DETAILED ACTION
This action is in response to Applicant’s Request for Continued Examination ("Response”) received on June 20, 2025 in response to the Office Action dated April 3, 2025. This action is made Non-Final.
Claims 1-29 are pending.
Claims 1, 11, 21, and 29 are independent claims.
Claims 1-29 are rejected.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Applicant’s Response
In Applicant’s Response, Applicant amended claims 1, 6, 11, 16, 21, 22, 24, 28, and 29, and submitted arguments against the prior art in the Office Action dated April 3, 2025.
Information Disclosure Statement
The information disclosure statement (IDS(s)) submitted on 08/28/2025, 08/29/2025, 09/05/2025, and 01/29/2026 is/are in compliance with the provisions of 37 C.F.R. 1.97. Accordingly, the IDS(s) is/are being considered by the examiner.
Claim Interpretation
The Specification recites “computer-readable storage media 824, as used herein, is not to be construed as being transitory signals per se.” Accordingly, Claim 29 is interpreted as statutory under §101 CRM analysis.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1, 6, 9-14, 20-22 and 28 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, Mohammad Hadi, et al. "Neural network for nanoscience scanning electron microscope image recognition." Scientific reports 7.1 (2017): 13282, (“Modarres”), in view of Kooiman et al., US Publication 2022/0342316 (“Kooiman”), and further in view of Savalia et al., US Patent 11,126,837 (“Savalia”).
Claim 1:
Modarres discloses a device screening system that comprises:
a computer system that comprises (see p. 2 - Training a neural network on SEM images; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified.);
a feature machine learning model ... configured to: receive images of quantum devices ... of quantum devices, identify features for the quantum devices in the images ..., and output the features (see p. 1 - extracting features from different types of microscope images; p. 2 - potential for feature extraction. feature extraction technique enables the application of learned features of a network. focused on the feature extraction technique., by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images (a) Feature extraction from an Inception-v3 model pre-trained on the ImageNet 2012 dataset, by retraining the last layers (softmax + fully connected layer) on the SEM dataset; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 -classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images.);
a classification neural network configured to receive, identify a set of characteristics for the quantum devices from a group of mutually exclusive characteristics based on the features, wherein the set of characteristics indicates whether the quantum devices will function, and output, the quantum device, the set of characteristics; and of the quantum device (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Modarres fails to explicitly disclose that comprises a 14-block structure; in different stages of fabrication of a wafer that comprises thousands; from the different stages of fabrication; grouped using global pooling into a tensor; that comprises three layers; receive the tensor; with a desired level of consistency; prior to completing the fabricating the device; and a screening manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the device.
Kooiman teaches or suggests that comprises a 14-block structure; in different stages of fabrication of a wafer that comprises thousands; from the different stages of fabrication; that comprises three layers; with a desired level of consistency; prior to completing the fabricating the device; and a screening manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the device (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0004 - determining of defectiveness is based on comparing the given feature in the after development image with a corresponding etch feature in the after etch image; para. 0102 - (i) an after development image 401 of the imaged substrate at a given location, the after development image including a plurality of features, and an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after-development image to classify whether the desired pattern will be defective after etching; para. 0139 - present disclosure is not limited to after development and after etch; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include that comprises a 14-block structure; in different stages of fabrication of a wafer that comprises thousands; from the different stages of fabrication; that comprises three layers; with a desired level of consistency; prior to completing the fabricating the device; and a screening manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the device for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Savalia teaches or suggests grouped using global pooling into a tensor; receive the tensor (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include grouped using global pooling into a tensor; receive the tensor for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Claim(s) 21:
Claim(s) 21 correspond to Claim 1, and thus, Modarres, Kooiman, and Savalia teach or suggest the limitations of claim(s) 21 as well.
Claim 6:
Modarres further discloses wherein the feature machine learning model system is a convolutional neural network trained to extract features for quantum devices (see p. 2 - lower layers of convolutional neural networks capture low-level image features, e.g. edges, while higher layers capture more complex details. we mainly focused on the feature extraction technique, by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images; p. 10 - presented a complete approach to automatically classify nanoscience SEM images by means of transfer learning technique and in particular feature extraction on deep convolutional neural networks.).
Kooiman further teaches or suggests and the undesired issue comprises a presence of a particle or missing gate (see para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include and the undesired issue comprises a presence of a particle or missing gate for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim 9:
Modarres further discloses wherein the quantum devices are selected from one of: a quantum dot device, a nanowire device, a quantum well device, a quantum information processing device, a quantum memory, a superconducting resonator, a Josephson junction, a quantum interference device, a topological quantum device, a waveguide, or an optical resonator (see Fig. 1, 2, 3, and 8; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Claim 10:
Modarres further discloses wherein the images of the quantum devices are generated during a development of a quantum device (see Fig. 1-33; p. 3 - 1D objects are often referred to as nanowires, rods, or fibres. These structures are often packed together, to form bundles, for instance, or aligned parallel to each another and along a specific direction (i.e., growth direction). 2D structures refer to films and coatings on a surface, which can be formed from a variety of different materials with a range of surface topologies: some surfaces can seem smooth and flat on a SEM, whilst others are made of small particles packed together, covering the entire surface. 3D structures can refer to pillars, or other devices like Micro Electro-Mechanical Systems (MEMS), typically fabricated using lithographic processes; p. 6 - MEMS devices are usually fabricated using the same technique as patterned surfaces; p. 7 - MEMS electrodes had nanowires dispersed over them. structures made from top down processes such as lithography.).
