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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 2025/12/31; 2024/12/09; 2024/07/17 & 2024/07/08. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1–34 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (an abstract idea) without reciting significantly more.
Regarding independent claims 1, 10, 19, and 27
Step 1 — whether the claim falls within any statutory category. See MPEP 2106.03.
Claim 1 is drawn to a method (process); claim 10 is drawn to a computer system (machine); claim 19 is drawn to a method (process); and claim 27 is drawn to a computer system (machine). Therefore, each of these claims falls under one of the four categories of statutory subject matter (process/method, machine/product/apparatus, manufacture, or composition of matter). The analysis proceeds to Step 2A.
Step 2A Prong 1 — whether the claim recites a judicial exception. See MPEP 2106.04, subsection II.
Regarding independent claim 1, the claim is directed to a method for supporting non-contact measurement of a structure, the method comprising obtaining a first set of metrology data, training a first machine learning model, obtaining a second set of metrology data, and performing transfer learning to produce a second machine learning model for predicting key parameters.
The limitations of "training a first machine learning model for the first one or more structures using the first set of metrology data" and "performing transfer learning from the first machine learning model to the second set of metrology data to produce a second machine learning model for predicting key parameters for the second one or more structures" recite an abstract idea.
These limitations recite mathematical concepts under MPEP 2106.04(a)(2), subsection I, because the training of a machine learning model and the performance of transfer learning to produce a model that predicts parameter values are accomplished through mathematical calculations and mathematical relationships, including the iterative adjustment of model weights to minimize a loss function and the mapping of feature inputs to predicted parameter outputs. Consistent with the 2024 Guidance Update on Patent Subject Matter Eligibility (AI) and Example 47, Claim 2, the recited training of a machine learning model and use of that model to generate predictions involve mathematical calculations and therefore recite a mathematical-concept abstract idea.
Accordingly, independent claim 1 recites a judicial exception, namely an abstract idea in the form of mathematical concepts under MPEP 2106.04(a)(2), subsection I. The analysis proceeds to Step 2A Prong 2.
Independent claim 10 is a computer system claim reciting limitations similar to claim 1 — obtaining first and second sets of metrology data, training a first machine learning model, and performing transfer learning to produce a second machine learning model for predicting key parameters — and recites the same mathematical-concept abstract idea for the same reasons.
Regarding independent claim 19, the claim is directed to a method for supporting measurement of a structure, comprising obtaining first and second sets of metrology data, selecting metrology data using a feature extractor, and training a machine learning model for predicting key parameters. The limitations of "selecting metrology data from the first set of metrology data and the second set of metrology data using a feature extractor" and "training a machine learning model with selected metrology data for predicting key parameters for the second one or more structures" recite mathematical concepts under MPEP 2106.04(a)(2), subsection I, because feature extraction/selection and the training of a machine learning model to predict parameter values are performed through mathematical calculations and mathematical relationships. Accordingly, independent claim 19 recites an abstract idea.
Independent claim 27 is a computer system claim reciting limitations similar to claim 19 and recites the same mathematical-concept abstract idea for the same reasons.
Step 2A Prong 2 — whether the claim as a whole integrates the recited judicial exception into a practical application. This evaluation is performed by (1) identifying whether there are any additional elements recited in the claim beyond the judicial exception, and (2) evaluating those additional elements individually and in combination to determine whether the claim as a whole integrates the exception into a practical application. See MPEP 2106.04(d).
Regarding independent claim 1, this claim recites the additional elements of "obtaining a first set of metrology data for a first one or more structures" and "obtaining a second set of metrology data for a second one or more structures," and the claim is generally limited to the field of "supporting non-contact measurement of a structure."
