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 .
DETAILED ACTION
This Office Action is in response to Application No. 18/950590 filed on November 18, 2024. Claims 1-20 are presented for examination and are currently pending.
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) 1-5, 8-12 and 15-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sahouria (US 20100229145 A1) in view of Besson (US 12354256 B2).
Regarding claim 8, Sahouria teaches a computer-implemented method comprising:
generating, by the processor, a class neighborhood graph for a lamella milling processes (“the master computer 103 is a multi-processor computer that includes a plurality of input and output devices 105 and a memory 107.” — ¶[0032]; “The processor units 111 may be any type of processor device that can be programmed to execute the software instructions 109A” — ¶[0035]; “a conflict graph is constructed that reflects these relationships. More particularly, each polygon segment is represented by a node in the conflict graph. Nodes representing polygon segments that have separated polygon edges are connected by a first type of graph edge…Nodes representing abutting polygon segments similarly are connected by a second type of edge” — ¶[0015]), wherein the generating comprises:
classifying, by the device, structures (“the separated edge identification module 303 identifies separated geometric or polygon edges in the initial layout design data 315. As used herein, the term ‘separated edges’ refers to geometric edges that should or must be formed using separate lithographic masks.” — ¶[0054]; “the segmentation module 305 generates proposed cut paths to segment the polygons in the initial layout design data 315.” — ¶[0048]) within a region of interest (“the separated edge identification module 303 identifies separated geometric or polygon edges in the initial layout design data 315. As used herein, the term ‘separated edges’ refers to geometric edges that should or must be formed using separate lithographic masks.” — ¶[0054]; “the segmentation module 305 generates proposed cut paths to segment the polygons in the initial layout design data 315.” — ¶[0048]) of a design schematic of a sample (“The layout design data may be in any desired format, such as, for example, the Graphic Data System II (GDSII) data format or the Open Artwork System Interchange Standard (OASIS) data format proposed by Semiconductor Equipment and Materials International (SEMI).” — ¶[0050]; “layout design data will include one or more geometric elements to be written to a mask or reticle. For conventional mask or reticle writing tools, the geometric elements typically will be polygons of various shapes.” — ¶[0051]); and
generating, by the device, a directed graph of the region of interest (“the graphing module 307 creates a conflict graph representing the separated edges and the proposed cut paths. More particularly, the graphing module 307 creates a conflict graph representing the relationship between the proposed cut paths, the polygon segments created by the proposed cut paths, and the separated edges shared by different polygon segments.” — ¶[0061]; “Next, a dual of the conflict graph is constructed.” — ¶[0016]; “Thus, each pair of separated polygon edges will have a corresponding separation graph edge in the conflict graph, and each cut path will have a corresponding cut path graph edge in the conflict graph.” — ¶[0015]; “the cut path selection module 309 will identify the nodes in the dual graph having an odd number of incident dual graph separation edges…identify the dual graph cut path edges that make up the dual graph cut path edge minimum-weight T-join of the identified nodes.” — ¶[0068]), wherein the directed graph comprises nodes representing the classified structures (“each polygon segment is represented by a node in the conflict graph” — ¶[0015]; “FIG. 9 illustrates a conflict graph created for the layout design data, separated edges and proposed cut paths shown in FIG. 8. As seen in this figure, each polygon segment 501A-501D has a corresponding node 501A’-501D’.” — ¶[0062]) and weighted directed edges between the nodes (“the cut path selection module 309 will…assign a weight to each of the edges in the dual graph. More particularly, the cut path selection module 309 may assign a relatively large weight, such as an essentially ‘infinite’ value, to each of the separation graph edges in the dual graph…It also will assign a relatively small weight, such as zero or a finite value, to each of the cut path graph edges in the dual graph.” — ¶[0068]).
Sahouria, however, does not explicitly teach wherein the weighted directed edges between the nodes representing distances between the classified structures.