Kooiman f more specifically teaches or suggests during a development of a resist on a device (see para. 0088 - resist layer on a substrate is exposed by the aerial image and the aerial image is transferred to the resist layer as a latent "resist image" (RI) therein. The resist image (RI) can be defined as a spatial distribution of solubility of the resist in the resist layer. A resist image 1250 can be simulated from the aerial image 1230 using a resist model 1240. The resist model can be used to calculate the resist image; para. 0139 - principle described herein works with any etch and combination of layers (e.g., a first resist, a second resist layer, etc.) of the substrate being patterned; para. 0220 – person skilled in art can perform the above methods using any images obtained before and after a particular process (e.g., OPC, optical process, resist process, etching, chemical mechanical polishing, etc.) or a combination of processes related to the patterning process. The model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0225 - image classified as defective after etching includes at least one of: a closed hole or a missing hole after etching due to resist blocking; para. 0249 - determine whether extra filtering step for a resist layer should be performed to reduce the failure rate; para. 0261 – produce a first image of the ADI feature, the ADI feature being a structure within a resist material; para. 0299 – resist screening, extra filtering steps for the resist; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0363 - when a SEM is used to measure the feature shape in the resist; para. 0376 - parameters associated with a resist process or the etch process; para. 0396 - whether, e.g., the structures are in a latent resist image, in a developed resist image.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include during a development of a resist on a device for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim 11:
Modarres teaches or suggests a device screening system that comprises a computer system that comprises:
a classification machine learning model system in the computer system, wherein the classification machine learning model system is configured to:
receive features for quantum devices identified in images of the quantum devices ... of quantum devices (see p. 1 - extracting features from different types of microscope images; p. 2 - potential for feature extraction. feature extraction technique enables the application of learned features of a network. focused on the feature extraction technique., by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images (a) Feature extraction from an Inception-v3 model pre-trained on the ImageNet 2012 dataset, by retraining the last layers (softmax + fully connected layer) on the SEM dataset; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 -classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images.);
identify a set of characteristics for the quantum devices (see p. 1 - extracting features from different types of microscope images; p. 2 - potential for feature extraction. feature extraction technique enables the application of learned features of a network. focused on the feature extraction technique., by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images (a) Feature extraction from an Inception-v3 model pre-trained on the ImageNet 2012 dataset, by retraining the last layers (softmax + fully connected layer) on the SEM dataset; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 -classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images.);
output the set of characteristics identified for the quantum devices based on the features identified in the images of the quantum devices, wherein the set of characteristics indicates whether the quantum devices will function (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Modarres does not explicitly disclose formatted as a 1024-dimension tensor; from different stages of fabrication of a wafer that comprises thousands; in a three layer classification neural network; in the different stages of fabrication; prior to completed fabrication of the quantum device; with a desired level of consistency; and a screen manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the quantum device.
Savalia further teaches or suggests formatted as a 1024-dimension tensor (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include formatted as a 1024-dimension tensor for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Kooiman further teaches or suggests from different stages of fabrication of a wafer that comprises thousands; in a three layer classification neural network; in the different stages of fabrication; prior to completed fabrication of the quantum device; with a desired level of consistency; and a screen manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the quantum device (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0004 - determining of defectiveness is based on comparing the given feature in the after development image with a corresponding etch feature in the after etch image; para. 0102 - (i) an after development image 401 of the imaged substrate at a given location, the after development image including a plurality of features, and an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after-development image to classify whether the desired pattern will be defective after etching; para. 0139 - present disclosure is not limited to after development and after etch; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include from different stages of fabrication of a wafer that comprises thousands; in a three layer classification neural network; in the different stages of fabrication; prior to completed fabrication of the quantum device; with a desired level of consistency; and a screen manager configured to, based upon the set of characteristics: determine, responsive to the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and initiate a release or a change to a fabrication process, of the quantum device for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim 12:
Modarres further teaches or suggests the classification machine learning model system is a classification neural network configured to identify the set of characteristics for the quantum devices from a group of mutually exclusive characteristics (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Savalia further teaches or suggests each layer of the three layers comprises 1024 neurons (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include each layer of the three layers comprises 1024 neurons for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Claim 13:
Modarres further teaches or suggests the classification machine learning system is configured to identify the set of characteristics for the quantum devices from a group of mutually exclusive characteristics (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Kooiman further teaches or suggests the change comprises a change to at least one of: a lithography, an etching time, or a dopant concentration (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0102 - an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after development image to classify whether the desired pattern will be defective after etching; para. 0139 - present disclosure is not limited to after development and after etch; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the change comprises a change to at least one of: a lithography, an etching time, or a dopant concentration for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim 14:
Modarres further discloses wherein the classification machine learning model system is selected from at least one of a neural network, a decision tree, a support vector machine, a Bayesion network, a genetic algorithm, or a cluster analysis algorithm (see p. 2 - lower layers of convolutional neural networks capture low-level image features, e.g. edges, while higher layers capture more complex details. we mainly focused on the feature extraction technique, by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images; p. 10 - presented a complete approach to automatically classify nanoscience SEM images by means of transfer learning technique and in particular feature extraction on deep convolutional neural networks.).
Claim 20:
Modarres further discloses wherein the quantum devices are selected from one of: a quantum dot device, a nanowire device, a quantum well device, a quantum information processing device, a quantum memory, a superconducting resonator, a Josephson junction, a quantum interference device, a topological quantum device, a waveguide, or an optical resonator (see Fig. 1, 2, 3, and 8; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Claim 22:
As indicated above, Modarres teaches or suggests the quantum devices; the quantum device.