These additional elements do not integrate the recited judicial exception into a practical application. The two "obtaining" limitations are mere data gathering that supplies the metrology data on which the abstract training and transfer-learning operations are performed, and therefore constitute insignificant extra-solution activity. See MPEP 2106.05(g). The preamble recitation that the method is for "supporting non-contact measurement of a structure" merely links the abstract idea to a particular technological environment or field of use; the claim does not recite actually measuring the structure, nor does it recite applying the predicted key parameters to control, adjust, or otherwise affect a metrology tool or a fabrication process. See MPEP 2106.05(h). Further, the claim does not recite an improvement to the functioning of a computer or to another technology or technical field, because it ends at producing a model that predicts key parameters and does not apply those predictions in any technological process. See MPEP 2106.04(d)(1) and MPEP 2106.05(a). The claim also does not apply the judicial exception with a particular machine, effect a particular transformation, or include other meaningful limitations beyond the abstract idea. See MPEP 2106.05(b), 2106.05(c), and 2106.05(e). This posture is analogous to Example 47, Claim 2, in which receiving training data, training the network, and outputting results did not integrate the mathematical-concept abstract idea into a practical application. Accordingly, claim 1 is directed to the abstract idea under Step 2A.
Regarding independent claim 10, this claim is drawn to a computer system reciting limitations similar to claim 1 and is directed to the abstract idea under the same rationale. Claim 10 additionally recites "a computer system … comprising: at least one processor, wherein the at least one processor is configured to" perform the operations. These additional elements are recited at a high level of generality and merely provide a generic computer component used as a tool to perform the abstract training and transfer-learning operations, amounting to no more than mere instructions to apply the exception using a generic computer component (see MPEP 2106.05(f)) and to generally linking the abstract idea to a computer environment (see MPEP 2106.05(h)). The claim therefore does not integrate the exception into a practical application.
Regarding independent claim 19, this claim recites the additional elements of "obtaining a first set of metrology data for a first one or more structures" and "obtaining a second set of metrology data for a second one or more structures," and is generally limited to the field of "supporting measurement of a structure." These additional elements do not integrate the exception into a practical application for the same reasons given for claim 1: the "obtaining" limitations are insignificant extra-solution data gathering (see MPEP 2106.05(g)), and the field-of-use preamble does not recite actually measuring the structure or applying the predicted key parameters to a metrology tool or fabrication process (see MPEP 2106.05(h), 2106.05(a)). Claim 19 is therefore directed to the abstract idea.
Regarding independent claim 27, this claim is drawn to a computer system reciting limitations similar to claim 19 and is directed to the abstract idea under the same rationale, further reciting "at least one processor … configured to" perform the operations, which amounts to no more than mere instructions to apply the exception on a generic computer (see MPEP 2106.05(f)) and generally links the exception to a computer environment (see MPEP 2106.05(h)).
Step 2B — whether the claim amounts to significantly more than the judicial exception. See MPEP 2106.05.
Regarding independent claim 1, the additional elements beyond the judicial exception are the "obtaining a first set of metrology data" and "obtaining a second set of metrology data" limitations and the field-of-use preamble. As explained in the Step 2A Prong 2 analysis, the "obtaining" limitations are mere data gathering used to supply the metrology data for the abstract analysis and therefore are insignificant extra-solution activity. Upon reevaluation in Step 2B, receiving or gathering data for processing is recognized as well-understood, routine, and conventional activity. See MPEP 2106.05(d)(II) and the court decisions discussed therein. Considered individually and in combination, these additional elements do not provide an inventive concept and do not amount to significantly more than the judicial exception.
Regarding independent claim 10, the additional elements are the "obtaining" limitations and the generic "computer system" and "at least one processor." For the reasons given for claim 1, the "obtaining" limitations are well-understood, routine, conventional data-gathering activity (see MPEP 2106.05(d) and 2106.05(g)), and the recited processor/computer system is generic computer implementation that the courts have recognized as well-understood, routine, and conventional (see MPEP 2106.05(d)(II) and 2106.05(f)). Considered individually and in combination, the additional elements of claim 10 do not provide an inventive concept.
Regarding independent claims 19 and 27, the additional elements are the "obtaining" data-gathering limitations (claims 19 and 27) and the generic "at least one processor"/computer system (claim 27). For the reasons given above, these amount only to insignificant extra-solution data gathering and generic computer implementation, which are well-understood, routine, and conventional. See MPEP 2106.05(d), 2106.05(f), and 2106.05(g). Considered individually and in combination, the additional elements do not provide an inventive concept.
Accordingly, independent claims 1, 10, 19, and 27 do not recite additional elements, individually or in combination, that amount to significantly more than the recited judicial exception, and are therefore ineligible under 35 U.S.C. § 101.
