PMML, Predictive Model Markup Language, Prognositcs, Use of Big Data in Manufacturing, Basic Architecture, Holonics, Agent Based Control, Advanced Process Control
Axa Assurance Maroc - Insurer Innovation Award 2024
2016 How to make big data productive in semicon manufacturing
1. HOW TO MAKE BIG DATA
PRODUCTIVE IN A SEMICON
MANUFACTURING ENVIRONMENT
Basic Architectures and Guidelines | Jan Eite Bullema
2. ENIAC INTEGRATE PROJECT
2 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Integrated Solutions for Agile Manufacturing in
High mix Semiconductor Fabs
Overall goal: Models, architecture and metrics for
agile manufacturing
Scope: From equipment to wafer to full fab
process control
Duration: Start Jan 2013 – End March 2016
Project 28 partners, total budget 35 M EURO
3. CHARACTERISTICS OF SEMICON MANUFACTURING
3 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Fan-Tien Cheng, Configuring AVM as a MES Component, IEEE 2010
Semiconductor is the state-of-the-art for manufacturing engineering
Heavily relies on Statistical Process Control (SPC) for quality monitoring
65 nm process has 36+ layers, 500+ operation steps
Long process times 50+ days of production time
Complex process routing
More than 14.000 SPC charts to monitor process status
4. BIG DATA IN SEMICON MANUFACTURING
OUT IN FRONT: EXPLOSION IN GENERATED DATA
4 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
http://electroiq.com/blog/2014/09/the-semiconductor-industry-out-in-front-but-lagging-behind/
2004
1 FAB produces
1 Terabyte/day
2014
1 FAB produces
1 Petabyte/day
2020
1 FAB
multiple
Petabytes/day
5. BIG DATA IN SEMICON MANUFACTURING
BUT LAGGING BEHIND: NO BIG DATA TOOLS USED
5 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
http://electroiq.com/blog/2014/09/the-semiconductor-industry-out-in-front-but-lagging-behind/
6. PREDICTIVE MODELS: IMPROVED LOGISTICS
6 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
In high mix model based dispatching can lead to 20% increase of throughput
"Lot sequence optimization for the wet bench … “ Bas de Kruif, APC|M 2015
7. PREDICTIVE: IMPROVED ASSETT UTILIZATION
7 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Charlie Pappis, Optimizing Fab Through Data Analytics, Nano chip Fab Solutions, V9/ Issue 2/2014
Case: Deposition Tool
Applied Materials claims that a customer
increased profits > $100,000 per chamber, per
year using predictive maintenance modelling
8. ITRS ROADMAP FOR FACTORY INTEGRATION
9 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
9. IMPLEMENTATION: EXISTING SEMI STANDARDS
24 June 201610 | How to Make Big Data Productive in Semicon Manufacturing?
James Moyne, SEMI E133, The Process Control System Standard, IEEE 2007
10. IMPLEMENTATION BARRIER: LEGACY FABS
11 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
https://objectpartners.com/2013/06/14/
12. BARRIER: A PLETHORA OF LEARNING TOOLS
13 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
http://www.kdnuggets.com//
13. IMPLEMENTATION BARRIER: COMMUNICATION GAP
14 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Data Scientist
Semicon
Professionals
IT
Professionals
Build
DeployUse
14. HOW TO BREAK DOWN THESE BARRIERS
15 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“A PMML based Data Model for Fab Wide Implementation of APC”, Jan Eite Bullema, APC 2015
15. PMML: PREDICTIVE MODEL MARKUP LANGUAGE
16 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“Enabling Architectures for Virtual Metrology”, Jan Eite Bullema, IIRC 2013
16. PMML IS XML BASED
17 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“Enabling Architectures for Virtual Metrology”, Jan Eite Bullema, IIRC 2013
Header
Data Description
Model
17. PMML SUPPORTS SEVERAL MODELS
18 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“Enabling Architectures for Virtual Metrology”, Jan Eite Bullema, IIRC 2013
• Association Rules
• Cluster Models
• Decision Trees
• Naïve Bayes Classifiers
• Neural Networks
• Regression
• Rule sets
• Sequences
• Support Vector Machines
• Text Models
• Time Series
• Neural Networks
18. BRIDGING THE COMMUNICATION GAP
19 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Predictive Models
- Data Mining
- Deep Learning
- SPSS, R, SAS
Real time scoring
Batch scoring
ADAPA:CLOUD
UPPI: Hadoop
XML Scoring
Build
DeployUse
Data Scientist
Semicon
Professionals IT Professionals
19. FASTER MODEL DEPLOYMENT
20 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Zementis, Predictive Analytics at Scale, Company Presentation, 2014
20. IMPLEMENTATION: INTEGRATION IN SEMI STANDARDS
21 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“A PMML based Data Model for Fab Wide Implementation of APC”, Jan Eite Bullema, SEMI Europe 2015
21. BUILDING AND IMPLEMENTING PREDICTIVE MODELS
22 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Raw DATA
Filter and
Pre-process
DATA
Build Model
R, Knime,
SPSS, etc.