Besson, in analogous art, teaches wherein the weighted directed edges between the nodes representing distances between the classified structures (“The nodes of the graph can be defined as the vertices of the brain surface mesh and the edges can be defined as the links of the triangles weighted as a function of the length of the link, such as (1) where A_i,j is the Euclidean distance between vertices i and j.” — c6 L35-65; “A scalar edge weight, w_ij, can be assigned to connect vertices v_i and v_j using their geodesic distance, w_ij, along the surface” — c14 L40-55).
Sahouria, on the one hand, teaches generating a graph from layout design data (GDS/OASIS format) where nodes represent classified polygon segments and edges between nodes are assigned weights. However, Sahouria’s edge weights represent decomposition cost values rather than physical distances between structures.
Besson, on the other hand, teaches that in graph-based analysis, edge weights can be computed as a function of the physical distance between nodes (e.g., Euclidean distance or geodesic distance), thereby encoding spatial proximity relationships into the graph structure.
Moreover, Sahouria’s weighted graph is a known method ready for improvement. Applying Besson’s known technique of distance-based edge weighting to Sahouria’s layout graph would yield the predictable result of a graph that encodes the physical spatial relationships (distances) between classified structures in the layout design data. Such distance-based weighting would predictably improve the graph’s ability to represent how closely or distantly structures are positioned relative to one another, which is useful information for any downstream process that relies on the spatial layout of the design.
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to modify Sahouria’s method of generating a weighted graph from layout design data by weighting the graph edges based on physical distances between the classified structures, as taught by Besson.
Moreover, weighting graph edges by physical distance is a fundamental and well-known concept in graph theory that is not unique to Besson’s specific application. One of ordinary skill in the art of graph-based analysis would readily recognize that substituting distance-based weights for arbitrary cost-based weights would predictably encode spatial proximity — a straightforward and well-understood improvement to any graph that represents spatially-arranged elements.
Regarding claim 9, Sahouria/Besson teach the computer-implemented method of claim 8, wherein the design schematic comprises a diagram of the sample and the region of interest (Sahouria: “the initial layout design data 315 may include geometric elements such as the polygons 501 and 503 shown in FIG. 5.” — ¶[0051]; “FIG. 5 illustrates an example of polygons in layout design data.” — ¶[0023]; “The geometric elements, which typically are polygons, define the shapes that will be created in various materials to manufacture the circuit.” — ¶[0006]) and further comprising determining, by the device, distances between the classified structures (Besson: “The nodes of the graph can be defined as the vertices of the brain surface mesh and the edges can be defined as the links of the triangles weighted as a function of the length of the link, such as (1) where A_i,j is the Euclidean distance between vertices i and j.” — c6 L35-65) based on measurements from center points of the classified structures to center points of other classified structures (Sahouria: “the separation directive polygon 601 separates edge 505 from edge 507, designating these geometric edges…as separated edges. Similarly, the separation directive polygon 603 separates edge 509 from edge 511, designating these polygon edges as separated edges.” — ¶[0056]; and Besson: “A_i,j is the Euclidean distance between vertices i and j.” — c6 L35-65).
Furthermore, regarding “measurements from center points of the classified structures to center points of other classified structures”; measuring distances between geometric shapes from their center points (centroids) is a well-known and elementary technique in computational geometry. One of ordinary skill would recognize this as a straightforward method for determining distances between polygon segments identified in layout design data. Therefore, the combination Sahouria/Besson clearly teaches this particular limitation.
Regarding claim 10, Sahouria/Besson teach the computer-implemented method of claim 9, further comprising classifying, by the device, the structures within the region of interest (Sahouria: “the separated edge identification module 303 identifies separated geometric or polygon edges in the initial layout design data 315.” — ¶[0054]; “the segmentation module 305 generates proposed cut paths to segment the polygons in the initial layout design data 315.” — ¶[0048]; “the cut path selection module 309 generates proposed cut paths for the geometric elements (i.e., the polygons) in the initial layout design data 315. Any desired technique can be used to generate these proposed cut paths.” — ¶[0057]).
