Slide 1: Data Science in Semiconductor Memory Manufacturing
Data Science in Semiconductor Memory Manufacturing Reviewed 2025 © 2022-2025 Micron Technology, Inc. All rights reserved. Information, products, and/or specifications are subject to change without notice. All information is provided on an “AS IS” basis without warranties of any kind. Statements rega
Slide 2: Copyright guidelines
Copyright guidelines By using any content provided by the Micron Educator Hub, you acknowledge that Micron Technology, Inc. (“Micron”) is the sole owner of the content and agree that any use of the content provided by the Micron Educator Hub must comply with applicable laws and require strict compli
Slide 3: How to cite sources from the Micron Educator Hub
How to cite sources from the Micron Educator Hub • Micron is committed to collaborate with Use case How to cite sources educators to make semiconductor memory education resources available through the Whole slide deck or whole No additional citation required document Micron Educator Hub Description:
Slide 4: Table of Contents
Table of Contents 1 Goal, Objectives and Target Audience 2 Data Science Big Picture 3 Data in Semiconductor Memory Manufacturing 4 Data Science Roles and Team Dynamics 5 Case Study 6 Key Terminology/Glossary 4
Slide 5: Data Science in memory manufacturing – goal and objectives
Data Science in memory manufacturing – goal and objectives Goal Objectives • Participants will gain understanding on source of • Identify the different sources of data in data and analytical methodologies in semiconductor memory manufacturing semiconductor manufacturing • Understand the different ro
Slide 6: Target Audience
Target Audience • Interns, NCGs (new college grads), and new employees in some Pro tip technical roles need to understand these concepts Everyone interviewing at Micron can • Examples of critical target audience roles at Micron that utilize use this presentation to prepare for these concepts the int
Slide 7: Data Science Big Picture
Data Science Big Picture 7
Slide 8: Industrial Revolution enablers by generation
Industrial Revolution enablers by generation 4th Industrial Revolution Early 21st Century 3rd Industrial Artificial Intelligence Revolution Information Technology 2nd Industrial Late 20th Century Revolution Intelligence Information 1st Industrial 19th ~ 20th Century Machine Learning Big Data Revolut
Slide 9: Key enablers for Big Data and Artificial Intelligence
Key enablers for Big Data and Artificial Intelligence • Massive data collection capability Artificial Intelligence • Highly scalable computational power and Information Technology storage Intelligence Information • Sophisticated machine learning algorithms Machine Learning Big Data Gen AI (LLM, etc.
Slide 10: Massive data collection
Massive data collection Thermography Vibration Sensor • Intelligent sensors with edge node – With data processing (ML) & transmittance A/D capability Converter • 5G wireless network – Super high band width for wireless real time communication which is 100X faster than 4G • Optical network – Wired ne
Slide 11: Highly Scalable Computational Power
Highly Scalable Computational Power Mid 1990’s Mid 2000’s Current • Personal Computer • Google’s advanced search engine • Non-IT Industries to (Intel/Microsoft/Apple) indexes massive internet data use scalable • Cloud technology clusters servers to computational power scale up data processing capabi
Slide 12: Highly Scalable Computing Power and Storage
Highly Scalable Computing Power and Storage • Corporations can now scale up both computational power and storage as their businesses grow utilizing cloud service • Enables individuals to have massive data processing and storage capability as collected data grows 12
Slide 13: Machine Learning
Machine Learning 1950’s ~ 1990’s 2000’s ~ early 2010’s Current • Neural Network is invented in the 1950’s • Decision Tree Algorithm • Enables nonparametric data analysis • A primitive neural network is developed • Random Forest • No prefixed distribution nor prefixed for the United States Postal Ser
Slide 14: Traditional Statistics: what traditional parametric statistics has been
Traditional Statistics What traditional parametric statistic has been… 1 1 (𝑙𝑛𝑥 − 𝜇)2 × exp(− ) Poisson Distribution 𝜆𝑘 𝑒 −𝜆 𝑥 𝜎 2𝜋 2𝜎 2 Asymmetric and skewed 𝑘! Log Normal Distribution Asymmetric and skewed E0 O3 N4 D4 N4 M4 E4 M0 O3 M0 O6 L0 M0 B5 Q0 B6 J4 Q0 B5 D5 J5 X0 M0 O3 O3 B4…
Slide 15: Traditional Statistics: Regression model e.g. polynomial