Kooiman further teaches or suggests the changing the fabrication process increasing a yield; modifying a process for fabricating the device based on the set of characteristics identified by the classification neural network (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0072 - advantage of such correlation-based process control will effectively be used to reduce defects after etching thereby improving the yield of the patterning process; para. 0102 - an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after development image to classify whether the desired pattern will be defective after etching; para. 0122 - modified etch conditions 905 can be further used to etch the imaged substrate thereby improving the yield (e.g., reduced failure of features/structures on the substrate) of the patterning process; para. 0139 - present disclosure is not limited to after development and after etch; para. 0198 - interpretation map or the pixel values therein can be further used to take actions such as adjusting a patterning process recipe (e.g., etch recipe) to improve yield of the patterning process; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the changing the fabrication process increasing a yield; modifying a process for fabricating the device based on the set of characteristics identified by the classification neural network for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Savalia further teaches or suggests the tensor comprising a 1024-dimension; each layer of the three layers comprising 1024 neurons (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include suggests the tensor comprising a 1024-dimension; each layer of the three layers comprising 1024 neurons for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Claim 28:
Modares further teaches or suggests wherien: and the image of the quantum device is generated ... on the quantum device (see Fig. 1-33; p. 3 - 1D objects are often referred to as nanowires, rods, or fibres. These structures are often packed together, to form bundles, for instance, or aligned parallel to each another and along a specific direction (i.e., growth direction). 2D structures refer to films and coatings on a surface, which can be formed from a variety of different materials with a range of surface topologies: some surfaces can seem smooth and flat on a SEM, whilst others are made of small particles packed together, covering the entire surface. 3D structures can refer to pillars, or other devices like Micro Electro-Mechanical Systems (MEMS), typically fabricated using lithographic processes; p. 6 - MEMS devices are usually fabricated using the same technique as patterned surfaces; p. 7 - MEMS electrodes had nanowires dispersed over them. structures made from top down processes such as lithography.).
Kooiman f more specifically teaches or suggests the undesired issue comprises a presence of a particle or a missing gate; and ... during a development of a resist on a device (see para. 0088 - resist layer on a substrate is exposed by the aerial image and the aerial image is transferred to the resist layer as a latent "resist image" (RI) therein. The resist image (RI) can be defined as a spatial distribution of solubility of the resist in the resist layer. A resist image 1250 can be simulated from the aerial image 1230 using a resist model 1240. The resist model can be used to calculate the resist image; para. 0139 - principle described herein works with any etch and combination of layers (e.g., a first resist, a second resist layer, etc.) of the substrate being patterned; para. 0220 – person skilled in art can perform the above methods using any images obtained before and after a particular process (e.g., OPC, optical process, resist process, etching, chemical mechanical polishing, etc.) or a combination of processes related to the patterning process. The model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0225 - image classified as defective after etching includes at least one of: a closed hole or a missing hole after etching due to resist blocking; para. 0249 - determine whether extra filtering step for a resist layer should be performed to reduce the failure rate; para. 0261 – produce a first image of the ADI feature, the ADI feature being a structure within a resist material; para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0363 - when a SEM is used to measure the feature shape in the resist; para. 0376 - parameters associated with a resist process or the etch process; para. 0396 - whether, e.g., the structures are in a latent resist image, in a developed resist image.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the undesired issue comprises a presence of a particle or a missing gate; and ... during a development of a resist on a device for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim(s) 2, 3, 15, 16, 23, and 24 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, in view of Kooiman, in view of Savalia, and further in view of Cheon, Sejune, et al. "Convolutional neural network for wafer surface defect classification and the detection of unknown defect class." IEEE Transactions on Semiconductor Manufacturing 32.2 (2019): 163-170, (“Cheon”).
Claim 2:
Modarres further teaches or suggests wherein the classification neural network in which the set of characteristics in a class in the classes is mutually exclusive (see p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 4 - features are sufficiently clear and distinct to be easily recognized by the network; p. 7 - Some MEMS electrodes had nanowires dispersed over them (Fig. 3a). The dominant feature in these images was the electrode like structure, since the nanowires were mostly dispersed and isolated from one another. p, 10 –network is best suited in identifying highly aligned from not-aligned images, as expected since it was trained on a binary choice. advantage of using our trained network is that a large batch of SEM images can be automatically classified based on the degree of ordering of the enclosed 1D nanostructures. The classification criterion was primarily based on the dimensionality of the nano- or micro scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_ Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powdernanowire categories, labelled either aligned (114 images) or not-aligned; p. 10 and 11 - the image classification to be highly effective in detecting distinct and representative images of each category.).
Savalia more specifically teaches or suggests the tensor comprises a 1024-dimension; each layer of the three layers comprises 1024 neurons (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the tensor comprises a 1024-dimension; each layer of the three layers comprises 1024 neurons for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Cheon more specifically teaches or suggests comprises a set of subnetworks of neurons corresponding to classes and also teaches the set of characteristics in a class in the classes is mutually exclusive (see Fig. 1, 5, 6; p. 165, §IIB - purpose of the full connection is to combine the global invariant features received from previous layers and use them as information for classification. The output layer is composed of as many nodes as there are class labels and outputs the class score for the input image; p. 166, §3A - increases the ability of the CNN to capture various geometric defect features with a rich set of feature maps; p. 167, §IIIB - image is input into the trained CNN (not the modified one) to receive its label from among the predetermined classes. uses CNN for defect classification; p. 167, IVA - 2,123 images of five defect classes (spot: 687; rock-shaped particle: 235; ring-shaped particle: 912; misalignment: 137; scratch: 152); p. 169, IVC - Fig. 6, which shows that the feature vectors of each defect class formed a cluster; therefore, the CNN successfully found information in defect images to achieve the goal of defect classification.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include comprises a set of subnetworks of neurons corresponding to classes for the purpose of efficiently implementing a neural network to classify images into separate classes using different portions of the neural network, improving neural network functionality and versatility, as taught by Cheon (§IIB, §IVC).
