Big Data & Analytics
for Semiconductor Manufacturing
半導体生産におけるビッグデータ活用
半導体生産におけるビッグデータ活用
半導体生産におけるビッグデータ活用
半導体生産におけるビッグデータ活用
Ryuichiro Hattori
服部
隆一郎
Intelligent SCM and MFG solution Leader
Global CoC (Center of Competence) Electronics team
General Business Services
Agenda
What is Big Data ?
Big Data in Semiconductor Manufacturing
Big Data and Analytics architecture
Volume
Velocity
Variety
Veracity*
Data at rest
Terabytes to
exabytes of existing
Data in motion
Streaming data,
milliseconds to
Data in many
forms
Structured,
unstructured, text,
Data in doubt
Uncertainty due to
data inconsistency
Fall 2013
Problem statement:
“Conventional or standard analytical methods and technologies are built for predictive
modeling on a small scale, not for investigation of hundreds or thousands of potential
factors and interactions
”
“Engineers with standard analytical techniques and tools have become the bottleneck,
outpaced by data volumes and complexity”
“New methods and software are needed to bridge the gap between analysis and action”
“Automated data mining and analysis tools are needed to explore and uncover problems
and opportunities that lead to action and potential manufacturing operation improvements
Big Data
Big Data
Big Data
Big Data
Hadoop
Hadoop
Hadoop
Hadoop
≠
“There’s a belief that if you want big data, you need to go out and buy Hadoop
and then you’re pretty much set. People shouldn’t get ideas about turning off
their relational systems and replacing them with Hadoop…
IBM PoV on Big Data and Analytics architecture
Information Integration & Governance
Systems
Security
On premise, Cloud, As a service
Storage
New/Enhanced
Applications
All Data
What action
should I
take?
Decision
management
Landing,
Exploration
and Archive
data zone
EDW and
data mart
zone
Operational
data zone
Real-time Data Processing & Analytics
What is
happening?
Discovery and
exploration
Why did it
happen?
Reporting and
analysis
What could
happen?
Predictive
analytics and
modeling
Deep
Analytics
data zone
What did
I learn,
what’s best?
Actionable insight
Reporting &
interactive
analysis
Data types Transaction and application data Predictive analytics and modeling Reporting and analysisOperational
systems
ArchiveEnterprise
Warehouse
Staging area
Transformation to target architecture - start
Actionable insight
Reporting &
interactive
analysis
Deep
analytics &
modeling
Data types Transaction and application data Predictive analytics and modeling Reporting and analysisOperational
systems
ArchiveEnterprise
Warehouse
Staging area
Transformation to target architecture – stage1
Actionable insight
Exploration and
landing
Trusted data
Reporting &
interactive
analysis
Deep
analytics &
modeling
Data types Transaction and application data Enterprise content Social data Image and videoThird-party data Predictive analytics and modeling Reporting, analysis, content analytics Discovery and exploration
Operational
systems
ArchiveTransformation to target architecture – stage2
Actionable insight
Exploration,
landing and
archive
Trusted data
Reporting &
interactive
analysis
Deep
analytics &
modeling
Data types Transaction and application data Enterprise content Social data Image and videoThird-party data Predictive analytics and modeling Reporting, analysis, content analytics Discovery and exploration
Operational
systems
ArchiveTransformation to target architecture – stage3
Actionable insight
Exploration,
landing and
archive
Trusted data
Reporting &
interactive
analysis
Deep
analytics &
modeling
Data types
Real-time processing & analytics
Transaction and application data Machine and sensor data Enterprise content Social data Image and video
Third-party data Decision management Predictive analytics and modeling Reporting, analysis, content analytics Discovery and exploration
Operational
systems
Transformation to target architecture – stage4
Information Integration & Governance
Actionable insightExploration,
landing and
archive
Trusted data
Reporting &
interactive
analysis
Deep
analytics &
modeling
Data types
Real-time processing & analytics
Transaction and application data Machine and sensor data Enterprise content Social data Image and video
Third-party data Decision management Predictive analytics and modeling Reporting, analysis, content analytics Discovery and exploration
Operational
systems