Regarding dependent claims 2–9, 11–18, 20–26, and 28–34
Step 1 — whether the claims fall within any statutory category. See MPEP 2106.03.
Dependent claims 2–9 and 19's dependents 20–26 depend from method claims 1 and 19 and are drawn to method claims within the process category; dependent claims 11–18 and 28–34 depend from computer system claims 10 and 27 and are drawn to machine claims within the machine category. Each therefore falls within a statutory category.
Step 2A Prong 1 — whether the claims recite a judicial exception. See MPEP 2106.04, subsection II.
Dependent claims 2–9, 11–18, 20–26, and 28–34 include the abstract idea limitations previously identified with respect to their base independent claims and further narrow that abstract idea with additional mathematical-concept limitations under MPEP 2106.04(a)(2), subsection I:
Claims 2, 5, 11, 14, 25, and 33 further characterize the data on which the model is trained (relative parameter variation; structure types/processes) and do not alter the mathematical-concept character of the training/prediction.
Claims 3, 4, 12, 13, 23, 24, 31, and 32 further characterize the training data as synthetic/experimental and labeled/unlabeled and thus further define the inputs to the mathematical training operation.
Claims 6–9, 15–18, 26, and 34 recite training a third machine learning model, performing further transfer learning, and transferring different model layers, which are additional mathematical training and transfer-learning operations.
Claims 20–22 and 28–30 recite minimizing domain differences, including by co-training, which are mathematical optimization operations.
Claims 21 and 29 recite minimizing domain differences using a domain classifier via a gradient reversal layer, which is a mathematical operation performed during model training.
These dependent-claim limitations recite mathematical concepts for the reasons given for the independent claims.
Step 2A Prong 2 and Step 2B
Claims 2–9, 11–18, 20–26, and 28–34 merely narrow the previously identified abstract idea. For the reasons described above with respect to independent claims 1, 10, 19, and 27, the judicial exceptions recited in these dependent claims are not meaningfully integrated into a practical application and do not amount to significantly more than the abstract idea. The additional elements recited in these dependent claims — to the extent any are present beyond the abstract idea — are limited to the same generic computer components and insignificant extra-solution data-gathering activity addressed above (see MPEP 2106.05(f), 2106.05(g), and 2106.05(h)), which are well-understood, routine, and conventional (see MPEP 2106.05(d)). None of the dependent claims recites actually measuring the structure or applying the predicted key parameters to control or adjust a metrology tool or fabrication process. Accordingly, dependent claims 2–9, 11–18, 20–26, and 28–34 also recite abstract ideas that do not integrate into a practical application or amount to significantly more than the judicial exception, and are rejected under 35 U.S.C. § 101.
Claim Rejections - 35 USC § 103
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 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.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 4, 5, 10, 13, and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2.
Regarding claim 1, (Yacoby) teaches a method for supporting non-contact measurement of a structure ((Yacoby), col. 1, ll. 6–10, Field of the Invention: "optical inspection of integrated circuit wafer patterns … measurement of wafer pattern parameters"; col. 1, ll. 38–43: "Optical critical dimension (OCD) metrology employs methods of scatterometry to measure scatterometric data, that is, reflected light radiation"; FIG. 1, items 10/12/14), the method comprising:
"obtaining a first set of metrology data for a first one or more structures";
"training a first machine learning model for the first one or more structures using the first set of metrology data";
"obtaining a second set of metrology data for a second one or more structures"; and
"performing transfer learning from the first machine learning model to the second set of metrology data to produce a second machine learning model for predicting key parameters for the second one or more structures."
With respect to "obtaining a first set of metrology data for a first one or more structures," (Yacoby) teaches receiving plural first sets of scatterometric (optical metrology) data measured from wafer patterns, which are the recited first one or more structures ((Yacoby), Abstract: "receiving multiple first sets of scatterometric data"; col. 3, ll. 3–4 (Summary); col. 5, ll. 62–64 and FIG. 2, step 214: "A first step 214 includes receiving multiple sets of scatterometric data, measured from respective wafer patterns"). (Yacoby) further teaches that the wafer patterns are structures having measurable pattern parameters ((Yacoby), col. 1, ll. 16–22: "stacked structures … critical dimensions (CDs) … height, width, and pitch of stacks"). Thus (Yacoby) teaches obtaining a first set of metrology data for a first one or more structures.
