Optical inspection-based automatic defect classification
Implementations disclosed describe, among other things, a systems and techniques for perform efficient inspection of a semiconductor manufacturing sample. The techniques include collecting optical inspection data for training sample(s) that have a plurality of defects. The techniques further include generating, using the optical inspection data, a training data set that includes descriptions, images, and ground truth classifications for the defects. The techniques further include using the training data set to train a plurality of machine learning (ML) classifiers to generate predicted classifications for the defects in the training sample(s). The techniques further include selecting, using the predicted classifications and the ground truth classifications, one or more ML classifiers that meet one or more accuracy criteria, and using the selected ML classifier(s) to classify defects in the semiconductor manufacturing sample.
1 . A method comprising:
generating a training data set comprising:
optical inspection training data collected for one or more training samples, each training sample having one or more known defects, and
one or more high-resolution target outputs, each target output comprising a ground truth defect classification, generated using sub-wavelength imaging of a respective sample of the one or more training samples;
training, using the training data set, a plurality of machine learning (ML) classifiers to predict, based on correlation of the optical inspection training data with corresponding high-resolution target outputs, defect classifications for the one or more training samples;
computing one or more evaluation metrics representative of learned ability of the trained plurality of ML classifiers to predict defect classifications;
selecting, based on the one or more evaluation metrics, an ensemble of ML classifiers from the plurality of ML classifiers; and
causing the ensemble of ML classifiers to be deployed to classify one or more defects in a semiconductor manufacturing sample.
2 . The method of claim 1 , wherein training an individual ML classifier of the ML classifiers comprises processing, using the individual ML classifier, a training input comprising:
optical inspection training data for a training sample, and
a data structure comprising a description of one or more known defects in the training sample, the description of the one or more known defects indexed by coordinates of the one or more known defects.
3 . The method of claim 1 , wherein the sub-wavelength imaging is performed using one or more of:
a scanning electron microscopy system,
an X-ray spectroscopy system,
a tunneling electron microscopy system,
an atomic force microscopy system, or
a neutron scattering system.
4 . The method of claim 2 , wherein the description of an individual known defect in the data structure comprises one or more of:
a signal-to-noise ratio (SNR) associated with the individual known defect,
one or more dimensions of the individual known defect,
a location of the individual known defect,
one or more cross-channel ratios for the individual known defect,
a total light intensity associated with the individual known defect,
an angular distribution of light intensity associated with the individual known defect, or
polarization data associated with the individual known defect.
5 . The method of claim 1 , wherein the optical inspection training data comprises light scattering data associated with light reflected or scattered from the one or more training samples.
6 . The method of claim 5 , wherein the light scattering data is collected for one or more of:
a plurality of scattering angles,
a plurality of scattered polarizations, or
a plurality of wavelengths.
7 . The method of claim 1 , wherein the predicted known classifications comprise one or more of:
a type of an individual defect, or
one or more dimensions of the individual defect.
8 . The method of claim 1 , wherein training the plurality of ML classifiers comprises:
processing, using the plurality of ML classifiers, (i) a first set of feature vectors representative of descriptions of the one or more known defects, and (ii) a second set of feature vectors representative of images of the one or more known defects, wherein the second set of feature vectors is generated using a convolutional neural network.
9 . The method of claim 1 , wherein the ensemble of ML classifiers comprises two or more of:
a decision tree ML classifier,
an adaptive boosting ML classifier,
a boosting ML classifier,
a K-nearest neighbor ML classifier,
a logistic regression ML classifier,
a support vector machine ML classifier,
a linear discriminant analysis classifier, or
a deep neural network ML classifier.
10 . The method of claim 1 , wherein selecting a first ML classifier of the plurality of ML classifiers is responsive to the first ML classifier having at least one evaluation metric of the one or more evaluation metrics that exceeds at least one of (i) a threshold metric, or (ii) a corresponding evaluation metric of a second ML classifier of the plurality of ML classifiers.
11 . The method of claim 1 , further comprising:
causing an individual defect of the one or more defects in the semiconductor manufacturing sample to undergo an additional inspection using a sub-wavelength resolution inspection system.
