Value Creation of AI in the Manufacturing Industry Janet George Fellow/Chief Data Scientist Western Digital Corporation September 28 th, 2016 2016 Western Digital Corporation or its affiliates. All rights reserved. 1
A Global Leader in Storage Technology $18.3 $15.9 LTM Revenue 1 $11.2 $11.2 $10.3 $5.5 $4.8 $3.4 $2.5 (Information Storage) (Storage & Memory) (NAND) (NAND) (NAND) (NAND) 1 LTM revenues based on most recent public filings and Wall Street research; Western Digital and SanDisk LTM as of 4/1/2016; Toshiba represents March 2016 LTM revenue. 2016 Western Digital Corporation or its affiliates. All rights reserved. 2
Moving Mountains of Data Core Register Core L1 Cache Core L2 Cache Shared L3 Cache DRAM HDD Source: Western Digital estimates 2016 Western Digital Corporation or its affiliates. All rights reserved. 3
Access Time (sec) Memory & Storage Hierarchy 1.E-02 1.E-03 Storage HDD 1.E-04 1.E-05 1.E-06 1.E-07 1.E-08 1.E-09 NAND ReRAM DRAM Memory 0.01 0.1 1 10 Die Cost ($/GB) PCM CBRAM Storage-Class Memory STT-MRAM SRAM Embedded ReRAM Non-Volatile Volatile 2016 Western Digital Corporation or its affiliates. All rights reserved. 4
The FAB 4 of Semiconductor Nirvana Case Study 2D NAND 3D NAND Transition Cost Scalability Scale Ecosystem 2016 Western Digital Corporation or its affiliates. All rights reserved. 5
When the Chips are Down, the Wafers are Up! How Tall Would it Be? Number of wafers produced in 2016 In Yokkaichi 2016 Western Digital Corporation or its affiliates. All rights reserved. 6
When the Chips are Down, the Wafers are Up! Mount Kilimanjaro 19,340ft Number of wafers produced in 2016 In Yokkaichi Eiffel Burj Khalifa Tower 2717ft 984ft Mount Fuji 12,388ft 2016 Western Digital Corporation or its affiliates. All rights reserved. 7
2D NAND Architecture BL BL Contact Floating Gate (memory cell) Source Plate SGD WL SGS NAND String Si-sub 2016 Western Digital Corporation or its affiliates. All rights reserved. 8
3D NAND Architecture SGD WL SGS Memory Holes Source Plate Memory Cell 2016 Western Digital Corporation or its affiliates. All rights reserved. 9
The BiCS Scaling March 48L 64L 24L BiCS5 BiCS1 BiCS2 BiCS3 BiCS4 2016 Western Digital Corporation or its affiliates. All rights reserved. 10
3D Technologies for SCM 3D XPoint 3D NAND Cell and Materials Selector Process Architecture Product Ecosystem 2016 Western Digital Corporation or its affiliates. All rights reserved. 11
Artificial Intelligence value creation First Class of problems New technology Node Creation Pattern Recognition & Machine learning 2016 Western Digital Corporation or its affiliates. All rights reserved. 12
Memory Hole (Background Knowledge) (3D) stacked memory structure referred to as a Bit Cost Scalable (BiCS) architecture. For example, a 3D NAND stacked memory device can be formed from an array of alternating conductive and dielectric layers. A memory hole is formed through the layers to define many memory layers simultaneously. A NAND string is then formed by filling the memory hole with appropriate materials. A straight NAND string extends in one memory hole, while a pipe- or U-shaped NAND string (p-bics) includes a pair of vertical columns of memory cells. Control gates of the memory cells may be provided by the conductive layers. 2016 Western Digital Corporation or its affiliates. All rights reserved. 13
Pattern Recognition -New Technology Node Creation Memory Hole Problems Class 1 3D NAND 2016 Western Digital Corporation or its affiliates. All rights reserved. 14
Pattern Recognition- New Technology Node Creation Memory Hole Problems Class 1 ML Analysis 3D NAND 2016 Western Digital Corporation or its affiliates. All rights reserved. 15
Pattern Recognition- New Technology Node Creation Memory Hole Shape Problems sub patterns Ex 1 2016 Western Digital Corporation or its affiliates. All rights reserved. 16
Pattern Recognition- New Technology Node Creation Memory Hole Shape Problems sub patterns Ex 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 17
Pattern Recognition- New Technology Node Creation Memory Hole Shape Solution Step 1- De-noising 2016 Western Digital Corporation or its affiliates. All rights reserved. 18
Pattern Recognition- New Technology Node Creation Memory Hole Shape Solution Step 2 Curation 2016 Western Digital Corporation or its affiliates. All rights reserved. 19
Pattern Recognition- New Technology Node Creation Memory Hole Shape Solution Step 3 Classification 2016 Western Digital Corporation or its affiliates. All rights reserved. 20
Pattern Recognition New Technology Node Creation Memory Hole Shape Solution Step 4 Complexity in Training 2016 Western Digital Corporation or its affiliates. All rights reserved. 21
Pattern Recognition AI Techniques Memory Hole Shape Solution Elastic Bunch Map Graphing 2016 Western Digital Corporation or its affiliates. All rights reserved. 22
Pattern Recognition New Technology Node Creation Memory Hole Shape Solution Step 5 Roundness 2016 Western Digital Corporation or its affiliates. All rights reserved. 23
Pattern Recognition New Technology Node Creation Memory Hole Shape Solution Step 5 Circularity 2016 Western Digital Corporation or its affiliates. All rights reserved. 24