Claim 3:
Modarres further teaches or suggests wherein the classes comprise at least one of an exposure, a feature collapse, a component, an etch amount, or an alignment (see p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 4 - features are sufficiently clear and distinct to be easily recognized by the network; p. 7 - Some MEMS electrodes had nanowires dispersed over them (Fig. 3a). The dominant feature in these images was the electrode like structure, since the nanowires were mostly dispersed and isolated from one another. p, 10 –network is best suited in identifying highly aligned from not-aligned images, as expected since it was trained on a binary choice. advantage of using our trained network is that a large batch of SEM images can be automatically classified based on the degree of ordering of the enclosed 1D nanostructures. The classification criterion was primarily based on the dimensionality of the nano- or micro scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_ Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powdernanowire categories, labelled either aligned (114 images) or not-aligned; p. 10 and 11 - the image classification to be highly effective in detecting distinct and representative images of each category.).
Kooiman further teaches or suggests the change comprises a change to at least one of: a lithography, an etching time, or a dopant concentration (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0102 - an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after development image to classify whether the desired pattern will be defective after etching; para. 0139 - present disclosure is not limited to after development and after etch; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the change comprises a change to at least one of: a lithography, an etching time, or a dopant concentration for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim(s) 23 and 24:
Claim(s) 23 and 24 correspond to Claim 3 and thus, Modarres, Kooiman, Savalia and Cheon teach or suggest the limitations of claim(s) 23 and 24 as well.
Claim 15:
As indicated above, Modarres teaches or suggests in which the set of characteristics in a class in the classes is mutually exclusive.
Cheon further teaches or suggests wherein the classification neural network comprises subnetworks of neurons corresponding to classes and also teaches the set of characteristics in a class in the classes is mutually exclusive (see Fig. 1, 5, 6; p. 165, §IIB - purpose of the full connection is to combine the global invariant features received from previous layers and use them as information for classification. The output layer is composed of as many nodes as there are class labels and outputs the class score for the input image; p. 166, §3A - increases the ability of the CNN to capture various geometric defect features with a rich set of feature maps; p. 167, §IIIB - image is input into the trained CNN (not the modified one) to receive its label from among the predetermined classes. uses CNN for defect classification; p. 167, IVA - 2,123 images of five defect classes (spot: 687; rock-shaped particle: 235; ring-shaped particle: 912; misalignment: 137; scratch: 152); p. 169, IVC - Fig. 6, which shows that the feature vectors of each defect class formed a cluster; therefore, the CNN successfully found information in defect images to achieve the goal of defect classification.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include comprises a set of subnetworks of neurons corresponding to classes for the purpose of efficiently implementing a neural network to classify images into separate classes using different portions of the neural network, improving neural network functionality and versatility, as taught by Cheon (§IIB, §IVC).
Claim 16:
Modarres further teaches or suggests wherein the classes comprise at least one of exposure, a feature collapse, a component, an etch amount, or alignment (see p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 4 - features are sufficiently clear and distinct to be easily recognized by the network; p. 7 - Some MEMS electrodes had nanowires dispersed over them (Fig. 3a). The dominant feature in these images was the electrode like structure, since the nanowires were mostly dispersed and isolated from one another. p, 10 –network is best suited in identifying highly aligned from not-aligned images, as expected since it was trained on a binary choice. advantage of using our trained network is that a large batch of SEM images can be automatically classified based on the degree of ordering of the enclosed 1D nanostructures. The classification criterion was primarily based on the dimensionality of the nano- or micro scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_ Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powdernanowire categories, labelled either aligned (114 images) or not-aligned; p. 10 and 11 - the image classification to be highly effective in detecting distinct and representative images of each category.).
Kooiman further teaches or suggests the undesired issue comprises a presence of a particle or a missing gate (see para. 0225 - image classified as defective after etching includes at least one of: a closed hole or a missing hole after etching due to resist blocking; para. 0249 - determine whether extra filtering step for a resist layer should be performed to reduce the failure rate; para. 0261 – produce a first image of the ADI feature, the ADI feature being a structure within a resist material; para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include the undesired issue comprises a presence of a particle or a missing gate for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Claim(s) 4, 17, and 25 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, in view of Kooiman, in view of Savalia, in view of Cheon, and further in view of Brauer et al, US Publication no. 2018/0157933 (“Brauer”).
Claim 4:
Modarres does not explicitly disclose wherein a set of neurons in the subnetworks of neurons is deactivated during training of the classification neural network such that overfitting is reduced.
Brauer teaches or suggests wherein a set of neurons in the subnetworks of neurons is deactivated during training of the classification neural network such that overfitting is reduced (see para. 0050 – CNN may be used for defect classification during runtime; para. 0063 - dropout techniques may be utilized to prevent overfitting. As referred to herein, dropout techniques are a regularization technique for reducing overfitting in neural networks by preventing complex co-adaptations on training data. The term "dropout" refers to dropping out units (both hidden and visible) in a neural network. For example, at each training stage, individual nodes may be either "dropped out" of the CNN with probability 1-p or kept with probability p, so that a reduced CNN remains. In some embodiments, incoming and outgoing edges to a droppedout node may also be removed. Only the reduced CNN is trained. Removed nodes may then be reinserted into the network with their original weights; para. 0065 - configuration of a CNN may change based on the wafer, image data acquisition subsystem, or predetermined parameters.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein a set of neurons in the subnetworks of neurons is deactivated during training of the classification neural network such that overfitting is reduced for the purpose of efficiently using dropout techniques during CNN training, reducing overfitting and improving CNN performance on new input, as taught by Brauer (0063-0065).