With respect to "training a first machine learning model for the first one or more structures using the first set of metrology data," (Yacoby) teaches training, in a self-supervised manner, auto-encoder neural networks using the first sets of scatterometric data, the auto-encoder neural networks constituting the first machine learning model ((Yacoby), col. 3, ll. 4–8 (Summary): "training, in a self-supervised manner, k² auto-encoder neural networks"; col. 6, ll. 30–47 and FIG. 2, step 220: each sub-vector pair is "applied to perform 'self-supervised' training of an auto-encoder neural network (NN)"; FIG. 3A, auto-encoder NN 300, encoder network 310, bottleneck layer 340). Thus (Yacoby) teaches training a first machine learning model for the first one or more structures using the first set of metrology data.
With respect to "obtaining a second set of metrology data for a second one or more structures," (Yacoby) teaches receiving plural second sets of scatterometric data measured from respective wafer patterns (the second one or more structures), together with corresponding reference parameters ((Yacoby), col. 3, ll. 9–12 (Summary): "receiving multiple respective sets of reference parameters and multiple corresponding second sets of scatterometric data, measured from multiple respective wafer patterns"; col. 6, l. 66–col. 7, l. 8 and FIG. 2, step 222/230). Thus (Yacoby) teaches obtaining a second set of metrology data for a second one or more structures.
With respect to "performing transfer learning from the first machine learning model to the second set of metrology data to produce a second machine learning model for predicting key parameters for the second one or more structures," (Yacoby) teaches a transfer-learning stage in which a transfer neural network — whose initial layers reuse the pretrained encoder networks of the first machine learning model — is trained using the second sets of scatterometric data as feature input and the reference parameters as target output, thereby producing a second machine learning model (the transfer NN) that estimates wafer pattern parameters (the recited key parameters) of the second one or more structures ((Yacoby), col. 7, ll. 1–35 and FIG. 2, step 230: "a transfer learning stage of NN training is performed"; "The k² encoder networks 310 … are combined in parallel into an input stage of a new neural network, referred herein as a 'transfer NN'"; col. 3, ll. 12–22 (Summary): "training a transfer neural network (NN) having initial layers including a parallel arrangement of the k² encoder neural networks … such that the transfer NN is trained to estimate new wafer pattern parameters from subsequently measured sets of scatterometric data"; FIG. 3B, transfer NN 400, output layer 410; col. 7, ll. 56–60: "three parameters, as shown in the figure, e.g., height, width, and pitch of a given wafer stack"). One of ordinary skill in the art would interpret the teaching of (Yacoby) as resulting in performing transfer learning from the first machine learning model to the second set of metrology data to produce a second machine learning model for predicting key parameters for the second one or more structures. The recited "key parameters" read on these wafer pattern parameters/critical dimensions ((Yacoby), col. 1, ll. 16–22). Thus (Yacoby) teaches performing transfer learning from the first machine learning model to the second set of metrology data to produce a second machine learning model for predicting key parameters for the second one or more structures.
Regarding claim 4, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1.
Regarding the limitation added by claim 4, (Yacoby) teaches:
"wherein at least a portion of the first set of metrology data is labeled and at least a portion of the second set of metrology data is labeled."
(Yacoby) teaches that the sets of scatterometric data used for training are paired with corresponding reference parameters that are used as the target output (labels) for machine-learning training ((Yacoby), col. 5, ll. ~30–37: "Reference parameters 44 may be used as target output for ML training. The reference parameters may be acquired from patterns of one or more wafers by high accuracy means known in the art, such as … CD-SEM, AFM, TEM, X-ray metrology"). (Yacoby) further teaches that, during transfer-NN training, "target output of the transfer NN training is set to the multiple sets of reference parameters and feature input is set to the multiple corresponding second sets of scatterometric data" ((Yacoby), col. 3, ll. ~17–22), such that each second set of scatterometric data is paired with, i.e., labeled by, its corresponding reference parameters ((Yacoby), col. 7, ll. ~26–33: "each set of feature input … is mapped to the corresponding reference parameters measured from the same wafer pattern"). One of ordinary skill in the art would interpret the teaching of (Yacoby) as resulting in at least a portion of the first set of metrology data is labeled and at least a portion of the second set of metrology data is labeled. Because the reference parameters constitute labels associated with the scatterometric data, (Yacoby) teaches that at least a portion of the metrology data used for training is labeled.