Regarding claim 11, Sahouria/Besson teach the computer-implemented method of claim 8, wherein the design schematic comprises a textual description of the sample (Sahouria: “The layout design data may be in any desired format, such as, for example, the Graphic Data System II (GDSII) data format or the Open Artwork System Interchange Standard (OASIS) data format proposed by Semiconductor Equipment and Materials International (SEMI). Other formats include an open source format named Open Access, Milkyway by Synopsys, Inc., and EDDM by Mentor Graphics, Inc.” — ¶[0050]) and further comprising determining, by the device, distances between the classified structures based on an extraction of the distances from the textual description (Sahouria: “the initial layout design data 315 may include information used to interpret the data describing the structures, or to specify special treatment for some subset of the structure data.” — ¶[0053]; “As will be appreciated by those of ordinary skill in the art, layout design data will include one or more geometric elements to be written to a mask or reticle…the layout design data usually include polygon data describing the features of polygons in the design.” — ¶[0051]).
Regarding claim 12, Sahouria/Besson teach the computer-implemented method of claim 11, further comprising classifying, by the device, the structures based on an extraction of class descriptions within the textual description (Sahouria: “the initial layout design data 315 may include information used to interpret the data describing the structures, or to specify special treatment for some subset of the structure data. For instance, it may be undesirable to generate a cut path in a polygon representing a transistor gate. Accordingly, some type of prohibition information may be included with that polygon to indicate to the tool 301 that no cut path may be generated within that polygon.” — ¶[0053]).
Regarding claims 1-5 and 15-19; these claim(s) limitations are significantly similar to those of claim(s) 8-12; and, thus, are rejected on the same grounds.
Claim(s) 6, 7, 13, 14 and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Sahouria (US 20100229145 A1) in view of Besson (US 12354256 B2); and further in view of Miller (US 20220067915 A1).
Regarding claim 13, Sahouria/Besson explicitly teach all the claim limitations except for the computer-implemented method of claim 8, wherein the region of interest comprises a left boundary, a right boundary, a first milling stage thickness and a final thickness.
Miller, on analogous art, teaches wherein the region of interest comprises a left boundary, a right boundary (“a surface of a sample is milled to remove a layer of material, the surface is then imaged and the image is analyzed by a neural network to determine if the structure indicates an end point to the milling process.” — ¶[0013]; “the chunk may be further milled and imaged in sequential steps to determine when a desired end point is reached.” — ¶[0027]; “the milling may remove other finFETs that are located between the surface of the chunk and the feature of interest.” — ¶[0028]), a first milling stage thickness (“The high beam energy milling may be used to remove relatively thick layers from a surface of the chunk, whereas the low beam energy milling may be used to obtain a final desired thickness for the lamella with the feature of interest centered therein.” — ¶[0027]; “an initial milling step of a sample may be performed at a high beam energy and an associated end point may be selected to prevent sample damage past a desired location. Upon end pointing the high beam energy process, a subsequent milling process at a lower beam energy may be performed to remove any damage and to mill to a new end point location.” — ¶[0018]) and a final thickness (“The chunk may be thicker than a final lamella, which will be used for imaging one or more features of interest. For example, the chunk may be around 100 nm thick while a final lamella may be around 15 to 25 nm thick, and less than 15 nm in some examples.” — ¶[0026]; “the end point may be a location to stop milling so that a desired feature to image is located at a central location of a thin lamella, e.g., ~15 nm in thickness.” — ¶[0018]).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to have combined the method of generating a graph from layout design data as taught by Sahouria, with edge weights representing physical distances as taught by Besson, and to have applied this combined method to a lamella milling process in which the region of interest comprises a left boundary, a right boundary, a first milling stage thickness and a final thickness, as taught by Miller.