Claim(s) 25:
Claim(s) 25 correspond to Claim 4 and thus, Modarres, Kooiman, Savalia, Cheon, and Brauer teach or suggest the limitations of claim(s) 25 as well.
Claim 17:
Brauer teaches or suggests wherein a set of neurons in the subnetworks of neurons is deactivated during training of the classification neural network such that overfitting is reduced (see para. 0050 – CNN may be used for defect classification during runtime; para. 0063 - dropout techniques may be utilized to prevent overfitting. As referred to herein, dropout techniques are a regularization technique for reducing overfitting in neural networks by preventing complex co-adaptations on training data. The term "dropout" refers to dropping out units (both hidden and visible) in a neural network. For example, at each training stage, individual nodes may be either "dropped out" of the CNN with probability 1-p or kept with probability p, so that a reduced CNN remains. In some embodiments, incoming and outgoing edges to a droppedout node may also be removed. Only the reduced CNN is trained. Removed nodes may then be reinserted into the network with their original weights; para. 0065 - configuration of a CNN may change based on the wafer, image data acquisition subsystem, or predetermined parameters.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein a set of neurons in the subnetworks of neurons is deactivated during training of the classification neural network such that overfitting is reduced for the purpose of efficiently using dropout techniques during CNN training, reducing overfitting and improving CNN performance on new input, as taught by Brauer (0063-0065).
Claim(s) 5, 18, and 26 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, in view of Kooiman, in view of Savalia, in view of Cheon, and further in view of Zhao et al, US Publication no. 2021/0241097 (“Zhao”).
Claim 5:
Modarres does not explicitly disclose wherein a set of neurons in the subnetworks of neurons have a set of amplitudes reduced during training of the classification neural network such that overfitting is reduced.
Zhao teaches or suggests wherein a set of neurons in the subnetworks of neurons have a set of amplitudes reduced during training of the classification neural network such that overfitting is reduced (see para. 0009 - improved training for a convolutional neural network model for object recognition, wherein, the optimization/updating amplitude, also known as the convergence gradient descent speed, for a convolutional neural network model is dynamically controlled during training, so as to adaptively match the progress of the training process, so that even for noisy training data sets, a high-performance training model can still be obtained; para. 0066 - order to constrain the optimization amplitude, the design idea of the weight function of the present disclosure is to design a mechanism which is effective for limiting the magnitude of the gradient, that is, it can flexibly control the gradient convergence speed suitable for a training data set with noise. That is to say, through usage of weight functions, variable amplitudes can be used to control the training convergence during training the convolutional neural network model, and the convergence speed will be slower and slower or even stop when the optimal training result is approached, therefore, instead of forcing fixed convergence as in the prior art, the convergence can appropriately stop or slow down, avoiding overfitting of noisy samples and ensuring that the model can effectively adapt to the training data set, thereby improving the performance of model training in terms of generalization; para. 0164 - using an improved weight function for dynamically controlling the amplitude of model updating/optimization, that is, the gradient descent speed, during the training process, it is possible to further optimize the training of a model for object recognition, such as a convolutional neural network model, compared with the prior art, so that a more optimized object recognition model can be obtained, and in turn the accuracy of object recognition/authentication is further improved.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein a set of neurons in the subnetworks of neurons have a set of amplitudes reduced during training of the classification neural network such that overfitting is reduced for the purpose of efficiently updating amplitude during CNN training, convergence can appropriately stop or slow down, avoiding overfitting of noisy samples and ensuring that the model can effectively adapt to the training data set, thus reducing overfitting and improving CNN performance on new input, as taught by Zhao (0009, 0066, and 0164).
Claim(s) 26:
Claim(s) 26 correspond to Claim 5 and thus, Modarres, Kooiman, Savalia, Cheon, and Zhao teach or suggest the limitations of claim(s) 26 as well.
Claim 18:
Zhao teaches or suggests wherein a set of neurons in the subnetworks of neurons have a set of amplitudes reduced during training of the classification neural network such that overfitting is reduced (see para. 0009 - improved training for a convolutional neural network model for object recognition, wherein, the optimization/updating amplitude, also known as the convergence gradient descent speed, for a convolutional neural network model is dynamically controlled during training, so as to adaptively match the progress of the training process, so that even for noisy training data sets, a high-performance training model can still be obtained; para. 0066 - order to constrain the optimization amplitude, the design idea of the weight function of the present disclosure is to design a mechanism which is effective for limiting the magnitude of the gradient, that is, it can flexibly control the gradient convergence speed suitable for a training data set with noise. That is to say, through usage of weight functions, variable amplitudes can be used to control the training convergence during training the convolutional neural network model, and the convergence speed will be slower and slower or even stop when the optimal training result is approached, therefore, instead of forcing fixed convergence as in the prior art, the convergence can appropriately stop or slow down, avoiding overfitting of noisy samples and ensuring that the model can effectively adapt to the training data set, thereby improving the performance of model training in terms of generalization; para. 0164 - using an improved weight function for dynamically controlling the amplitude of model updating/optimization, that is, the gradient descent speed, during the training process, it is possible to further optimize the training of a model for object recognition, such as a convolutional neural network model, compared with the prior art, so that a more optimized object recognition model can be obtained, and in turn the accuracy of object recognition/authentication is further improved.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein a set of neurons in the subnetworks of neurons have a set of amplitudes reduced during training of the classification neural network such that overfitting is reduced for the purpose of efficiently updating amplitude during CNN training, convergence can appropriately stop or slow down, avoiding overfitting of noisy samples and ensuring that the model can effectively adapt to the training data set, thus reducing overfitting and improving CNN performance on new input, as taught by Zhao (0009, 0066, and 0164).