Information Integration Data Matching & MDM Security & Privacy Lifecycle Management Metadata & LineageTransformation to target architecture – stage5
Watson Foundations
Watson Foundations
Exploration,
landing and
archive
Trusted data
Reporting & interactive analysisDeep
analytics &
modeling
Data types
Real-time processing & analytics
STREAMS, DATA REPLICATIONTransaction and application data Machine and sensor data Enterprise content Social data Image and video
Third-party data
Operational
systems
BIGINSIGHTS PUREDATA HADOOP DB2, INFORMIX PUREDATA TRANSACTIONS PUREDATA ANALYTICS DB2 BLU PUREDATA ANALYTICS DB2 WAREHOUSE PUREDATA OPERATIONA L ANALYTICS Actionable insight Decision management Predictive analytics and modeling Reporting, analysis, content analytics Discovery and exploration SPSS MODELER COGNOS BI COGNOS TM1 DATA EXPLORER SPSS MODELER GOLDLeverages all data available in fab:
logistics, metrology, inspection, test, tool sensors
Combination of :
1) IBM’s Big Data platform and 2) custom applications
largely developed, built and driven by IBM Research expertise
Equipment Sensor Data
Yield analysis routines
Demand/Supply
Planning
Manufacturing Execution System
(MES)
Sensor
Systems
Equipment
Control
AMHS
Control
Advanced
Process
Controls
Information
Warehouse
&
E-biz interface
Product
Demand
Management
Equipment
Maintenance
And
Scheduling
Recipe
Mgt
Energy
Management
Part Number
Build
Product
Dispatch
Engineering
Analysis
Enterprise
Factory
Several real use cases are described on following pages
Use Case 1: Big Data approach to the problem of large
dataset analysis
Traditional
Tester
Data
Ware
house
Large
dataset
retrieval
Large
analysis
routine
Review
reports
Tester
InfoSphere StreamsInteractive
review
Near real-time analysis
Model results
New approach
Challenge: Existing analysis methods struggle with current data volumes
pulling and manipulating data takes too long
thousands of charts and graphs that require manual review
analysis may not be complete before product is shipped
Partial Least Squares (PLS) model compares actual yield to previous results
analysis output highlights what has changed
Automated Streams solution:
• compares yield by test pattern to historical
data
• identifies unusual yield behavior, based on
multivariate model
• larger bars indicate larger deviation from
historical yield
• has been used to immediately identify
problems on leading edge of new production
• problem identified before the first wafer had
completed testing
• new data added to existing model and kept in
memory for fast and easy analysis
Yield Contribution By Pattern
What we did:
Collected and enabled quick review of massive amounts of sensor data, in a simple dashboard
Identified tool issues and parameters that influence critical product measurements
Developed scoring algorithms, including advanced info theory to highlight relationships
ease of use, guides analyst to significant findings
Fully automated, with linked reports for full drill-down capability
Challenge:
Yield learning is the most direct contributor to fab
profitability and time to market
Huge volume of data (billions of points per day) with
many subtle interpretations
Want to maximize usefulness of semi-structured tool
sensor data for variety of problem solving
Large engineering team, with varying skills in analysis,
statistics, data mining
Challenge
Challenge
Solution
Solution
Quality and supply chain
managers need advanced
techniques to examine
quality date from tens of
thousands of parts
(incoming, manufactured,
deployed) and to provide
better, more proactive
quality management
Software system which uses
proprietary IBM technology to
detect & prioritize quality
problems earlier with fewer false
alarms, coupled with push alert
functionality for IBM & suppliers
to proactively detect & manage
quality issues at any stage of
product lifecycle
Key Innovations
Earlier identification of quality issues
through proprietary analytic techniques
Fewer false alarms
Business Value at IBM
Business Value at IBM
Cost savings – $39M in hard warranty
savings, with additional soft savings and
benefits in other areas
Proactive quality mgt – identify and
resolve issues before they become
problems, up to 6 weeks earlier than
traditional SPC
Improved quality processes – improves
quality process efficiency & effectiveness
Results from QEWS Proof of Concept at external client
GPS