The combination would have been obvious because a person of ordinary skill in the art would understand that each element (graph generation from design data, distance-based edge weighting, and ROI definition with boundaries and thicknesses for lamella milling) was known in the prior art and performed its respective function independently. Sahouria’s graph generation method operates on a defined region of layout data. Miller’s lamella milling process defines a region of interest with specific dimensional parameters (lateral boundaries and depth thicknesses) for controlled material removal. Combining these elements — applying Sahouria’s graph generation to a region of interest defined by Miller’s lamella milling parameters — yields the predictable result of a class neighborhood graph that represents the classified structures within a region bounded laterally and by depth stages, as would be encountered in a multi-stage lamella milling process.
Regarding claim 14, Sahouria/Besson explicitly teach all the claim limitations except for the computer-implemented method of claim 8, further comprising: classifying, by the device, using a milling neural network, one or more structures in an image of a cutface of the sample; modifying, by the device, the classifications based on the class neighborhood graph; and instructing, by the device, a scientific instrument to mill the sample based on the modified classifications.
Miller, in analogous art, teach further comprising: classifying, by the device, using a milling neural network, one or more structures in an image of a cutface of the sample (“the MLS 114, which may also be referred to as a deep learning system, is a machine-learning computing system. In some embodiments, the MLS 114 is an artificial neural network (ANN) that includes a collection of connected units or nodes, which are called artificial neurons.” — ¶[0030]; “the MLS 114 is a CNN configured to detect and/or identify, e.g., classify or segment, objects of interest shown in an image of the sample, e.g., chunk or lamella. A feature/structure of interest is a portion of the sample that is imaged” — ¶[0032]; “the image may be analyzed by the MLS 114 to determine whether the end point has been reached…the ANN may be a CNN or an RCNN that segments or classifies the images to determine whether the end point has been reached.” — ¶[0040]; “Process block 205 may be followed by process block 207, which includes analyzing the image using artificial intelligence to determine if the end point has been reached…In either embodiment, the end point may be based on features in the image that indicate a location of the face of the chunk/lamella with respect to the feature(s) of interest.” — ¶[0040]);
modifying, by the device, the classifications based on the class neighborhood graph (“the ANN or machine learning system/algorithm is used to both analyze images and to provide feedback for charged particle microscope control. The aim of the analysis and feedback is to provide additional or full automation of lamella preparation processes” — ¶[0019]); and
instructing, by the device, a scientific instrument to mill the sample based on the modified classifications (“In the disclosed solution, the ANN or machine learning system/algorithm is used to both analyze images and to provide feedback for charged particle microscope control.” — ¶[0019]; “If the neural network determines the structure is indicative of the end point, the process ends, else the milling process may be repeated.” — ¶[0013]; “based on the end point not being reached, milling, by the focused ion beam, the surface of the sample to remove a layer of material; and based on the end point being reached, cease material removal.” — Claim 13).
Miller’s method of using a neural network to classify structures in images of a sample during lamella milling is a known method that is ready for improvement. Miller explicitly acknowledges the difficulty of accurately classifying structures due to the variability and complexity of features in the images. This difficulty creates a recognized need for additional information or constraints to improve classification accuracy.
Sahouria, on the one hand, teaches that a graph representation of the design layout — capturing the known structural relationships and spatial arrangement of classified structures — can be generated from design data and used to guide downstream processing decisions. Besson, on the other hand, teaches that graph structure (nodes, edges, and edge weights) can be used to influence and modify neural network classification outputs, i.e., the known technique of graph-informed neural network classification.
Hence, Applying the known technique of using a graph to modify neural network classifications (as taught by Besson) to Miller’s neural network classification during lamella milling (a known method ready for improvement), where the graph is derived from design layout data (as taught by Sahouria with distance-based weights as taught by Besson), would yield the predictable result of improved classification accuracy. Specifically, the graph encodes the known spatial arrangement and relationships of structures within the sample. When the neural network generates a raw classification of a structure visible in a cutface image, the graph can be consulted to modify that classification based on which structures are expected to be present at that location given what has already been identified — thereby constraining the neural network’s output with prior design knowledge.