Claim(s) 7 and 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, in view of Kooiman, in view of Savalia, and further in view of Stein et al., US Publication 2019/0325595 (“Stein”).
Claim 7:
Modarres further discloses wherein the convolutional neural network comprises: layers of neurons that identify the features in images, wherein a subsequent layer in the layers of neurons identifies larger features as compared to a prior layer in the layers of neurons (see p. 2 - lower layers of convolutional neural networks capture low-level image features, e.g. edges, while higher layers capture more complex details. we mainly focused on the feature extraction technique, by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images; p. 10 - presented a complete approach to automatically classify nanoscience SEM images by means of transfer learning technique and in particular feature extraction on deep convolutional neural networks.).
Stein more specifically teaches or suggests and wherein an output of the prior layer is input into the subsequent layer (see Fig. 5 and 12; para. 0067 - set of preprocessed images 530 are provided as input 506 to convolutional network portion 502. Each layer produces a feature map, which is in tum passed to the subsequent layer for further processing along forward propagation path 508. As depicted, the operations of convolutional network portion 502 operate to progressively reduce resolution of the feature maps, while increasing the number of channels (dimensionality) of the feature maps along convolutional forward propagation path SOSA; para. 0102 - convolutional stages 1204 then reduce resolution of the images 1202 but increase the number of channels, which creates complex features; para. 0501 - successive layers of the convolution portion are to operate to progressively reduce resolution of their corresponding feature maps while increasing dimensionality of the feature maps along the forward propagation path.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include and wherein an output of the prior layer is input into the subsequent layer for the purpose of efficiently implementing convolutional stages of a neural network to create complex features, improving neural network performance, as taught by Stein (0102 and 0501).
Claim 8:
Stein further teaches or suggests wherein each layer in the layers of neurons increases a number of channels and reduces a resolution of an image (see Fig. 5 and 12; para. 0067 - set of preprocessed images 530 are provided as input 506 to convolutional network portion 502. Each layer produces a feature map, which is in tum passed to the subsequent layer for further processing along forward propagation path 508. As depicted, the operations of convolutional network portion 502 operate to progressively reduce resolution of the feature maps, while increasing the number of channels (dimensionality) of the feature maps along convolutional forward propagation path SOSA; para. 0102 - convolutional stages 1204 then reduce resolution of the images 1202 but increase the number of channels, which creates complex features; para. 0501 - successive layers of the convolution portion are to operate to progressively reduce resolution of their corresponding feature maps while increasing dimensionality of the feature maps along the forward propagation path.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein each layer in the layers of neurons increases a number of channels and reduces a resolution of an image for the purpose of efficiently implementing convolutional stages of a neural network to create complex features, improving neural network performance, as taught by Stein (0102 and 0501).
Claim(s) 19, 27, and 29 is/are rejected under 35 U.S.C. 103 as being unpatentable over Modarres, in view of Kooiman, in view of Savalia, and further in view of Y. Yuan-Fu and S. Min, "Double Feature Extraction Method for Wafer Map Classification Based on Convolution Neural Network," 2020 31st Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC), Saratoga Springs, NY, USA, 2020, pp. 1-6, doi: 10.1109/ASMC49169.2020.9185393, Date of Conference: 24-26 August 2020, Date Added to IEEE Xplore: 03 September 2020, (“Yang”).
Claim 19:
Modarres further teaches or suggests neural network configured to receive images of the quantum devices, identify features for the quantum devices in the images, and output features identified for the quantum devices in the images of the quantum devices (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Yang more specifically teaches or suggests wherein the features are received in an output from a feature extraction neural network (see Fig. 1, 4; Abstract - then input this feature into the convolution layer for the second feature extraction; §II - then input this feature into the convolution layer for the second feature extraction. During this process, we can learn rich features at each layer, which used as good descriptors for image retrieval. Finally, our method can be extended to multi-class classification to recognize various types of failure patterns simultaneously; §C - CNN attempts to imitate this structure by extracting features in a similar way from the input space and then performing classification. Each convolutional layer in the network contains many feature maps. neurons in different feature maps extract different features; §E - learn rich features at each convolution layer.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include wherein the features are received in an output from a feature extraction neural network for the purpose of efficiently identifying important features using a neural network for classification purposes and/or further training, improving neural network performance, as taught by Yang (II).
Claim 27:
Modarres further teaches or suggests receiving features ... neural network configured to receive the image of the quantum device and output the features identified for the quantum device in the image of the quantum device (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Yang more specifically teaches or suggests receiving the features received in an output from a feature extraction neural network configured to receive the image (see Fig. 1, 4; Abstract - then input this feature into the convolution layer for the second feature extraction; §II - then input this feature into the convolution layer for the second feature extraction. During this process, we can learn rich features at each layer, which used as good descriptors for image retrieval. Finally, our method can be extended to multi-class classification to recognize various types of failure patterns simultaneously; §C - CNN attempts to imitate this structure by extracting features in a similar way from the input space and then performing classification. Each convolutional layer in the network contains many feature maps. neurons in different feature maps extract different features; §E - learn rich features at each convolution layer.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include receiving the features received in an output from a feature extraction neural network configured to receive the image for the purpose of efficiently identifying important features using a neural network for classification purposes and/or further training, improving neural network performance, as taught by Yang (II).