Traditional Statistics Regression model e.g. polynomial The data is known to have no correlation with Q-time The data is known that cliff is present at 3.7 hours [%] [%] When you use regression model, you presume such physical Yield Yield mechanism Q-Time Q-Time Fitting by 3rd order of polynomial •
Slide 16: What machine learning does (e.g. Random Forest)
Machine Learning What machine learning does (e.g. ”Random Forest”) Your Data Sheet to Explore y x01 x02 x03 x04 x05 x06 x07 x08 x09 x10 x11 x12 x13 x14 x15 7 1 ~10 10 2 13 3 16 4 y y 19 5 22 6 ~25 25 28 7 8 31 9 34 10 37 11 40 12 43 13 46 14 𝑥08 49 15 𝑥02 x02 [1] Equation: y = 𝑓(𝑥01, 𝑥02, 𝑥03…
Slide 17: Random Forest to find highly correlated parameter
Machine Learning Random Forest to find highly correlated parameter Node by relevant variable Terminal Node of irrelevant indicator [7] Repeat [2] to [5] to create multiple trees to form a “forest” The forest and/or trees with node of relevant variable have better predictability at the terminals on “
Slide 18: Machine Learning transforms data analysis
Machine Learning transforms data analysis • In memory manufacturing, massive number of parameters Light Source Reticle (Mask) are controlled and understood ~5000 parameters – 𝑌 = 𝑓(𝑋1, 𝑋2, 𝑋3 … … … . 𝑋𝑛) from photolithography – Where n can be hundreds to thousands per parameter Lens system Y…
Slide 19: Machine Learning vs. Traditional Statistics
Machine Learning vs. Traditional Statistics Traditional Statistics approach • Expectations for Machine Learning in manufacturing are very high 0.4 • However, manufacturing engineers are deeply rooted in p-value based 0.3 statistical methods 0.2 – p-value ( p < 0.05) based judgement has long served a
Slide 20: Generative AI (Gen AI)
Generative AI (Gen AI) https://www.micron.com/about/micron-glossary/generative-ai • Definition: • Prompt engineering: – Generative AI - commonly known as Gen AI - refers – The process by which humans interact with generative to artificial intelligence systems capable of creating AI, crafting, and re
Slide 21: Data in memory manufacturing
Data in memory manufacturing 21 Image created for Micron by Copilot
Slide 22: Micron delivers a broad portfolio of memory and storage
Micron delivers a broad portfolio of memory and storage Memory Storage High-bandwidth Multichip memory (HBM), Low-power DRAM packages Solid-state drives Flash storage and graphics and DRAM (DRAM and (SSD) (NAND and NOR) memory (GDDR) NAND) solutions Examples of Micron portfolio products 22
Slide 23: DRAM memory cell
Let’s explore DRAM Memory Speed？ IO Timing ? DRAM Controller • Dynamic Capacitor • Random To store 1/0 bit Read Write • Access • Memory How many Transistor DRAM bits working ? products DRAM Memory Cell • Billions of DRAM memory cells, each with a transistor and a capacitor, store binary information
Slide 24: Where does the data come from within DRAM?
Where does the data come from within DRAM? Capacitor (Cylinder) Capacitor This is a digital Transistor composite image, not a real SEM image (Scanning Electron Microscopy). Capacitor IO Transistor Controller 24
Slide 25: Proportional accuracy
Proportional accuracy 2 𝑛𝑚 2.6 𝑢𝑚 = 300 𝑚𝑚 400 𝑚 Photolithography Tool A 400m long bullet train with multiple cars would have z Source Light to stop within a 2.6μm precision (less than the width of Reticle a human hair) at a platform to match the precision of the photolithography tool!…
Slide 26: Proportional aspect ratio
Proportional aspect ratio An example from Japan: 634 𝑚𝑒𝑡𝑒𝑟𝑠 (~2080 𝑓𝑒𝑒𝑡) 68 𝑚𝑒𝑡𝑒𝑟𝑠 (~223 𝑓𝑒𝑒𝑡) 5 Tokyo Skytree buildings! Aspect Ratio (AR) = height/width height width AR for the Tokyo Skytree ~ 10 AR for a DRAM capacitor cylinder ~50! 26
Slide 27: Proportional aspect ratio
Proportional aspect ratio An example from the United States: 5 Washington Monuments! Aspect Ratio (AR) = height/width 555 ft height Washington Monument, Washington DC Aspect ratio for DRAM capacitor cylinder ~50! Aspect ratio Washington width Monument ~10 55.5 ft 27
Slide 28: How small is the nanometer scale
How small is the nanometer scale It is difficult to imagine the nanometer scale. …that is similar to comparing a 0.5 mm line from a Let’s say that a DRAM line width is mechanical pencil to 10 kilometers, which is the approximately 15 nm. If we compare the line distance from Shinagawa to Ikebukuro st
Slide 29: Data in Semiconductor manufacturing
Data in Semiconductor manufacturing Prime Data Equipment Sensor Equipment History • Data that has immediate impact to yield Sensor data collected during wafer ❑ Which equipment processed a wafer at & quality of products processing a specific process step? • Collected inline & real time during ❑ When
Slide 30: Other data in Semiconductor manufacturing
