DETAILED ACTION
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 .
Response to Amendment & Claim Interpretation
In the Reply filed 24 August 2026 Applicant amended the independent claims 1, 16, and 20 to recite:
wherein, in the training phase, the machine learning model is trained to generate semantic segmentation maps without relying on height measurements as inputs to the machine learning model; and
wherein the trained machine learning model is operative to generate the semantic segmentation map from the 2D inspection image without relying on height measurements of the inspection area as inputs to the trained machine learning model.
Initially, it is noted that the highlighted language are negative limitations due to the use of the word “without relying on height measurements as input”. Guidance on negative limitations may be found in MPEP 2173.05(i) stating in relevant part that:
The current view of the courts is that there is nothing inherently ambiguous or uncertain about a negative limitation. So long as the boundaries of the patent protection sought are set forth definitely, albeit negatively, the claim complies with the requirements of 35 USC 112, second paragraph … Any negative limitation or exclusionary proviso must have basis in the original disclosure… Any claim containing a negative limitation which does not have basis in the original disclosure should be rejected under 35 U.S.C. 112, first paragraph, as failing to comply with the written description requirement
The instant specification supports “semantic” segmentation techniques that rely on height measurements of the of the inspection area including using AFM (atomic force microscope) to measure the heights that are used as inputs in the training phase to train the machine learning model. See, for example, [0117], [0129]. As such, the amended claim language disclaims such embodiments (training with actual height measurements) by requiring “without relying on height measurements” and thus limits the invention to training the model using label data associated with a height profile that is inferred from the training images or labels manually provided by a user as disclosed in [0119] or via supervised feedback in [0117] both of which train the machine learning model such that “in the training phase, the machine learning model is trained to generate semantic segmentation maps without relying on height measurements as inputs to the machine learning model” and thereby “generate the semantic segmentation map from the 2D inspection image without relying on height measurements of the inspection area”. In other words, the specification supports both “with relying on” and “without relying on” height measurements such that the negative limitations in the amendments disclaim or otherwise do not encompass the embodiments that rely on height measurements.
Response to Arguments
Applicant’s arguments with respect to claims 1, 16, and 20 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
More specifically, Applicant an RCE together with a Reply on 24 August 2026 including substantial amendments to the independent claims that necessitate new grounds of rejection and render the arguments moot.
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1-4, 9, 11, and 13-20 are rejected under 35 U.S.C. 102(a)(1) and (a)(2) as being anticipated by Young (US 2022/0043357 A1).
Claim 1
In regards to claim 1, Young discloses a system comprising one or more processing circuitries {see Figs. 1, 5, computer systems 100, 102, 45 which contain processing circuitries including computer readable medium 600, program instructions and processor 605, ]to implement the disclosed method, [0016]-[0017], [0042], [0077], [0106]-[0107], [0111]-[0115]} configured to:
obtain a 2D inspection image of an inspection area of a semiconductor specimen
{see Fig. 1a imaging system 10 using an optical or light-based imaging to generate a 2D inspection image of a semiconductor specimen such as 2D image of an inspection area used for inspecting bump heights on a semiconductor wafer as per [0001]-[0010], [0028]-[0039]. Fig. 4 illustrates the 2D inspection image 412. Note that the 2D inspection image may be a 2D optical bright field (BF) and/or dark field (DF) images and that such BF and DF 2D images do not contain or include height information but nonetheless optically reveal and are indicative of height profile patterns such that they are leveraged by Young to semantically segment the 2D image as further explained below}, and
feed the 2D inspection image to a trained machine learning model to generate a semantic segmentation map of at least one given structural element represented in the 2D inspection image
{Fig. 5 illustrates feeding inspection image (e.g. DF image 500, BF image 502, and bump patch images 510, 512) into DL (deep learning) model 514 to generate a semantic segmentation map based on “predicted height”.
It is noted that “predicted height” is not a ground truth measurement of the actual height of the bump but semantic match with, for example, the nominal (height) bump image 414 or the short (height) bump image 416. See [0065]. Fig. 4 illustrates two such semantic categories such that the DL model effectively semantically segments the bump patch images into two (height-based) categories/classes.