Claim 29:
Modarres further teaches or suggests a computer program product for screening quantum devices, wherein the computer program product comprises: a computer readable storage media; first program code, stored on the computer-readable storage media, executable by a computer system to cause the computer system to receive features received in an output ... configured to receive, respectively, an image of a quantum device ... of quantum devices and output the features identified in the image of the quantum device (see . 1 - extracting features from different types of microscope images; p. 2 - potential for feature extraction. feature extraction technique enables the application of learned features of a network. focused on the feature extraction technique., by retraining the final layers of an Inception-v312 deep convolutional neural network, implemented with the TensorFlow (TF) library13, to perform the classification task on SEM images (a) Feature extraction from an Inception-v3 model pre-trained on the ImageNet 2012 dataset, by retraining the last layers (softmax + fully connected layer) on the SEM dataset; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 -classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images.);
second program code, stored on the computer-readable storage media, executable by the computer system to cause the computer system to send features extracted from an image of a quantum device into a classification neural network ... wherein the classification neural network is configured to output ... of the quantum device, a set of characteristics for the quantum devices based on the features identified in the image of the quantum device (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.);
receive the set of characteristics identified by the classification neural network for the quantum devices ... based on the features identified in the image of the quantum device, wherein the set of characteristics indicates whether the quantum device will function with a desired level of consistency; and ... the quantum device ... the quantum device (see Fig. 1, 2, 8; p. 3 - One way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it; p. 9 - such as nanowires are used in a diverse range of applications from photonics to energy devices17,18. Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation; p. 10 - using our trained network is that a large batch of SEM images can be automatically classified. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. The following ten categories were identified and populated: Particles, Fibres, Biological, Patterned_surface, Nanowires, Tips, Films_Coated_Surface, Porous_sponge, MEMS_devices_and_electrodes, and Powder.).
Savalia further teaches or suggests grouped using a global pooling into a tensor that comprises a 1024-dimension; that each comprise 1024 neurons (see Fig. 5E and 7; Abstract - convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node; col. 25, lines 11-63 - a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include grouped using a global pooling into a tensor that comprises a 1024-dimension; that each comprise 1024 neurons for the purpose of efficiently reducing dimensionality, improving model performance, as taught by Savalia (0111 and 0707).
Kooiman further teaches or suggests from different stages of fabrication of a wafer that comprises thousands; that comprises three layers; prior to completed fabrication; in the different stages of fabrication; with a desired level of consistency; and initiate, based upon the characteristics: a determination, based upon the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and a release ... or a change to a fabrication process of ... that comprises a change to at least one of: lithography, an etching time, or a dopant concentration (see Fig. 6, 30; para. 0003 - substrate may then undergo various processes such as etching, ion-implantation (doping), metallization, oxidation, chemo-mechanical polishing, etc., all intended to finish off the individual layer of the device; para. 0004 - determining of defectiveness is based on comparing the given feature in the after development image with a corresponding etch feature in the after etch image; para. 0102 - (i) an after development image 401 of the imaged substrate at a given location, the after development image including a plurality of features, and an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features; para. 0103 – features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective; para. 0109 - classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature; para. 0111 – predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist; para. 0115 - model is a machine learning model such as a convolution neural network; para. 0119 - initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after-development image to classify whether the desired pattern will be defective after etching; para. 0139 - present disclosure is not limited to after development and after etch; para. 0220 - model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed; para. 0249 - used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Can be used to rework, based on the failure rate, a certain substrate or a lot of substrate before etching; para. 0299 - model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes); para. 0300 -accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes; para. 0308 - simulate a patterning process ( e.g., lithography step, etching, resist development, etc.). Then, based on the simulation results, it is possible to calibrate individual parameters according to, e.g., the correlation between results of different process (e.g., after resist development and after etch development; para. 0707 - adjusting, via simulating a patterning process using the correlation, parameters related to a lithographic process to cause a performance metric of a lithographic apparatus to be within a specified performance threshold; para. 0806 - concepts disclosed herein may be used for imaging on a substrate such as a silicon wafer, it shall be understood that the disclosed concepts may be used with any type of lithographic imaging systems, e.g., those used for imaging on substrates other than silicon wafers; para. 0807 - employed in the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, liquid-crystal displays (LCDs), thin film magnetic heads, micromechanical systems (MEMs ), etc.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include suggests from different stages of fabrication of a wafer that comprises thousands; that comprises three layers; prior to completed fabrication; in the different stages of fabrication; with a desired level of consistency; and initiate, based upon the characteristics: a determination, based upon the set of characteristics comprising an undesired issue, a root cause of the undesired issue in the images; and a release ... or a change to a fabrication process of ... that comprises a change to at least one of: lithography, an etching time, or a dopant concentration for the purpose of efficiently adjusting a fabrication process based on product functionality failure predictions, improving product fabrication, as taught by Kooiman (0111 and 0707).
Yang further teaches or suggests from feature extraction neural network (see Fig. 1, 4; Abstract - then input this feature into the convolution layer for the second feature extraction; §II - then input this feature into the convolution layer for the second feature extraction. During this process, we can learn rich features at each layer, which used as good descriptors for image retrieval. Finally, our method can be extended to multi-class classification to recognize various types of failure patterns simultaneously; §C - CNN attempts to imitate this structure by extracting features in a similar way from the input space and then performing classification. Each convolutional layer in the network contains many feature maps. neurons in different feature maps extract different features; §E - learn rich features at each convolution layer.).
Accordingly, it would have been obvious to one having ordinary skill before the effective filing date of the claimed invention to modify the system and method, taught in Modarres, to include from feature extraction neural network for the purpose of efficiently identifying important features using a neural network for classification purposes and/or further training, improving neural network performance, as taught by Yang (II).
Response to Arguments
Rejections under 35 USC 103:
Applicant argues that Modarres does not disclose “identify a set of characteristics for the quantum devices from a group of mutually exclusive characteristics based on these features, wherein the set of characteristics indicates whether the quantum devices will function”.