To the extent applicant argues that (Yacoby) does not expressly label "the first set of metrology data," (Yacoby) teaches that the second sets of scatterometric data corresponding to reference parameters "may be a subset of the … first sets of scatterometric data" ((Yacoby), col. 3, ll. ~30–33; col. 7, ll. ~1–5), such that the labeled second-set portion is also a labeled portion of the first set. (Yacoby) thereby teaches that at least a portion of the first set of metrology data is labeled and at least a portion of the second set of metrology data is labeled.
Regarding claim 5, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1.
Regarding the limitation added by claim 5, (Yacoby) teaches the limitation, broken down into its sub-limitations below:
"wherein the first one or more structures and the second one or more structures are different types of structures"; or
"are a same type of structures produced using a same or different processes."
With respect to "wherein the first one or more structures and the second one or more structures are different types of structures," (Yacoby) teaches that the patterned stacks measured for OCD metrology may be of differing types, including two-dimensional and three-dimensional profile types ((Yacoby), col. 1, ll. ~30–35: pattern parameters may include "over-fill/under-fill of 2-dimentional (HKMG), 3-dimentional profile (FinFETs)"), and that these stacked structures are "formed in repetitive patterns" of differing geometry and material composition ((Yacoby), col. 1, ll. ~13–22). Because the first and second sets of scatterometric data are measured from respective wafer patterns ((Yacoby), col. 3, ll. ~9–12), (Yacoby) teaches that the first one or more structures and the second one or more structures may be different types of structures.
With respect to "are a same type of structures produced using a same or different processes," (Yacoby) teaches that the wafer patterns from which the first and second sets are measured "are typically fabricated dies, which may be measured from one or more wafers" ((Yacoby), col. 7, ll. ~5–8), and that "the multiple respective wafer patterns may be located on one or more wafers" ((Yacoby), col. 3, ll. ~38–40). Because dies of the same pattern type may be fabricated on the same wafer or on different wafers (i.e., by the same or different fabrication processes), (Yacoby) teaches that the first and second structures may be a same type of structure produced using a same or different process. One of ordinary skill in the art would interpret the teaching of (Yacoby) as resulting in the first one or more structures and the second one or more structures are different types of structures or are a same type of structures produced using a same or different processes.
The two sub-limitations are recited in the alternative ("different types of structures or are a same type of structures produced using a same or different processes"). Under the broadest reasonable interpretation, the limitation is satisfied by teaching either alternative; (Yacoby) teaches both.
Because (Yacoby) teaches each and every limitation of claim 5, the claim is unpatentable under 35 U.S.C. 103, anticipation being the epitome of obviousness. Accordingly, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to practice the method of claim 5 as taught by (Yacoby).
Claim 10 recites a computer system comprising at least one processor configured to perform the steps of claim 1. Claim 10 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 1.
Claim 13 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 4.
Claim 14 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 5.
Claims 2, 3, 11, and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2, in view of Bhaskar et al. (Bhaskar), US 2017/0193400 A1, cited in the IDS filed 7/17/2024.
Regarding claim 2, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1. Claim 2 adds a limitation that is addressed by its sub-limitations below:
"wherein one of the first one or more structures or the second one or more structures includes larger variations in structural parameters, layer property parameters, material property parameters, or a combination thereof"; and
"than an other of the first one or more structures or the second one or more structures."
(Yacoby) teaches that variations between sets of scatterometric data are indicative of differing pattern parameters of the wafer structures ((Yacoby), col. 5, ll. 30–34; col. 1, ll. 16–22). However, (Yacoby) does not teach the added limitation. In the same field of endeavor, (Bhaskar) teaches the added limitation, as set forth below.