IOT = “Internet of Things” Other data in Semiconductor manufacturing Weather IoT evolution Temperature, humidity, IoT: Internet of Things lighting Has enabled additional equipment/processing data collection Potentially impact yield & quality Clean Room Equipment Parts Environment Transport robot Hum
Slide 31: Priorities in Memory Manufacturing
Priorities in Memory Manufacturing Yield Make it functional Yield & Volume Quality Make better Yield/Maintain Development Yield Ramp Cycle Time Prototype Make faster Cost Spend less time Data shown in previous slides is analyzed to optimize these key priorities This data has been produced solely for
Slide 32: Big Data/AI integration in semiconductor industry
Big Data/AI integration in semiconductor industry New data Environment Real Fab Facility Clean room Digital Twin Virtual Fab Equipment enables virtual fab-wide experiment without physical impact Drive Yield/Quality Cost/Cycle Time improvement 32
Slide 33: Data Science Roles and Team Dynamics
Data Science Roles and Team Dynamics 33
Slide 34: How Data Science Team is formed
How Data Science Team is formed Business/Finance/Supply Chain Business Process Business Analyst Domain Expertise Process Analyst Engineering/Factory/Medical Domain Knowledge Mathematics/Statistics/Machine Full Stack Data Scientist Data Science Methods Learning/Visualization Scientist Solution Archit
Slide 35: Data Science Dynamics: Skills/Tools Coverage
Domain Expertise Data Science Dynamics: Skills/Tools Coverage Data Science Fab Equipment Equipment / sensors / IofT ML coding proficiency Data Engineering Python, R, etc. Database architecture Scikit- Tensor Robots Fabrication process from wafer to product Learn Flow API O3 E0 N4 D4 N4 development M
Slide 36: Dynamics of Data Science Project
Dynamics of Data Science Project Domain Knowledge Driven Request Prototype Needs Technology Development Data Scientist Data Engineer Engineers Business UI Engineer Manufacturing Process Analyst Engineers Integrated Productization Solution IT Skill & Knowledge Driven How Data Science Project is Execu
Slide 37: Prerequisite for Data Science in memory manufacturing
Prerequisite for Data Science in memory manufacturing • Can you tell how DRAM memory stores single bit “0/1” information? – Hint: Can use Micron Educator Hub to learn more, e.g., Intro to Memory • Can you tell how DRAM memory is made? – Hint: Can use Micron Educator Hub to learn more, e.g., Intro to
Slide 38: Prerequisite for Data Science Continued
Prerequisite for Data Science Continued • Can you explain the difference between mean and median? – Hint: what if data distribution is skewed? What if there is excursion point? • Can you explain difference between supervised and unsupervised learning? – Hint: e.g. learning with labeled data for lear
Slide 39: Prerequisite for Data Science continued
Prerequisite for Data Science continued • Can you develop analytical scripts with Python and/or R? – Example: print (‘Hello World’) in Python • Can you extract data from a database using SQL? – Example: “select * from data_table” • Can you create an interactive user interface (UI) and what tools can
Slide 40: Integrating Data Science Jigsaw Puzzle Pieces into One
Integrating Data Science Jigsaw Puzzle Pieces into One Problem Algorithm Statement Statistics Integrated Data Visualization DS (Data Science) Solution Sourcing User Interface Image created for Micron by Adobe Firefly 40
Slide 41: Case Studies
Case Studies Image created for Micron by Copilot 41
Slide 42: Case Study: Wafer Map Analytic
Case Study: Wafer Map Analytic Domain knowledge • Each memory die on wafer is electrically tested Wafer Map – Check every die functionality to ensure it is sellable to customer • A Wafer Map is generated with the test results. Each die is represented by a rectangle in the wafer map. Colors and codes
Slide 43: Electrical Functional Test Flow
Electrical Functional Test Flow Domain knowledge Probe Tests -> hundreds of tests to validate capacitors correctly store 1s and 0s: includes data reads and writes Test Data to be characterization Repair Pass or Failure collected data 1 1 1 1 1 1 1 1 1 Fails can be categorized Tester estimates if the
Slide 44: Wafer Map of Probe Test Data: Bit Map
Wafer Map of Probe Test Data: Bit Map Domain knowledge Single bit fails • Bit Map – Fails can be categorized into different fail Double bit fail types, e.g., single bit, double bit, column/row Row fail – Fails can also be visualized in a bit map – Fail data can be used to determine how to repair usi
Slide 45: Wafer Map of Probe Test Data: Repair Density