Use in
Production
Upload
PMML model
Generate
PMML
22. AUTOMATED GENERATION OF PMML FROM DATA
23 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
“A PMML based Data Model for Fab Wide Implementation of APC”, Jan Eite Bullema, APC 2015
23. AGENT BASED CONTROL
A STORY OF TWO WATCHMAKERS
24 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Herbert Simon, The Architecture of Complexity, Proceedings of the American philosophical society, vol 106, no. 6., December, 1962
24. AGENT BASED CONTROL
HOLONIC MANUFACTURING SYSTEMS
25 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
Mihaela Ulieru, The Holonic Enterprise, 2005
25. AGENT BASED CONTROL
HOLONIC MANUFACTURING SYSTEMS
Start with simple white box models or existing decision rules
Growing to more advanced models
26 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
26. AGENT BASED CONTROL
STATISTICAL PROCESS CONTROL HOLON
24 June 201627 | How to Make Big Data Productive in Semicon Manufacturing?
Decision: In Control or Out of Control
27. 24 June 201628 | How to Make Big Data Productive in Semicon Manufacturing?
In Control or Out of Control
PMML Model for
Statistical Control Chart
Decision
28. 24 June 201629 | How to Make Big Data Productive in Semicon Manufacturing?
PMML Model for
Statistical Control Chart
Visual Manufacturing
Signal Light
29. 24 June 201630 | How to Make Big Data Productive in Semicon Manufacturing?
Remaining Useful Life
PMML Model for
Artificial Neural Network
Prediction
30. 24 June 201631 | How to Make Big Data Productive in Semicon Manufacturing?
Dispatch or No Dispatch
PMML Model for
Decision Tree
Decision
31. 24 June 201632 | How to Make Big Data Productive in Semicon Manufacturing?
In Control
RUL > 4 hr
Dispatch OK
32. 24 June 201633 | How to Make Big Data Productive in Semicon Manufacturing?
XML
Orders / Job List
33. 24 June 201634 | How to Make Big Data Productive in Semicon Manufacturing?
Dispatching Sequence
PMML Model for
Discrete Event Simulation
Decision
34. 24 June 201635 | How to Make Big Data Productive in Semicon Manufacturing?
Orders
Dispatch OK
Dispatch Sequence
35. 24 June 201636 | How to Make Big Data Productive in Semicon Manufacturing?
Product OK or Product Defect
PMML Model for
Product Classification
Decision
36. 24 June 201637 | How to Make Big Data Productive in Semicon Manufacturing?
In Control
Product Defect
Adjust Setting:
Reduce Platform Speed
with 10%
37. AGENT BASED CONTROL
PREDICTIVE MAINTENANCE HOLON
24 June 201638 | How to Make Big Data Productive in Semicon Manufacturing?
Prediction : Remaining Useful Life
38. AGENT BASED CONTROL
BUILDING A HOLARCHY
24 June 201639 | How to Make Big Data Productive in Semicon Manufacturing?
39. TRAINING PHASE
40 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
gregmakowski/kamanja-driving-business-value-through-realtime-decisioning-solutions 2015
Data Mining
Tool
Training & Test
Data
(Batch)
PMML
Producer
PMML File
Full Model
Specification
40. Scoring Engine
e.g. Kamanja
OutputScoring Data
PRODUCTION SCORING
41 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
gregmakowski/kamanja-driving-business-value-through-realtime-decisioning-solutions 2015
Scoring Data PMML
Consumer
Control Actions
Decisions
Planning
41. WHY IS PMML ATTRACTIVE AS THE BASIC
ELEMENT IN A HOLARCHY
Industry Standard – for over 15 years
Supported by several (~18) PMML producers and (~12) PMML Consumers
Developments towards Big Data implementations
- Kamanja => run several hundreds models simultaneously for real time scoring
- Spark support for PMML => score Petabytes of Data
After set-up – Maintenance and expanding a PMML based holarchy could be
run at the costs of running the traditional SPC
42 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016
42. AGENT BASED CONTROL
IS THE ONLY WAY
TO MAKE BIG DATA PRODUCTIVE
IN A COMPLEX MANUFACTURING ENVIRONMENT
AT AFFORDABLE COSTS
43 | How to Make Big Data Productive in Semicon Manufacturing? 24 June 2016