With respect to "wherein one of the first one or more structures or the second one or more structures includes larger variations in structural parameters, layer property parameters, material property parameters, or a combination thereof," (Bhaskar) teaches acquiring information for "non-nominal instances" of one or more specimens — instances in which one or more defects, abnormalities, or process-induced parameter variations are present — such that one set of specimens includes larger variations in structural, layer, and material parameters ((Bhaskar), Abstract: "acquiring information for non-nominal instances of specimen(s)"; Detailed Description, definition of "non-nominal instances": the non-nominal instances of the specimen(s) are instances in which one or more defects, abnormalities, or variations are present).
With respect to "than an other of the first one or more structures or the second one or more structures," (Bhaskar) teaches that the machine-learning-based model is trained with only information for "nominal instances" of additional specimens — instances free of such defects, abnormalities, or variation — such that the other set of specimens (the nominal instances) includes smaller parameter variation than the non-nominal set ((Bhaskar), Abstract: "trained with only information for nominal instances of additional specimen(s)"; Detailed Description, definition of "nominal instances": the nominal instances of the specimen(s) are instances in which no defect or abnormality is present). (Bhaskar) thereby teaches one set of structures having larger parameter variation relative to an other set of structures.
(Yacoby) and (Bhaskar) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based optical metrology and characterization of semiconductor wafer structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-supervised and transfer-learning OCD-metrology method of (Yacoby) with the nominal/non-nominal training of (Bhaskar), such that one of the first or second sets of structures includes larger parameter variations than the other. The motivation to combine is as recited by (Bhaskar), which teaches that training the model to account for non-nominal instances exhibiting larger parameter variation improves the accuracy and robustness of the model when applied to specimens produced under varying process conditions ((Bhaskar), Abstract); accordingly, one of ordinary skill would have incorporated (Bhaskar)'s structures having larger parameter variation into (Yacoby)'s framework — in which a wider-variation first set is transferred to a narrower second set — to reduce the quantity of high-accuracy reference data and re-training required when fabrication conditions change, with a reasonable expectation of success because both references train machine learning models on metrology data of semiconductor wafer structures.
Regarding claim 3, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1.
Regarding the limitations added by claim 3, (Yacoby) teaches:
"experimental metrology data generated from the first one or more structures" ((Yacoby), col. 5, ll. ~62–64 and FIG. 2, step 214: "receiving multiple sets of scatterometric data, measured from respective wafer patterns"; col. 5, ll. ~14–18: "The scatterometry data 32 generated by the metrology system 30"); and
"experimental metrology data generated from the second one or more structures" ((Yacoby), col. 3, ll. ~9–12: "multiple corresponding second sets of scatterometric data, measured from multiple respective wafer patterns").
(Yacoby) teaches something related to generating estimated scatterometry data from an optical model of a structure having known parameters ((Yacoby), col. 2, ll. ~8–17: optical models "can … be applied to generate, from a set of known pattern parameters, an estimate of scatterometry data that would be measured during spectrographic testing"). However, (Yacoby) does not teach "synthetic metrology data generated from one or more models of the first one or more structures … and synthetic metrology data generated from one or more models of the second one or more structures."
In the same field of endeavor, (Bhaskar) teaches this limitation. (Bhaskar) teaches a machine-learning-based model "configured for performing simulation(s) for the specimens" ((Bhaskar), Abstract), and teaches generating the training input by "empirical simulation of real defect events on wafers and reticles using DOEs" ((Bhaskar), col. 16, ll. ~14–17) and by models that generate the instance information used for training ((Bhaskar), col. 15–16), thereby teaching synthetic metrology data generated from one or more models of the structures. Combined with the experimental metrology data taught by (Yacoby), (Bhaskar) also satisfies the "or a combination thereof" alternative recited for each set.
(Yacoby) and (Bhaskar) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based optical metrology and characterization of semiconductor wafer structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-supervised-and-transfer-learning OCD-metrology method of (Yacoby) with the model-generated (synthetic) training data of (Bhaskar), such that the first and second sets of metrology data each comprise synthetic metrology data generated from one or more models of the respective structures, experimental metrology data, or a combination thereof. The motivation to combine (Yacoby) and (Bhaskar) is as recited by (Bhaskar), which teaches that generating training information by simulation allows the model to be adequately trained where actual specimen data is limited or where variation/defect instances are rare, thereby reducing the amount of measured data required ((Bhaskar), Abstract; col. 16–17), with a reasonable expectation of success because both references train machine learning models on metrology data of semiconductor wafer structures.