Further as to “semantic segmentation map” note that the machine learning system of Fig. 5 spatially segments the input images 500/502 using a template and mask 508 to generate isolated bump patch images 510/512 which are used to produce a map of bump heights (semantic segmentation map), [0100], [0105] further describing a map of bump heights across the specimen including locations of each “height profile pattern” (bump object) and structure classifications including class labels. Taken together these elements describe all of the essential technical details of a semantic segmentation map and are thus considered as within the common understanding and meaning of these terms as well as the BRI of the claim language.
NOTE: the instant specification does not define or even mention a single time the terms “sematic”, “semantic segmentation”, or “semantic segmentation map”. While the specification supports these broadly recited and conventional concepts, the BRI of these terms is thus consistent with the plain and ordinary meaning. As such, the fact that Young also does not happen to use these same terms is of no import and does not affect the alignment between Young’s disclosure and the claimed invention both of which are “semantic segmentation” generating a “semantic segmentation map”},
wherein the semantic segmentation map comprises:
a first spatial segment assigned a first feature identity semantically identifying the first spatial segment as a first feature type of the given structural element; and a second spatial segment assigned a second feature identity, different from the first feature identity, semantically identifying the second spatial segment as a second feature type, different from the first feature type, of the same given structural element, wherein the first and second feature types are associated, respectively, with a first height-profile pattern and a second height-profile pattern different from the first height-profile pattern,
{See Fig. 4 (copied below) and associated explanations.
PNG
media_image1.png
484
656
media_image1.png
Greyscale
Young classifies spatial segments corresponding to structural elements such as semiconductor bump structures 408, 410 into semantic height categories to generate a semantic segmentation map comprising:
a first spatial segment (e.g. bump image 414) assigned a first feature identity (first class label, first ID, [0105]) semantically identifying the first spatial segment as a first feature type (nominal height) of the given structural element (first bump) and
a second spatial segment (e.g. bump image 416) assigned a second feature identity (second class label, second ID, [0105]) semantically identifying the second spatial segment as a second feature type (short height) of the given structural element (second bump).
See Fig. 4 illustrating these two semantic height categories, Note that, per [0065], the “predicted height” generated by Young’s classification/segmentation is “not a ground truth measurement of the height” but instead a semantic category. In other words, the height of the bumps is “predicted”/classified via a learned relationship between training bump images and (semantic) labels indicating two trained classes/categories (nominal height category and short height category which are first and second feature types being respectively associated with first and second height-profile patterns as claimed. See also Fig. 4 graphically illustrating this correspondence and mapping to the claim language. Thus, while Young does not utilize the term “semantic segmentation”, the system of Fig. 5 including DL model 514 performs semantic segmentation and generates a semantic segmentation map, [0087]-[0092], [0094]-[0105]).
Note also that Young’s segmentation further corresponds to the first and second “spatial segment”-based segmentation claimed because Youn’g classification/semantic height label is applied to the structural element (e.g. semiconductor bump structures).
In sum, Young’s system performs 2(or more)-level classification of whole structural elements into, e.g. 2 semantic categories, each of these semantic categories being respectively associated with first and second “height-profile patterns” (nominal height profile pattern and short height profile pattern)}.
wherein the machine learning model has been trained, in a training phase, using a plurality of 2D training images and segmentation label data, each of the plurality of 2D training images representing a respective training structural element corresponding to the given structural element, the segmentation label data identifying, for each respective training structural element represented in each of the plurality of 2D training images:
a first training segment assigned the first feature identity, thereby identifying the first training segment as belonging to the first feature type of the respective training structural element, the first feature type being associated with the first height-profile pattern; and a second training segment assigned the second feature identity, thereby identifying the second training segment as belonging to the second feature type of the same respective training structural element, the second feature type being associated with the second height-profile pattern
{see above including Fig. 4 illustrating 2D training images. Further as to training, see [0078]-[0083] and the mapping of claim 16 below. Further as to segmentation label data, see the ground truth training in [0076]-[0083], [0105]}
wherein, in the training phase, the machine learning model is trained to generate semantic segmentation maps without relying on height measurements as inputs to the machine learning model; and wherein the trained machine learning model is operative to generate the semantic segmentation map from the 2D inspection image without relying on height measurements of the inspection area as inputs to the trained machine learning model
{see above including using inferred height from the training images or known (not measured) bump heights such as in [0081], [0082] known from design values of the wafer and not measured}.
