Sampling Plan for Wafer Inspection (Small Lot)

[Reyov]

Hi,
Newbie on this forum and in QA…

I have been asked to create a sample plan for the inspection of wafers we buy from our supplier.
We buy very small lots of wafers (6 – 8). Each wafer has 42 dies.
From past experience, the yield is about 50% good dies (electrically) on each wafer.

In the past we used to test ~30% of the dies on each wafer to estimate the yield. Each die is either GOOD or BAD (attribute response).

Assuming that we want to reject a wafer if the defect rate is > 60%, how many dies should we test on each wafer?

I thought of using a AQL of 40% and RQL of 60% with alpha and Beta = 0.05.
With Minitab, I calculated that we need to test 67 dies and accept the lot if the number of defective dies is < 33. However, I have 50% chance to reject a wafer if the real yield is 50%...

Is it a good approach?

Or... how can I determine the sample size if I want to make sure that the % of defective dies on a wafer is < 60% ? with 95% confidence.

Any suggestions welcomed.
Thanks.

 

[Mikishots]

Hi,
Newbie on this forum and in QA…

I have been asked to create a sample plan for the inspection of wafers we buy from our supplier.
We buy very small lots of wafers (6 – 8). Each wafer has 42 dies.
From past experience, the yield is about 50% good dies (electrically) on each wafer.

In the past we used to test ~30% of the dies on each wafer to estimate the yield. Each die is either GOOD or BAD (attribute response).

Assuming that we want to reject a wafer if the defect rate is > 60%, how many dies should we test on each wafer?

I thought of using a AQL of 40% and RQL of 60% with alpha and Beta = 0.05.
With Minitab, I calculated that we need to test 67 dies and accept the lot if the number of defective dies is < 33. However, I have 50% chance to reject a wafer if the real yield is 50%...

Is it a good approach?

Or... how can I determine the sample size if I want to make sure that the % of defective dies on a wafer is < 60% ? with 95% confidence.

Any suggestions welcomed.
Thanks.

A quick question: how long does it take to test each die? If the attribute response is GOOD or BAD, the test is not time-consuming and especially if the yield rate is at 50% - is it unreasonable to perform a 100% inspection? With a yield rate like that in such small batches, sampling won't tell you much if you're trying to identify usable product.

 

[Reyov]

Hi Mikishots,

To answer your question : the test is done manually and take approximately 15 minutes / die. (production volume are too small to justify an automated testing equipement for this step).

So ~10 hrs / wafer; 80 hrs / batch.

We would like to minimize inspection step but at the same time, ensure that the wafers provided by the supplier have at least 50% good dies.

 

[Barbara B]

We buy very small lots of wafers (6 – 8). Each wafer has 42 dies. From past experience, the yield is about 50% good dies (electrically) on each wafer.

In the past we used to test ~30% of the dies on each wafer to estimate the yield. Each die is either GOOD or BAD (attribute response).

Assuming that we want to reject a wafer if the defect rate is > 60%, how many dies should we test on each wafer?

I thought of using a AQL of 40% and RQL of 60% with alpha and Beta = 0.05.
With Minitab, I calculated that we need to test 67 dies and accept the lot if the number of defective dies is < 33. However, I have 50% chance to reject a wafer if the real yield is 50%...

Is it a good approach?

Or... how can I determine the sample size if I want to make sure that the % of defective dies on a wafer is < 60% ? with 95% confidence.

If you change the aim of your inspection a little bit from

• from "How many dies should be tested on each wafer?"
• to "How many wafers should be tested?"

you can work with the number of defective dies on a wafer instead of the nonconformance rate (poisson distribution instead of binomial). This approach avoids also the struggle with die selection if you use the number of dies.

• Binomial approach / nonconformance rate: dies should be selected randomly out of the whole lot of wafers.
• Poisson approach / no. of defects per unit (=wafer): Only the appropriate number of wafers have to be selected randomly and each die on the selected wafer is tested.

But for your small lot size there doesn't exist a plan which has a maximum of 5% alpha and 5% beta (95% confidence / 95% reliability). If you increase the allowed alpha and beta risks (e.g. to 20%) you can take a sample of 5 wafers with alpha=15.3% and beta=19.9% (see attachments). To have a higher confidence/reliability, the number of sample has to be bigger (e.g. 16 wafers). And if you work with a smaller sample (e.g. 1 wafer) alpha and beta increase (e.g. to alpha=28.4% and beta=39.4%) - smaller than in your plan, but still too high to assure a good die quality.

Hope this helps

Barbara

 

[Reyov]

Hi Barbara,
Thanks for the suggestion. I like the idea of using the Poisson approach even though it means that we need to test more dies. It sure seems to be a more "solid" approach then what we did in the past.