12 . The method of claim 11 , wherein the individual defect is selected for the additional inspection based on one or more of:
a random selection, or
the individual defect being classified as a target-class defect by at least one ML classifier of the ensemble of ML classifiers.
13 . The method of claim 11 , further comprising:
using an output of the additional inspection to validate the ensemble of ML classifiers.
14 . The method of claim 1 , further comprising:
selecting a processing operation for the semiconductor manufacturing sample in view of the one or more defects in the semiconductor manufacturing sample.
15 . A system comprising:
a memory device; and
a processing device communicatively coupled to the memory device, to:
generate a training data set comprising (i) optical inspection training data collected for one or more training samples, each training sample having one or more known defects, and (ii) one or more high-resolution target outputs, each target output comprising a ground truth defect classification, generated using sub-wavelength imaging of a respective sample of the one or more training samples;
train, using the training data set, a plurality of machine learning (ML) classifiers to predict, based on correlation of the optical inspection training data with corresponding high-resolution target outputs, defect classifications for the one or more training samples;
compute one or more evaluation metrics representative of learned ability of the trained plurality of ML classifiers to predict defect classifications;
select, based on the one or more evaluation metrics, an ensemble of ML classifiers from the plurality of ML classifiers; and
cause the ensemble of ML classifiers to be deployed to classify one or more defects in a semiconductor manufacturing sample.
16 . The system of claim 15 , wherein the optical inspection training data comprises light scattering data associated with light reflected or scattered from the one or more training samples, and wherein the light scattering data is collected for one or more of: a plurality of scattering angles, a plurality of scattered polarizations, or a plurality of wavelengths.
17 . The system of claim 15 , wherein the training data set generated by the system includes a data structure comprising a description of one or more defects, and wherein the description of an individual known defect in the data structure comprises one or more of:
a signal-to-noise ratio (SNR) associated with the individual known defect,
one or more dimensions of the individual known defect,
a location of the individual known defect,
one or more cross-channel ratios for the individual known defect,
a total light intensity associated with the individual known defect,
an angular distribution of light intensity associated with the individual known defect, or
polarization data associated with the individual known defect.
18 . The system of claim 15 , wherein to train the plurality of ML classifiers, the processing device is to:
process, using the plurality of ML classifiers, (i) a first set of feature vectors representative of descriptions of the one or more known defects, and (ii) a second set of feature vectors representative of images of the one or more known defects, wherein the second set of feature vectors is generated using a convolutional neural network.
19 . The system of claim 15 , wherein the ensemble of ML classifiers comprises two or more of: a decision tree ML classifier, an adaptive boosting ML classifier, a boosting ML classifier, a K-nearest neighbor ML classifier, a logistic regression ML classifier, a support vector machine ML classifier, a linear discriminant analysis classifier, or a deep neural network ML classifier.
20 . The system of claim 15 , wherein selecting a first ML classifier of the plurality of ML classifiers is responsive to the first ML classifier having at least one evaluation metric of the one or more evaluation metrics that exceeds at least one of (i) a threshold metric, or (ii) a second corresponding evaluation metric of a second ML classifier of the plurality of ML classifiers.
21 . A non-transitory computer-readable storage medium storing instructions thereon that, when executed by a processing device, cause the processing device to perform operations comprising:
generating a training data set comprising (i) optical inspection training data collected for one or more training samples, each training sample having one or more known defects, and (ii) one or more high-resolution target outputs, each target output comprising a ground truth defect classification, generated using sub-wavelength imaging of a respective sample of the one or more training samples;
training, using the training data set, a plurality of machine learning (ML) classifiers to predict, based on correlation of the optical inspection training data with corresponding high-resolution target outputs, defect classifications for the one or more training samples;
computing one or more evaluation metrics representative of learned ability of the trained plurality of ML classifiers to predict defect classifications;
selecting, based on the one or more evaluation metric an ensemble of ML classifiers from the plurality of ML classifiers; and
causing the ensemble of ML classifiers to be deployed to classify one or more defects in a semiconductor manufacturing sample.