Claim 2
In regards to claim 2, Young discloses the system configured to, during run-time examination of the specimen, use the trained machine learning model to determine, in the inspection image, a plurality of segments corresponding to different types of features of interest of the same given structural element present in the inspection area.
{see above including, e.g., wherein the different feature types of features correspond to the nominal bump height and short bump height and the segments correspond to the labeled bump objects and/or bump images 510, 512}.
Claim 3
In regards to claim 3, Young discloses wherein the types of features of interest, or data informative thereof, have been used in training of the machine learning model {see above cites for claim 1 wherein the features of interest (images of different bump heights/categories/types, Fig. 4) have been used in training the ML model}.
Claim 4
In regards to claim 4, Young discloses the system configured to:
for each given structural element of a plurality of structural elements present in the inspection area, use the trained machine learning model to determine segments of the inspection image corresponding to different types of features of interest of said given structural element, thereby obtaining a set of segments, and use the set of segments to determine metrology data informative of the plurality of the structural elements
{see above mapping of claim 1, wherein the template 506 and mask 508 are used to determine segments of the inspection image corresponding to different types of features of interest (nominal or short bump height) and uses this set of segments to determine metrology data (predicted height) informative of the plurality of structural elements 408, 410}.
Claim 9
In regards to claim 9, Young discloses wherein the segmentation label data
have been obtained using three- dimensional data informative of one or more areas of one or more semiconductor specimens, wherein the three-dimensional data have been acquired by an examination tool {see above mapping of clam 1 including [0083] in which the ground truth measurements (label data) for bump height is acquired by a bump height measurement tool}.
Claim 11
In regards to claim 11, Young discloses wherein at least one of the first feature type or the second feature type is informative of at least one of: a foot of an element present in the inspection area, a slope of an element present in the inspection area, an edge of an element present in the inspection area, a round edge of an element present in the inspection area, a top edge of an element present in the inspection area {see above cites for claim 1 including Fig. 4 in which the segment(s)/features are at least broadly “informative” of an edge and top edge and a round edge of an element present in the inspection area}.
Claim 13
In regards to claim 13, Young discloses wherein, for each given 2D training image informative of a given area, the given segmentation label data has been obtained using an image representative of 3D information of the given area acquired by an examination tool
{see claim 1 mapping. See also Fig. 1a imaging system 10 using an optical or light-based imaging to generate a 2D inspection image of a semiconductor specimen such as 2D image of an inspection area used for inspecting bump heights on a semiconductor wafer as per [0001]-[0010], [0028]-[0039]. Fig. 4 illustrates the 2D inspection image 412. See also the dark field (DF) image 400 and BF (bright field) image 502 which are also images representative of 3D information of the given area acquired by an examination tool}.
Claim 14
In regards to claim 14, Young discloses wherein the image representative of 3D information of the given area has been acquired by an Atomic Force Microscope or a Scanning Transmission Electron Microscope
{see scanning electron microscope SEM in [0005] which may be the tool used in [0083] to measure bump height}.
Claim 15
In regards to claim 15, Young discloses configured to use at least one of the first spatial segment S'1 or the second spatial segment S'2 to determine metrology data informative of the inspection area {see above wherein the determination of height features and predicted heigh are metrology data informative of the inspection area as claimed}.