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, faced with the recognized difficulty of accurately classifying structures during lamella milling (as acknowledged by Miller), would have been motivated to incorporate known structural relationships from the sample’s design data (as taught by Sahouria) to improve classification accuracy, using the known technique of graph-informed neural network processing (as taught by Besson).
Regarding claims 6, 7 and 20; these claim(s) limitations are significantly similar to those of claim(s) 13 and 14; and, thus, are rejected on the same grounds.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Sun (CN 113592008 A) teaches a small-sample classification framework in which an encoder-decoder learns intra-class difference information, reconstructed support samples are generated, and those samples are combined with query samples in a graph neural network for edge-label prediction and class inference. It is a useful graph-learning reference for few-shot classification, but it does not teach design schematics, lamella milling, directed graphs, or edge weights representing distances between classified structures in a sample region.
Flanagan (US 20230034667 A1), this patent discloses segmenting a first image of structure into classes, associating a class with a second image, and using the association to perform metrology via edge finders or other boundary-locating analytics. It is relevant to AI-assisted metrology and image segmentation, but it does not teach building a class neighborhood graph from a design schematic, using directed graph edges, or controlling a milling instrument based on modified classifications.
Sharma (US 20210201474 A1), this application teaches generating synthetic training images for visual inspection by combining compliant component images with synthetically generated defect images, then training a machine-learning model for defect detection or counting tasks. It is useful as a synthetic-data training reference, but it is not directed to design schematics, lamella milling, graph construction, or distance-weighted directed edges between classified structures.
Bradley (US 20230154101 A1), this reference discloses a morphable radiance field / NeRF-style model for generating photorealistic head images from multiple viewpoints using identity, deformation, and canonical appearance codes. It is relevant to multi-view generative modeling and rendering, but it does not address graph neighborhood construction, design schematics, lamella milling, or graph-based modification of classifications for a scientific instrument.
Sun (US 20230351215 A1), this patent teaches dynamic graph representation learning by extracting structural embeddings from graph snapshots and then applying temporal convolution over those embeddings to predict graph context. It is relevant to dynamic graph embedding generally, but it does not teach design schematics, directed graph generation from classified structures, lamella milling, or controlling a milling instrument based on modified classifications.
Terliuc (US 20230014490 A1), this application discloses endoscope-based imaging with a mechanical enhancement element, such as an inflatable balloon, to stretch tissue and improve detection of regions of interest using a trained machine-learning model. It is relevant to mechanically enhanced imaging and AI detection, but it does not teach graph generation from design schematics, directed weighted edges based on structural distances, or lamella milling/instrument control.
Miller (US 20220067915 A1), this patent teaches obtaining images of a sample surface, analyzing the images with a machine-learning system to determine whether a milling end point has been reached, and then milling or ceasing material removal accordingly. It is directly relevant to AI-controlled lamella preparation and milling endpointing, but it does not teach class neighborhood graphs from design schematics or distance-based graph construction.
Cacella (US 5851413 A), this patent discloses a gas delivery system for particle beam processing using a shroud-type concentrator with an axial passage to deliver reactant material to a workpiece while allowing a particle beam to pass through. It is relevant to FIB/gas-delivery hardware and particle-beam processing, but it does not teach graph generation, classification of structures from a schematic, or graph-based neural-network modification of classifications.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to RAMON A MERCADO whose telephone number is (571)270-5744. The examiner can normally be reached Mo-Th: 5:30AM-4PM.
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, David Yi can be reached on 5712707519. 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.
/Ramon A. Mercado/Supervisory Patent Examiner, Art Unit 3658
Prosecution Insights