Method for detecting defects in semiconductor device

1. A method for detecting defects in a semiconductor device, the method comprising:

pre-training a pre-trained convolutional neural network model using a sampled clean data set extracted from a first data set;

training a normal convolutional neural network model and a label-noise convolutional neural network model using first data of the first data set and the pre-trained convolutional neural network model;

outputting a first prediction result on whether second data of a second data set is good or bad using the second data and the normal convolutional neural network model;

outputting a second prediction result on whether the second data is good or bad using the second data and the label-noise convolutional neural network model;

comparing the first prediction result with the second prediction result to perform noise correction when there is a label difference;

adding third data created as results of the noise correction to the sampled clean data set; and

additionally training the normal convolutional neural network model and the label-noise convolutional neural network model using the sampled clean data set with the third data added.

2. The method of claim 1 , wherein the noise correction performed when there is a label difference comprises:

determining whether there is a defect in the second data using an attention map created for the label difference, and

correcting a label for the second data when it is determined that there is a defect in the second data.

3. The method of claim 1 , wherein the label-noise convolutional neural network model uses a bootstrapping loss function as a loss function.

4. The method of claim 1 , further comprising:

after additionally training the normal convolutional neural network model and the label-noise convolutional neural network model using the sampled clean data set, outputting a third prediction result on whether the first data is good or bad using the first data and the normal convolutional neural network mode; and

outputting a fourth prediction result on whether the first data is good or bad using the first data and the label-noise convolutional neural network model.

5. The method of claim 4 , further comprising:

comparing the third prediction result with the fourth prediction result to additionally perform the noise correction when there is the label difference; and

adding fourth data created as results of the additionally performed noise correction to the sampled clean data set.

6. The method of claim 1 , further comprising:

generating a visualization of bad data based on a clean data set created after the sampled clean data set has been updated when there is no label difference,

wherein the visualization of the bad data is based on performing clustering on the bad data according to attributes of defects detected from the bad data, and

wherein the visualization of the bad data is further based on creating a defect classification model that classifies the defects of the bad data by the attributes of the defects based on the results of the clustering.

7. The method of claim 1 , further comprising:

generating a visualization of bad data based on a clean data set created after the sampled clean data set has been updated when there is no label difference,

wherein the visualization of the bad data is based on performing clustering on the bad data according to attributes of defects detected from the bad data, and

wherein the performing the clustering on the bad data comprises performing clustering on the bad data using a convolutional neural network auto-encoder.

8. A method for detecting defects in a semiconductor device, the method comprising:

first training a normal convolutional neural network (CNN) model and a label-noise convolutional neural network model using first data of a first data set and a pre-trained convolutional neural network model;

outputting a first prediction result on whether second data of a second data set is good or bad using the second data and the normal convolutional neural network model;

outputting a second prediction result on whether the second data is good or bad using the second data and the label-noise convolutional neural network model;

comparing the first prediction result with the second prediction result to perform noise correction when there is a label difference;

additionally training the pre-trained convolutional neural network model for third data created as results of the noise correction;

second training the normal convolutional neural network model and the label-noise convolutional neural network model using the pre-trained convolutional neural network model that is additionally trained for the third data;

outputting a third prediction result on whether the first data is good or bad using the first data and the normal convolutional neural network model;

outputting a fourth prediction result on whether the first data is good or bad using the first data and the label-noise convolutional neural network model;

comparing the third prediction result with the fourth prediction result to additionally perform the noise correction when there is the label difference; and

additionally training the pre-trained convolutional neural network model for fourth data created as results of the additionally performed noise correction.

9. The method of claim 8 , wherein the noise correction performed when there is a label difference comprises:

determining whether there is a defect in the second data using an attention map created for the label difference for the first prediction result and the second prediction result, and

correcting a label for the second data when it is determined that there is a defect, and

wherein the additionally performing the noise correction comprises:

determining whether there is a defect in the first data using an attention map created for the label difference for the third prediction result and the fourth prediction result, and

correcting a label for the first data when it is determined that there is a defect.

10. The method of claim 8 , further comprising:

generating a visualization of bad data based on a clean data set created after the sampled clean data set has been corrected when there is no label difference,

wherein the visualization of the bad data is based on performing clustering on the bad data according to attributes of defects detected from the bad data, and

wherein the visualization of the bad data is further based on creating a defect classification model that classifies the defects of the bad data by the attributes of the defects based on the results of the clustering.

11. The method of claim 8 , further comprising:

generating a visualization of bad data based on a clean data set created after the sampled clean data set has been corrected when there is no label difference,

wherein the visualization of the bad data is based on performing clustering on the bad data according to attributes of defects detected from the bad data, and

wherein the performing the clustering on the bad data comprises performing clustering on the bad data using a convolutional neural network auto-encoder.

12. A method for detecting defects in a semiconductor device, the method comprising:

pre-training a pre-trained convolutional neural network (CNN) model using a sampled clean data set extracted from a first data set;

training a normal convolutional neural network model and a label-noise convolutional neural network model using first data of the first data set and the pre-trained convolutional neural network model;

outputting a first prediction result on whether second data of a second data set is good or bad using the second data and the normal convolutional neural network model;

outputting a second prediction result on whether the second data is good or bad using the second data and the label-noise convolutional neural network model; and

comparing the first prediction result with the second prediction result to generate a visualization of the bad data based on a clean data set created after a label has been corrected, when there is no label difference.

13. The method of claim 12 , wherein visualization of the bad data visualizes the bad data using an attention map created for the clean data set.

14. The method of claim 12 , wherein visualization of the bad data based on a clean data set is based on performing clustering on the bad data according to attributes of defects detected from the bad data.

15. The method of claim 14 , wherein visualization of the bad data based on a clean data set is further based on creating a defect classification model that classifies the defects of the bad data by the attributes of the defects based on the results of the clustering.

16. The method of claim 14 , wherein the performing the clustering on the bad data comprises performing clustering on the bad data using a convolutional neural network auto-encoder.

17. The method of claim 12 , further comprising:

comparing the first prediction result with the second prediction result;

performing noise correction for the second data when there is a label difference;

adding third data created as results of the noise correction to the sampled clean data set; and

additionally training the normal convolutional neural network model and the label-noise convolutional neural network model using the sampled clean data set.

18. The method of claim 17 , wherein the performing the noise correction comprises:

determining whether there is a defect in the second data using an attention map created for the label difference, and

correcting a label for the second data when it is determined that there is a defect.

19. The method of claim 17 , after additionally training the normal convolutional neural network model and the label-noise convolutional neural network model using the sampled clean data set, further comprising:

outputting a third prediction result on whether the first data is good or bad using the first data and the normal convolutional neural network mode; and

outputting a fourth prediction result on whether the first data is good or bad using the first data and the label-noise convolutional neural network model.

20. The method of claim 19 , further comprising:

comparing the third prediction result with the fourth prediction result to additionally perform the noise correction when there is the label difference; and

adding fourth data created as results of the additionally performed noise correction to the sampled clean data set.