Claim 16
In regards to claim 16, Young discloses a system comprising one or more processing circuitries {see Figs. 1, 5, computer systems 100, 102, 45 which contain processing circuitries including computer readable medium 600, program instructions and processor 605, ]to implement the disclosed method, [0016]-[0017], [0042], [0077], [0106]-[0107], [0111]-[0115]} configured to:
obtain a plurality of 2D training images and segmentation label data, the plurality of 2D training images representing respective training structural elements belonging to a same category
{see Fig. 1a imaging system 10 using an optical or light-based imaging to generate 2D images including a “set” (plurality) of training images 414 of bump 408 and a set of training images 416 of bump 410 and represent respective training structural training elements, [0063]-[0065] and Fig. 4. Further as to the 2D imaging system used to obtain the training images and the run-time image for segmentation of a 2D semiconductor wafer image see [0001]-[0010], [0028]-[0039]. Note that the 2D training and inspection images may be 2D optical bright field (BF) and/or dark field (DF) images and that such BF and DF 2D images do not contain or include height information but nonetheless optically reveal and are indicative of height profile patterns such that they are leveraged by Young to semantically segment the 2D image as further explained herein},
the segmentation label data identifying, for each respective training structural element: a first training segment assigned a first feature identity semantically identifying the first training segment as belonging to a first feature type of the respective training structural element; and a second training segment assigned a second feature identity, different from the first feature identity, semantically identifying the second training segment as belonging to a second feature type, different from the first feature type, of the same respective training structural element
{see above including Fig. 4 illustrating 2D training segments (image portions)
Including
a first training segment (e.g. bump image 414) assigned a first feature identity (first class label, first ID, [0105]) semantically identifying the first training segment as belonging to a first feature type (nominal height) of the respective training structural element (first bump)
a second training segment (e.g. bump image 416) assigned a second feature identity (second class label, second ID, [0105]) semantically identifying the second training segment as belonging to a second feature type (short height) of the respective training structural element (second bump).
Further as to training, see [0078]-[0083]. Further as to segmentation label data, see the ground truth training in [0076]-[0083], [0105]},
wherein the first and second feature types are associated, respectively, with a first height-profile pattern and a second height-profile pattern different from the first height-profile pattern {see above, nominal and short height profile patterns which are different},
feed the plurality of 2D training images and the segmentation label data to the machine learning model to train the machine learning model to generate semantic segmentation maps from 2D images without relying on height measurements as inputs to the machine learning model
{see [0078]-[0083]. Further as to segmentation label data, see the ground truth training in [0076]-[0083], [0105]} See mapping of claim 1 above for “without relying on height measurements as inputs”},
wherein the trained machine learning model is operative to generate, based on a 2D inspection image of an inspection area comprising a given structural element belonging to the same category, a semantic segmentation map comprising a first spatial segment assigned the first feature identity semantically identifying the first spatial segment as belonging to the first feature type of the given structural element; and a second spatial segment assigned the second feature identity semantically identifying the second spatial segment as belonging to the second feature type of the same given structural element, without relying on height measurements of the inspection area as inputs to the trained machine learning model
{Fig. 5 illustrates feeding inspection image (e.g. DF image 500, BF image 502, and bump patch images 510, 512) into DL (deep learning) model 514 to generate a semantic segmentation map based on “predicted height”.
It is noted that “predicted height” is not a ground truth measurement of the actual height of the bump but semantic match with, for example, the nominal (height) bump image 414 or the short (height) bump image 416. See [0065]. Fig. 4 illustrates two such semantic categories such that the DL model effectively semantically segments the bump patch images into two (height-based) categories/classes.
Further as to “semantic segmentation map” note that the machine learning system of Fig. 5 spatially segments the input images 500/502 using a template and mask 508 to generate isolated bump patch images 510/512 which may be used to produce a map of bump heights (semantic segmentation map), [0100], [0105] further describing a map of bump heights across the specimen including locations of each “height profile pattern” (bump object) and structure classifications including class labels. Taken together these elements describe all of the essential technical details of a semantic segmentation map and are thus considered as within the common understand and meaning of these terms as well as the BRI of the claim language.
