervised DL)
Today’s 3 PM talk “Making digital twins using the Deep Learning Kit (DLK)” will cover digital twins in more
details
https://www.attendmia.com/blog/2019/02/27/explore-the-latest-manufacturing-technology/
Confidential and Proprietary to Center for Deep Learning in Electronics Manufacturing and its IP Members. 27
SEM digital twin
https://www.pinterest.co.uk/pin/849421179688401557/