Modarres teaches one way of classifying nanostructures based on their shape and structure is their dimensionality, which is visually distinct and easy to be classified by the neural network, as well as a relevant property, since many of the functionalities of the nanostructures are strongly related to it. p3. Further, Factors such as the length, diameter, number density and orientation of the nanowires can determine the performance of such devices19,20. p. 9. Alignment of nanowires in a SEM image is determined by their orientation relative to a reference direction. Id. Aligned nanowires are oriented in the same direction with a low spread in angular distribution, whilst non-aligned nanowires have random directions with no prominent common orientation. Further, using our trained network is that a large batch of SEM images can be automatically classified. p. 10. classification criterion was primarily based on the dimensionality of the nano- or micro-scale objects represented in the images. Id. The Examiner notes Modarres teaches identifying characteristics based on features that indicate performance or functionality of the quantum devices such including nanowire characteristics such as alignment or non-alignment, and low spread or high spread angular distribution. Accordingly, Modarres teaches or suggests “identify a set of characteristics for the quantum devices from a group of mutually exclusive characteristics based on these features, wherein the set of characteristics indicates whether the quantum devices will function”.
Applicant argues the Office Action errs in asserting that Modarres discloses “a 14-block structure configured to: ... output, prior to completing fabricating the quantum device, the set of characteristics.” The Examiner notes the Office Action does not assert that Modarres discloses “a 14-block structure configured to: ... output, prior to completing fabricating the quantum device, the set of characteristics.”
Applicant argues Kooiman cannot be combined with Modarres. The Examiner respectfully disagrees. Modarres and Kooiman both deal with classifying devices based on characteristics of features identified in images. As indicated above, Modarres teaches or suggests classification pertaining to device functionality and performance. Kooiman also teaches or suggests classification pertaining to device functionality and performance. The Examiner further notes that though Modarres does not appear to use the term defect, Modarres discloses identifying characteristics based on features in images indicating defects.
Applicant argues the Office Action errs in asserting that Kooiman discloses “14-block structure ... and a screening manager configured to, based on the set of characteristics.”
The Examiner respectfully disagrees.
Kooiman discloses FIG. 30 is a block diagram that illustrates a computer system 100 which can assist in implementing methods and flows disclosed herein. para. 0403. The Examiner notes Figure 30 includes 14 blocks for implementing the methods and flows disclosed in Kooiman including a feature machine learning algorithm. Further, Kooiman teaches determining of defectiveness is based on comparing the given feature in the after development image with a corresponding etch feature in the after etch image. Para. 0004.
Further, (i) an after development image 401 of the imaged substrate at a given location, the after development image including a plurality of features, and an after etch image 402 of the imaged substrate at the given location, the after etch image including etched features corresponding to the plurality of features. Para. 0102. Further, features classified as defective in ADI will have high likelihood of failure after etch compared to features that were not classified as defective. Para. 0103. Further, classifying the identified feature as defective; and adjusting model parameter value of the model based on the defectiveness of the identified feature. Para. 0109. Further, predict defective features more accurately and appropriate adjustments to patterning process (e.g., an etch process) may be performed to improve the yield of the patterning process. Para. 0111. In an embodiment, the adjustments may involve changing the focus or dose of the lithographic apparatus, or adjusting the chemical composition of the resist. Id. Further, initial etch conditions 413 may be obtained. Procedure P415 involves executing the training model 403 using the after-development image to classify whether the desired pattern will be defective after etching. Para. 0119. Further, model then establishes are relationship between using such images to determine contribution of a process recipe (e.g., optical process recipe, resist process recipe, etch recipe, etc.) towards probability of failure after the process is performed. Para. 0220. Further, used to tune a lithographic process to reduce the failure rate of ADI features after etching, wherein the tuning comprises adjusting dose, focus ,or both. Para. 0249. Further, model 2210 can be employed in or associated with a lithographic apparatus to tune lithographic parameters, based on model-predicted failure rates, to reduce the number of feature failures (e.g., filled contact holes). Para. 0299. Further, accurate defect classification based on ADI can help to find the root cause of AEI failures of e.g., contact holes. a fraction of filled contact holes can be used to assess whether extra descumming or punch-through should be used before etch to reduce the impact of filled contact holes. Para. 0300. The Examiner notes Kooiman teaches screening management based upon characteristics. Accordingly, Kooiman teaches or suggests “14-block structure ... and a screening manager configured to, based on the set of characteristics.”
Applicant argues the Office Action errs when it asserts Savalia discloses “output features grouped using global pooling into a tensor.”
Savalia teaches convolutional operations followed by a global average pooling layer, a fully connected layer with 1024 node. Abstract. Further, a global average pooling layer, a fully connected layer with 1024 node. use of the global average pooling operation (GAP) 508.1 operates to minimize overfitting by reducing the total number of parameters in the 55 model. Col. 25. GAP layers are similar to max pooling layers in that they are used to reduce the spatial dimensions of a three-dimensional tensor as shown in FIG. 7. Id. However, GAP layers perform a more extreme type of dimensionality reduction, where a tensor with dimensions hxwxd is reduced 60 in size to have dimensions lxlxd. GAP layers reduce each hxw feature map to a single number by taking the average of all hw values. Id. The Examiner notes Savilia uses global pooling to arrive at reduced tensor containing values corresponding to features. Accordingly, Savalia teaches or suggests “output features grouped using global pooling into a tensor.” The Examiner notes that Savalia does not teach away is analogous art that analyzes images and performs at least classification.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Andrew T McIntosh whose telephone number is (571)270-7790. The examiner can normally be reached M-Th 8:00am-5:30pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tamara Kyle can be reached at 571-272-4241. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/ANDREW T MCINTOSH/Primary Examiner, Art Unit 2144