NOTE: the instant specification does not define or even mention a single time the terms “sematic”, “semantic segmentation”, or “semantic segmentation map”. While the specification supports these broadly recited and conventional concepts, the BRI of these terms is thus consistent with the plain and ordinary meaning. As such, the fact that Young also does not happen to use these same terms is of no import and does not affect the alignment between Young’s disclosure and the claimed invention both of which are “semantic segmentation” generating a “semantic segmentation map”
Young classifies spatial segments corresponding to structural elements such as semiconductor bump structures into semantic height categories to generate semantic segmentation map comprising:
a first spatial segment (e.g. bump image 414) assigned a first feature identity (first class label, first ID, [0105]) semantically identifying the first spatial segment as a first feature type (nominal height) of the given structural element (first bump) and
a second spatial segment (e.g. bump image 416) assigned a second feature identity (second class label, second ID, [0105]) semantically identifying the second spatial segment as a second feature type (short height) of the given structural element (second bump).
See Fig. 4 illustrating these two semantic height categories, Note that, per [0065], the “predicted height” generated by Young’s classification/segmentation is “not a ground truth measurement of the height” but instead a semantic category. In other words, the height of the bumps is “predicted”/classified via a learned relationship between training bump images and (semantic) labels indicating two trained classes/categories (nominal height category and short height category which are first and second feature types being respectively associated with first and second height-profile patterns as claimed. See also Fig. 4 graphically illustrating this correspondence and mapping to the claim language. Thus, while Young does not utilize the term “semantic segmentation”, the system of Fig. 5 including DL model 514 performs semantic segmentation and generates a semantic segmentation map, [0087]-[0092], [0094]-[0105]).
Note also that Young’s segmentation further corresponds to the first and second “spatial segment” based segmentation claimed because Youn’g classification/semantic height label is applied to the structural element (e.g. semiconductor bump structures).
In sum, Young’s system performs 2-level classification of whole structural elements into two semantic categories, each of these semantic categories being respectively associated with first and second “height-profile patterns” (nominal height profile pattern and short height profile pattern)}.
Claim 17
In regards to claim 17, Young discloses wherein the trained machine learning model is operative to identify, in the 2D inspection image of the inspection area, the first feature type and the second feature type of the same given structural element, without receiving 3D information on the inspection area {see above cites including mapping of claim 1}.
Claim 18
In regards to claim 18, Young discloses performing (i) or (ii):
i) for each given 2D training image area, obtaining a given second image representative of 3D information of a given area represented in the given 2D training image, wherein for said given area, at least part of the segmentation label data is determined using the given second image {see the BF and DF images which may respectively considered the first and second given training images representative of 3D information and used for training as claimed}, or
(ii) for each given 2D training image, obtaining a given second image representative of 3D information of the given area acquired by an Atomic Force Microscope or a Scanning Transmission Electron Microscope, wherein, for each said given area, at least part of the segmentation label data is determined using the given second image
{see scanning electron microscope SEM in [0005] which may be the tool used in [0083] to measure bump height for training label data}.
As Young, discloses both options i) and ii)}.
Claim 19
In regards to claim 19, Young discloses wherein the system is configured to:
segment the given 2D training image is using the given second image, data informative of the first height-profile pattern and data information of the second height profile pattern {see mapping of claims 1 and 18 in which segmentation of the given image is performed by processing circuitry using the identified data in the claim}; or
segment the given 2D training image using a feedback of a user and the given second image {supervised feedback [0067].
Claim 20
The rejection of device system claim 1 above applies mutatis mutandis to the corresponding limitations computer readable medium claim 20 while noting that the rejection above cites to both device and method disclosures as well as computer readable storage medium implementations.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Michael R Cammarata whose telephone number is (571)272-0113. The examiner can normally be reached M-Th 7am-5pm EST.
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, Matthew Bella can be reached at 571-272-7778. 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.
/MICHAEL ROBERT CAMMARATA/ Primary Examiner, Art Unit 2667