Pattern Recognition New Technology Node Creation Memory Hole Shape Solution Step 6 Curvature 2016 Western Digital Corporation or its affiliates. All rights reserved. 25
Pattern Recognition New Technology Node Creation Memory Hole Shape Solution Step 7 & More Diameter Radius Inner & Outer Patterns detection Sub-pattern detection 3D NAND All known AI/ML techniques Published or in Open Source (leading edge) applied/used 2016 Western Digital Corporation or its affiliates. All rights reserved. 26
Artificial Intelligence value creation Second Class of problems Failure Pattern Recognition & Machine learning 2016 Western Digital Corporation or its affiliates. All rights reserved. 27
Pattern Recognition Shot Map Millions of Wafer Failure Patterns Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 28
Pattern Recognition Heat Map Millions of Wafer Failure Patterns Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 29
Pattern Recognition Millions of Wafer Failure Patterns Modes Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 30
Pattern Recognition Millions of Wafer Failure Patterns Modes Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 31
Pattern Recognition Zone (AI/ML) Analysis Millions of Wafer Failure Patterns Modes Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 32
Pattern Recognition Failure Modes Stacked Complexity Millions of Wafer Failure Patterns Modes Class 2 Impact = (occurrence rate) x (defects rate) 2016 Western Digital Corporation or its affiliates. All rights reserved. 33
Pattern Recognition Millions of Wafer Failure Patterns Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 34
Decision Tree and Random Forest Comparisons Millions of Wafer Failure Patterns Class 2 2016 Western Digital Corporation or its affiliates. All rights reserved. 35
Artificial Intelligence value creation Third Class of problems Machine learning & Correlations 2016 Western Digital Corporation or its affiliates. All rights reserved. 36
Correlations in Single Factor Machine Learning problems Class 3 2016 Western Digital Corporation or its affiliates. All rights reserved. 37
Finding Nemo! Correlation and value (non-deterministic) Permutation and Combination of every known & unknown correlation Prep Manufacturing Data Measure Test Par 1 Screen Screen 1 Clean Height Par 2 Screen 2 Tool Config Film Oxide Par 3 Screen 3 ALO Cover Par 4 Screen 4 Wet Thickness Par 5 Screen 5 2016 Western Digital Corporation or its affiliates. All rights reserved. 38
Going Deeper into the Device Physics Data Heteroskedastic data Heteroskedastic: A measure in statistics that refers to the variance of errors over a sample. Heteroskedasticity is present in samples where random variables display differing variabilities than other subsets of the variables. 2016 Western Digital Corporation or its affiliates. All rights reserved. 39
Studying the Complexity of Data 2016 Western Digital Corporation or its affiliates. All rights reserved. 40
Higher Orders of Complexity 2016 Western Digital Corporation or its affiliates. All rights reserved. 41
Challenges in manufacturing data (New technology Node Creation Material instability. Deformation, does not bond Material build-up Y4 Substrate over heating. Heat sink Resistive loss Deposition issues Thickness Warpage Wetting layers Oxidation Diffusion barrier DPPM known and unknown causes, complex error recovery Coupling effects, adjacent track interferences. Y3 Y5 2016 Western Digital Corporation or its affiliates. All rights reserved. 42
Dealing with Heteroskedastic data - Challenges Machine learning model building requires constant optimization training and re-training with change. Ranking and weighting correlations for DPPM Page Rank Model. Linear regression versus Random forest What works for the data/value creation 2016 Western Digital Corporation or its affiliates. All rights reserved. 43
Key Take away: Internet of Machines- High Variability (Heteroskedastic, leading edge, higher order complexity in data) 2016 Western Digital Corporation or its affiliates. All rights reserved. 44
Bringing the possibilities of data to life. 2016 Western Digital Corporation or its affiliates. All rights reserved.
Janet George Short Bio Janet George: Fellow/Chief Data Scientist Big Data Platform/Data Science/Cognitive Computing Background/Relevant Experience/Expertise At Western Digital Building global core competencies Shaping, driving, implementation of the Big Data platform Data Science, AI, Machine learning, Pattern recognition in Storage/Flash Memory manufacturing Industry Experience in Big Data Platform, Data Science/Cognitive Computing/ Artificial Intelligence /Machine learning and Distributed Computing/Compliers Prior served as Managing Director at Accenture technology labs Prior served as Head of Yahoo Labs/Research Engineering inventing Next Generation Platforms, Cloud Infrastructures and Data Science AI/Machine Learning Prior @ ebay and Apple Computer Education: Bachelors and Advanced Master Degree with distinction in Computer Science, Mathematics, with a thesis focus on Artificial Intelligence and Machine learning.