Data Modeling

16
Preliminary Early, we may reduce the impact scope, may not predict accurately. There are different performance between distinct tools/products (Quality of output,frequency of maintenance, PMconditions, designs…) Find the key factors via models (process parameters,variation of materials, Critical dimensions…) Predict the variation before events and control the parameters to reduce the variation or compensate for the imperfections to reach customers’ specifications. Unhealthy tool with specific factor Poor quality What are the unhealthy tools and the specific factors? Can we control the factors? 2015/2/5 1 Yun-Hsuan Yeh

Transcript of Data Modeling

Page 1: Data Modeling

Preliminary

• Early, we may reduce the impact scope, may not predict accurately.

• There are different performance between distinct tools/products (Quality • There are different performance between distinct tools/products (Quality

of output, frequency of maintenance, PM conditions, designs…)

• Find the key factors via models (process parameters, variation of materials,

Critical dimensions…)

• Predict the variation before events and control the parameters to reduce

the variation or compensate for the imperfections to reach customers’

specifications.

Unhealthy

tool with

specific

factor

Poor quality

•What are the unhealthy

tools and the specific

factors?

•Can we control the

factors?

2015/2/5 1Yun-Hsuan Yeh

Page 2: Data Modeling

Data Modeling

Definition:

•Yield improve

•Cost down

Efficiency & Accuracy:

•R^2

•Number of correct •Cost down

Observed data:

•Objects

•Indexes

•Number of correct

•Confidence interval

Review:

•Samples

•Variables

•Fitness of model

Hypotheses:

•Regression

•Classification

•Clustering

Implementation:

•Real-time insights

•Improve operations

•Innovations

2015/2/5 2Yun-Hsuan Yeh

Page 3: Data Modeling

Improved by Learning Process

New

samples

Model

Re-Fit with

Keep the samples

Add variable or

Change algorithm

Accuracy do not

change or

become better

Re-Fit with

new sample

Review the

samples

Change algorithm

Accuracy is worse.Exclude the sample

effective

ineffective

2015/2/5 3Yun-Hsuan Yeh

Page 4: Data Modeling

Model of classification

Index 1 Index 2 Index 3 State …

1 1 4.3% 123 A …

2 1.23 3.3% 113 B …

•Separates the classes

•Prediction is made by plugging in

observed values of the attributes 2 1.23 3.3% 113 B …

3 1.56 6% 156 B …

State B

State

State A

observed values of the attributes

into the expression.

Index 1

Index 2

Index 3

Good

Bad

BadGood

≥1.33<1.33

≥120 <120

<10%

State B

≥10%

Classification tree

Bad

We can define the

health of tools

(Good/Bad) by PM

conditions, Cpk,

LRR, GRR,…

To fit the line2015/2/5 4Yun-Hsuan Yeh

Page 5: Data Modeling

Products / Customers

Product features /

Defects / Services

data

Fit model

Figure out the potential factors

that cause defects / customer

complains

2015/2/5 Yun-Hsuan Yeh 5

Improve imperfect design of the

product or Service method.

Page 6: Data Modeling

• Xs: Condition of Materials, Health of tools (Monitor items,

PM conditions, Records of maintenance, Cpk…) ,…

Correlations between process parameters,

CTQs, and Yield

PM conditions, Records of maintenance, Cpk…) ,…

• Ys: CTQs, UPH, Cost,…

• Zs: Yield, Reliability, Customer complaints ,…

Material

condition

Parameter

1

Parameter

2

Variations CTQs

Fit

Model

Fit

Model

Reliability Yield

2015/2/5 6Yun-Hsuan Yeh

Page 7: Data Modeling

Controlled and Uncontrolled factors

• There are many failed DOEs, since the user did not consider the uncontrolled factors.

• It collects all potential factors (uncontrolled factors) for big data analysis (Weather forecast).

Variation 1

Defect 1

Quality

index

Pa

ram

ete

r 2

Controlled factors Uncontrolled factors

Variation 2

Variation 3

Defect 2

Parameter 1

0.60.91.6

Pa

ram

ete

r 2

After low yield occurring, we can map out

the impact processes fast, and reduce the

time for troubleshooting or compensate for

the imperfections .

The factors (parameters) can be controlled, we

may control them to optimize the UPH or cost.2015/2/5 7Yun-Hsuan Yeh

Page 8: Data Modeling

Real case

2015/2/5 8Yun-Hsuan Yeh

Page 9: Data Modeling

Define Target and Assumptions

�We want to reduce the primary defect in FA monitor.

�Skip the samples that suffered dicing shift.

�Since the crack cannot be quantified, we apply the logistic

regression.

�Use binary quadratic equation to fit the model

(the variables may have nonlinear relationship).

2015/2/5 9Yun-Hsuan Yeh

Page 10: Data Modeling

Observed data and Hypotheses

�Since Length and Depth (potential factors ) have significance by

parameter estimates, we take that as variables.

�Determines the impact of multiple independent variables presented �Determines the impact of multiple independent variables presented

simultaneously predict correlation between variables and crack ratio.

Fail rate

Logistic Depth

Length

Fail rate

Logistic

function

Profile statement2015/2/5 10Yun-Hsuan Yeh

Page 11: Data Modeling

Efficiency & Accuracy

( )P

eCE

−+=1

1025.14019.0055.0113.0602.115.0

22

+−−+−= LDDLLDP, where

�R^2=0.77, p-value<0.05, and predict with 72.6% accuracy.

Depth

e+1

Length

2015/2/5 11Yun-Hsuan Yeh

Page 12: Data Modeling

�Polar coordinate.

Θ

r

Θπ/180

< 36.21°

> 55.52°

r

Θπ/180

2015/2/5 12Yun-Hsuan Yeh

Page 13: Data Modeling

Review and Other Analysis

G0308 G0309

Mean 46.9 41.5

�The Angles are different between Produce G0308 and G0309

to show that G0309 is better G0308.

Mean 46.9° 41.5°

STD 5° 5.54°

Crack rate 64%(60/94) 43%(13/30)

P-value=0.0001166

�When the products were reworked over 2 times for LC, they

suffered serious Ni finger, but there is no crack issue. Thus,

Θ

r

suffered serious Ni finger, but there is no crack issue. Thus,

shows that Ni finger and crack is not linear relationship.

2015/2/5 13Yun-Hsuan Yeh

Page 14: Data Modeling

Suggestions

Project CAvailabilit

Depth

Project CProject Risk Cost

Availabilit

y

A High MiddleNew

structure

B Low High

New tool

and

accessorie

s

Length

Project A

Project B

s

C Middle LowTuning

recipes

2015/2/5 14Yun-Hsuan Yeh

Page 15: Data Modeling

Conclusion

�Correlation between SMDG crack and Ni finger is nonlinear.

�The model is not closed-form solution, we only provide �The model is not closed-form solution, we only provide

phenomenon and trend (do not consider all variations).

� Implement project C and verify by trial run under limit

resources.

2015/2/5 15Yun-Hsuan Yeh

Page 16: Data Modeling

Implementation

�Improve the defect ratio, yield, and reliability base on the logistic regression.

30

40

50

60

70

80

90

100

1

1.5

2

2.5

3

0

10

20

30

70um (STD) 70um + LTR 50um Deeper Notch Depth

0

0.5

1

SMDG crack %(T0) T-open%(T0) Related function fail %

2015/2/5 16Yun-Hsuan Yeh