現在正在上Standford 的線上課程.
http://www.ml-class.org/
講的滿清楚的
學ml還可以練習英文
一舉兩得啊
2012/02/06
2010/12/02
Defect/Failure Pattern Yield impact model
From the US patent : Patent number: 6367040
System and method for determining yield impact for semiconductor devices
It's estimate the yield impact by two step:
Step 1: calculate each defect's kill probability(the chance for die fail if this die own the defec)
Step 2: calculate the total yield loss according to the above kill probability
We can use the patent's idea for defect or pattern type's yield impact modeling.
First, use logistic regression to get the kill probability for each defect/pattern
Then, use the above kill probability to estimate the yield impact.
For defect type data, we have following data:
Each die's final result(pass or fail), which defect(s) fall in this die.
Summary data format as following:
Die's Result/Defect1/Defect2/.../DefectN
0 0 1 .... 0
1 0 1 1
...
where 0 for pass die or no defect
1 for fail die or own that defect
First we fit the logistic regression with response variable as die's result, explain variables as defect1 ... defectN
Then each defect's kill probability equals the proportion of the odds
For each fail die, we assign each defect to response the yield loss by it's kill probability.
for example, if we have 5 defect types, defect1,..., defect5 with kill probability (0.1,0.1,0.3,0.2,0.3)
if die1 is fail and with defect1 and defect5 located,
then we assign defect1 has yield loss 0.1/0.4 and defect5 has yield loss 0.3/0.4 for this die.
After summary all the fail dies, we can get the estimated yield for each defect type.
For pattern type data, replace 0 or 1 with the true yield loss, and we still have the same result.
The example R core as following
System and method for determining yield impact for semiconductor devices
It's estimate the yield impact by two step:
Step 1: calculate each defect's kill probability(the chance for die fail if this die own the defec)
Step 2: calculate the total yield loss according to the above kill probability
We can use the patent's idea for defect or pattern type's yield impact modeling.
First, use logistic regression to get the kill probability for each defect/pattern
Then, use the above kill probability to estimate the yield impact.
For defect type data, we have following data:
Each die's final result(pass or fail), which defect(s) fall in this die.
Summary data format as following:
Die's Result/Defect1/Defect2/.../DefectN
0 0 1 .... 0
1 0 1 1
...
where 0 for pass die or no defect
1 for fail die or own that defect
First we fit the logistic regression with response variable as die's result, explain variables as defect1 ... defectN
Then each defect's kill probability equals the proportion of the odds
For each fail die, we assign each defect to response the yield loss by it's kill probability.
for example, if we have 5 defect types, defect1,..., defect5 with kill probability (0.1,0.1,0.3,0.2,0.3)
if die1 is fail and with defect1 and defect5 located,
then we assign defect1 has yield loss 0.1/0.4 and defect5 has yield loss 0.3/0.4 for this die.
After summary all the fail dies, we can get the estimated yield for each defect type.
For pattern type data, replace 0 or 1 with the true yield loss, and we still have the same result.
The example R core as following
2010/09/05
2010/02/25
Tips for enhance R code
After reading Writing Efficient Programs in R and R Code optimization and Packages Creation, 3 tips by now.
1. avoid data frame (from 2nd)
2. ifelse is slower than if () { } else { } (from 1st)
3. aovid using rbind, cbind in loops, predifine a NA array or matrix(from 2nd)
Examples:
1. avoid data frame (from 2nd)
2. ifelse is slower than if () { } else { } (from 1st)
3. aovid using rbind, cbind in loops, predifine a NA array or matrix(from 2nd)
Examples:
2009/11/09
EDA 常用的統計方法
先整理一下 有空再詳細寫
1.基本統計量 mean/std/min/max/percentile/
2.distribution verify and display
3.testing hypotheses
4.DOE
5.regression
6.SPC
進階分析
1.MVA(PCA/FDA)
2.Data mining
問題分類
1.Process control
2.Root cause analysis
3.Yield impact
4.Yield prediction
5.Wafer map analysis
1.基本統計量 mean/std/min/max/percentile/
2.distribution verify and display
3.testing hypotheses
4.DOE
5.regression
6.SPC
進階分析
1.MVA(PCA/FDA)
2.Data mining
問題分類
1.Process control
2.Root cause analysis
3.Yield impact
4.Yield prediction
5.Wafer map analysis
2009/07/02
訂閱：
文章 (Atom)