Wafer Map of Probe Test Data: Repair Density Domain knowledge • Repair Density – How many repair elements were used to repair fails found during testing – Data levels: - Lot level data (e.g., wafer average repair count) - Wafer level data (e.g., die average repair count) - Die level data (e.g., repa
Slide 46: Wafer Map of Probe Test Data: Fail Bin
Wafer Map of Probe Test Data: Fail Bin Domain knowledge • Fail Bin Pass – The probe testing consists of a series of tests, each with specific conditions that test different circuit functionality – Each test has a name and an associated color and character - For example, if a test finds open contacts
Slide 47: Wafer Map by resolution of the data
Wafer Map by resolution of the data Domain knowledge Bit Map Repair Density Fail Bin Map Rich Information on yield/quality improvement opportunity High Level Summary e.g., Yield This data has been produced solely for educational purposes and does not represent real Micron data. 47
Slide 48: Problem statement
Problem Problem statement Statement • Semiconductor Fabs produce a remarkably high volume of wafer outs daily—often in the hundreds or even thousands. • Within this output, some 'signatures' may emerge, warranting further investigation • Given time constraints, how would you efficiently classify or
Slide 49: Wafer Map Classification by Human
Problem Wafer Map Classification by Human Statement • A Micron engineer in the Yield Enhancement Group was asked to monitor & summarize bit map of “pizza slice” signature issue every two weeks. • The engineer has to look at the “wafer map” for hundreds of wafers individually, to find which wafers ha
Slide 50: Wafer Map Analysis Data vs. Algorithm
Algorithm Algorithm Wafer Map Analysis Data vs. Algorithm Statistics Statistics Data Data Type Analytic Approach Numeric clustering algorithm Bit Map Summary Numeric e.g. K-mean Expectation Maximize Numeric clustering algorithm Repair Density Numeric e.g. K-mean Expectation Maximize Deep Learning Pa
Slide 51: Wafer Map Classification Approach
Algorithm Wafer Map Classification Approach Statistics WAFER_ID LOT_ID Zone10 Zone11 Zone12 Zone13 Zone14 Zone15 Zone16 Zone17 Zone18 Zone19 Zone20 Zone1 Zone2 Zone3 Zone4 Zone5 Zone6 Zone7 Zone8 Zone9 A 1 A 2 A 3 A 4 A 5 A 6 A 7 A 8 A 9 A 10 A 11 A 12 A 13 A 14 A 15 Divide into 20 equal area zones
Slide 52: Wafer Map Clustering
Integrated DS Wafer Map Clustering Solution Pizza Slice Pizza Slice Pizza Slice Pizza Slice • Classification on 2000 wafers failure mode • Visual inspection by human: Hours to a day • By algorithm: 5 min This data has been produced solely for educational purposes and does not represent real Micron d
Slide 53: Comparison of human classification vs. algorithm
Integrated Comparison of human classification vs. algorithm DS Solution Classification by date by algorithm Week 5 – Week 6: X Week 5 – Week 6: X: high occurrence Week 7 – Week 8: ✓ Week 7 – Week 8: ✓: little occurrence Week 9 – Week 10: ✓ Week 9 – Week 10: ✓: little occurrence Week 11 – Week 12: Δ
Slide 54: Further Drilldown
Integrated Further Drilldown DS Solution Bad Wafers • Subsequent drilldown analysis of manufacturing data Good (pizza slice) revealed a clear distinction between good wafers and Wafers those exhibiting 'pizza slice’ signature, which were linked to a specific process step and equipment • This insight
Slide 55: Summary on Wafer Map Signature Analysis
Summary on Wafer Map Signature Analysis • Clustering algorithm automated manual cognitive work by human • Visually identifies possible root cause of quality degradation • The outcome is objective, robust and more reproducible than human classification • With further drilldown the result can be presc
Slide 56: Summary & Conclusion
Summary & Conclusion • Catalyst for Big Data & AI in Smart Manufacturing – Enabled by highly scalable computing, advanced machine learning algorithms, and ultra-high bandwidth IoT networks • Essential Role of Data in Memory Manufacturing – Characterized by extremely small, precise dimensions requiri
Slide 57: Key Terminology Glossary
Key Terminology Glossary 57
Slide 58: Glossary
Glossary Term or acronym Definition/description AI Artificial Intelligence AR Aspect Ratio Big Data Large and complex data sets that require advanced methods and technologies to store, process, and analyze. Capacitance The ability of a system to store an electric charge, measured in farads. Critical
Slide 59: Micron copyright and trademark notice
© 2022-2025 Micron Technology, Inc. All rights reserved. Information, products, and/or specifications are subject to change without notice. All information is provided on an “AS IS” basis without warranties of any kind. Statements regarding products, including statements regarding product features,