AI on edge device for laser chip defect detection

2019 IEEE 9th Annual Computing and Communication Workshop and Conference (CCWC)
Date of Conference: 07-09 January 2019
Conference Location: Las Vegas, NV, USA

Transcript

1. 1 AI on edge device for laser chip defect detection Dennis Hou, Tuo Liu, Yen-Ting Pan, Janpu Hou January 8, 2019
2. 2 Abstract Machine learning has been a major driver for improving semiconductor laser chip manufacture process. The virtual metrology system was used to enable the manufacturers to conjecture the wafer quality and deduce the causes of defects without performing physical metrology. However, building the virtual metrology system required a large amount of classified chip images. Therefore, a fast, accurate, portable image classifier was needed to fit modern flexible semiconductor laser manufacture setup, even without Internet connection. Based on a few pre-trained deep learning modes(AlexNet, ZFNet, and GoogLeNet), we use transfer learning to train the classifier on semiconductor distributed feedback (DFB) laser chip images. The GoogLeNet was identified to outperform the other two, and a portable image classifier was built. This paper has two main contributions: (1) A GoogLeNet-based semiconductor laser chip defect detection and classification network was developed with better than 97% accuracy in manufacturing production test. (2) The inference network is implemented on single board computer with an Intel Movidius Neural Compute Stick and USB digital microscope to form a low-power off-line handheld laser chip defect image classifier.
3. 3 Laser Chip Virtual Metrology : need defect detection Build model with machine learning Deep Learning Model Physical metrology result Virtual metrology result Processed Wafers Sampled Wafers Process equipment Metrology equipment
4. 4 What are the machine trying to learn? Looks like there’s dirt on my Dalmatian Dog vs Cat ? ➢ Dogs and cats looks different ➢ Every dogs and cats come from different family Dirt spot on Dalmatian ! ➢ All Dalmatians looks alike ➢ All our Dalmatian come from the same family (same manufacture process) ➢ Spot the dirt on the Dalmatian(defect detection) Feature Engineering
5. 5 What are the machine trying to learn? Different oncologist has different diagnostic on the same biopsy Diagnostic Medical Image ? ➢ The tumor is noncancerous (benign) ➢ Benign tumors press on blood vessels or nerves ➢ Cancerous tumor but not life threaten ➢ Cancerous tumor already crashed the bone, Rush to the emergency room? Inspect Defect Image ! ➢ None killer defect ➢ Low risk killer defect ➢ High risk killer defect ➢ True failure killer defect Label Engineering
6. 6 Traditional Sampled Laser Chip Defect Detection Digital Microscope Laser Chip Defects Dataset Domain Expert Laser Chips Labelled Images A Rule Image Created for each class Automatic Defect Detection Feature Engineering Label Engineering
7. 7 ML Workflow for Laser Chip Defect Classifier 1. Sample a Wafer 2. Crop a Chip image 3. Classified with CNN Coating Active Surface Edge Pass 4. Inference Output
8. 8 Model Building Dataset Class Image Number No Defects 10,000 Surface Defect 1,200 Edge Defect 1,200 Active Defect 1,200 Coating Defect 1,200
9. 9 0 25 50 75 100 0 10 20 30 Training Epoch Test Accuracy(%) Model Training Results
10. 10 Transfer Learning Based on Deep Learning Model Image Size Accuracy AlexNet 5 convolution layers 227x227x3 96.888% ZFNet 5 convolution layers 227x227x3 96.832% GoogLeNet 8 inception module layers 224x224x3 97.716% Using pre-trained model do the feature engineering, Re-train only the label engineering layers
11. 11 Data Pipeline and Workflow for Model Building Single Board Computer with Digital Microscope Laser Chip Defects Dataset Single Board Computer w/ VPU Image Classifier Multi-Core Computer w/ GPU Model Training Laser Chips Labelled Images Updated Model Labelled Images
12. 12 Smart Laser Chip Detection and Classification Portable, fast, without internet connection Digital Microscope Inference Engine Single Board Computer
13. 13 Single Board Computer with VPU and Digital Microscope Multi-Core Computer w/ GPU Model Training Laser Chips Updated Model Labeled Image Production Sampling Classification Workflow Wafer bin map defect diagnosis for continuous process improve
14. 14 Expert Defined ML Classifier Defect No Defect Defect No Defect 4000 (True Positives) 2 (False Positives) 5998 (True Negatives) 0 (False Negatives) Performance of Laser Chip Defect Detection
15. 15 Estimate Purity and Performance of Our Classifier Coating Active Surface Edge No Defect Purity Coating 994 3 2 1 1 99.301% Active 5 990 3 0 1 99.099% Surface 1 3 994 1 0 99.499% Edge 0 4 1 998 0 99.501% No Defect 0 0 0 0 5998 100.000% Class Performance 99.4% 99.0% 99.4% 99.8% 99.966% Expert Defined Classifier Total Purity 99.480% Total Performance 99.513%
16. 16 Kernel Learning + Deep Neural Networks. We advocate a methodic learning paradigm MINDnet (S.Y. Kung, Princeton University) ➢ Computationally efficient in training the networks ➢ Mathematically feasible to analyze Monotonically INcrease the Discriminative power, MIND-net uses a Discriminant Information (DI) optimization metric. On CIFAR-10 dataset: 93.57% (ResNet, 2015) < 96.01% (DenseNet, 2016) < 97.35% (NAS-Net, 2018) < 98.26% (MINDnet, 2018) Learning Deeper, Understand Better!
17. 17 Who Has Cited This Publication?