Automatic defect classification without sampling and feature selection

• US 10,650,508 B2
• Filed: 12/01/2015
• Issued: 05/12/2020
• Est. Priority Date: 12/03/2014
• Status: Active Grant

• Current Assignee: KLA-Tencor Corporation
• Original Assignee: KLA-Tencor Corporation
• Inventors: Chang, Wei, Olavarria, Ramon, Rao, Krishna
• Primary Examiner(s): Rudolph, Vincent
• Assistant Examiner(s): Schwartz, Raphael
• Application Number:
  US14/956,326
• Publication Number:
  US 20160163035A1
• Time in Patent Office: 1,624 Days
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06F 18/28 Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06T 2207/20081 Training; Learning
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 2207/30148 Semiconductor; IC; Wafer
• CPC Class Codes: G06T 7/0004 Industrial image inspection
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 2201/06 Recognition of objects for industrial automation
• CPC Class Codes: G06V 30/19173 Classification techniques

Abstract

Systems and methods for defection classification in a semiconductor process are provided. The system includes a communication line configured to receive a defect image of a wafer from the semiconductor process and a deep-architecture neural network in electronic communication with the communication line. The neural network has a first convolution layer of neurons configured to convolve pixels from the defect image with a filter to generate a first feature map. The neural network also includes a first subsampling layer configured to reduce the size and variation of the first feature map. A classifier is provided for determining a defect classification based on the feature map. The system may include more than one convolution layers and/or subsampling layers. A method includes extracting one or more features from a defect image using a deep-architecture neural network, for example a convolutional neural network.

Claims

1. A system for defect classification in a semiconductor process, comprising:

a communication line configured to receive a defect image of a wafer from the semiconductor process;

a deep architecture neural network in electronic communication with the communication line, comprising;

a first convolution layer of neurons, each neuron configured to convolve a corresponding receptive field of pixels from the defect image with a filter to generate a first feature map;

a first subsampling layer configured to reduce the size and variation of the first feature map; and

a classifier for determining a defect classification based on the feature map; and

wherein the system is configured to inject one or more features learned from local descriptors at one or more higher convolution layers of the deep architecture network, wherein the features learned from local descriptors are determined by;

extracting a plurality of local descriptors at each pixel of each of a plurality of defect images, wherein each of the local descriptors is a defect classifier of the defect images, and wherein each of the local descriptors is external to the deep architecture neural network; and

generating the one or more features learned from local descriptors based on the extracted local descriptors.

2. The system of claim 1, wherein the system is configured to replace a pixel in the defect image with a random pixel from a neighboring region.

3. A method for defect classification in a semiconductor process, comprising:

extracting, using a deep-architecture neural network, one or more features from a defect image from the semiconductor process;

injecting one or more features learned from local descriptors at one or more higher layers of the deep-architecture neural network, wherein the features learned from local descriptors are determined by;

extracting, using a processor, a plurality of local descriptors at each pixel of each of a plurality of defect images, wherein each of the local descriptors is a defect classifier of the defect images, and wherein each of the local descriptors is external to the deep architecture neural network; and

generating, using the processor, the one or more features learned from local descriptors based on the extracted local descriptors; and

classifying, using the deep-architecture neural network, the defect image based on the extracted one or more features.

4. The method of claim 3, wherein the one or more features are extracted from the defect image using a convolutional neural network (CNN) having one or more convolutional layers, each convolutional layer of the one or more convolutional layers followed by a subsampling layer.

5. The method of claim 4, wherein extracting one or more features further comprises convolving, at a first convolutional layer of the CNN, each of a plurality of overlapping receptive fields with a filter to generate a first feature map having a plurality of map units, wherein each receptive field comprises a subset of pixels of the defect image.

6. The method of claim 5, wherein extracting one or more features further comprises applying a nonlinearity to the map units of the first feature map.

7. The method of claim 5, wherein extracting one or more features further comprises combining, at a first subsampling layer of the CNN, two or more map units of the first feature map into a unit of a subsampling feature map.

8. The method of claim 7, wherein a size of a unit of the first feature map is reduced by averaging the values of the map units, finding the maximum value of the map units, or adding the map units.

9. The method of claim 8, wherein at each subsequent convolutional layer of the CNN, a subset of map units of the feature map of the preceding layer are convolved with a filter to create a corresponding feature map.

10. The method of claim 9, wherein the defect image is classified using a fully-connected layer of the CNN.

11. The method of claim 4, comprising training the CNN using back propagation to refine the filters of each convolution layer of the CNN.

12. The method of claim 3, wherein the plurality of local descriptors include one or more of a histogram of gradients (HOG), a local binary pattern, and a DAISY descriptor.

13. The method of claim 12, wherein the one or more high-level features are generated by principle component analysis (PCA).

14. The method of claim 3, wherein the defect image is processed using an auto-encoder, a Deep Belief Network (DBN), or a Restricted Boltzmann Machine (RBM).

15. The method of claim 3, wherein the one or more defect images are stored in a centralized data server.

16. The method of claim 15, where defect class codes are generated based on the one or more defect images.

17. The method of claim 16, wherein the defect images are classified based on the defect codes.

18. The method of claim 3, further comprising replacing a pixel in the defect image with a random pixel from a neighboring region.

19. A system for deriving features, comprising:

an electronic storage device;

a feature library stored on the storage device;

a deep-architecture neural network in electronic communication with the storage device, the neural network having one or more convolutional layers, each convolutional layer of the one or more convolutional layers followed by a subsampling layer, the neural network configured to;

derive a feature from one or more defect images of a semiconductor wafer, wherein the feature is for classifying a defect of the defect images;

encapsulate the feature with a set of calculations used to determine the feature;

add the encapsulated feature to the feature library of the storage device; and

inject one or more features learned from local descriptors at the one or more higher convolution layers, wherein the features learned from local descriptors are determined by;

extracting a plurality of local descriptors at each pixel of each of a plurality of defect images, wherein each of the local descriptors is a defect classifier of the defect images, and wherein each of the local descriptors is external to the deep architecture neural network; and

generating the one or more features learned from local descriptors based on the extracted local descriptors.

20. The system of claim 19, wherein the neural network is further configured to replace a pixel in one or more of the defect images with a random pixel from a neighboring region.