Claim 11 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 2.
Claim 12 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 3.
Claims 6-9 and 15-18 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2, in view of Honda et al. (Honda), US 2018/0356807 A1.
Regarding claim 6, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1. Claim 6 adds a limitation addressed by its sub-limitations below:
"obtaining a third set of metrology data for a third one or more structures";
"training a third machine learning model for the third one or more structures using the third set of metrology data"; and
"performing transfer learning from the third machine learning model with the first machine learning model to the second set of metrology data to produce the second machine learning model for predicting key parameters for the second one or more structures."
Regarding claim 6, (Yacoby) teaches:
"performing transfer learning from the … machine learning model with the … machine learning model to the second set of metrology data to produce the second machine learning model for predicting key parameters for the second one or more structures" ((Yacoby) teaches that the layers of plural trained source models are merged in parallel into the input stage of the second model (the transfer NN), which is then trained on the second sets of scatterometric data to predict the wafer pattern parameters — (Yacoby), col. 7, ll. ~9–13: "The k² encoder networks 310 … are combined in parallel into an input stage of a new neural network, referred herein as a 'transfer NN'"; col. 3, ll. ~12–22: "training a transfer neural network (NN) having initial layers including a parallel arrangement of the k² encoder neural networks … such that the transfer NN is trained to estimate new wafer pattern parameters").
(Yacoby) teaches something related to training plural machine learning models on sets of scatterometric data and reusing (transferring) their layers into the second model ((Yacoby), col. 6, ll. ~30–47 and FIG. 3A: "k² auto-encoder NNs are trained"). However, (Yacoby) does not teach "obtaining a third set of metrology data for a third one or more structures" and "training a third machine learning model for the third one or more structures using the third set of metrology data," nor does (Yacoby) expressly teach that one of the plural source models transferred to the second model is "the third machine learning model" trained for "a third one or more structures."
In the same field of endeavor, (Honda) teaches "obtaining a third set of metrology data for a third one or more structures" and "training a third machine learning model for the third one or more structures using the third set of metrology data." Specifically, (Honda) teaches identifying the targets of interest of a semiconductor manufacturing process — "variables relating to specific features of the semiconductor device" ((Honda), ¶ [0046]) — and teaches that "a plurality of ML models are used to predict the target(s) using the current training set data," including "new models created for this purpose," wherein "a variety of different types of models … can be employed on the basis that an evaluation of all the different predictions of the various different models may provide a better overall prediction" ((Honda), ¶ [0047]). (Honda) thereby teaches obtaining metrology/training data for additional (e.g., third) structures or targets and training an additional (third) machine learning model on that data, in addition to the first machine learning model.
(Yacoby) and (Honda) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based characterization and prediction of parameters of semiconductor structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the multiple-source transfer-NN architecture of (Yacoby) — in which the layers of plural trained source models are merged into the second model — with the plurality of models trained for additional structures/targets of (Honda), such that the source models merged into the second model include both the first machine learning model and a third machine learning model trained for a third one or more structures, thereby performing transfer learning from the third machine learning model with the first machine learning model to the second set of metrology data to produce the second machine learning model. The motivation to combine (Yacoby) and (Honda) is as recited by (Honda), which teaches that employing and evaluating a plurality of different models "may provide a better overall prediction" for the target features ((Honda), ¶ [0047]), such that incorporating an additional (third) model trained for an additional structure into (Yacoby)'s transfer-NN would improve prediction accuracy and robustness across multiple structures, with a reasonable expectation of success because both references train machine learning models to predict parameters/features of semiconductor structures.
Regarding claim 7, the limitations of claim 6 are rejected under the same rationale set forth above for claim 6, the combination of (Yacoby) and (Honda) teaching the method of claim 6.
Regarding the limitation added by claim 7, the claim adds:
"wherein different layers from the first machine learning model and the third machine learning model are transferred to the second machine learning model."
(Yacoby) teaches that particular, selected layers of the trained source models — specifically the encoder networks, comprising the input layer, the hidden layers, and the internal bottleneck (code) layer — are transferred into the second model (the transfer NN), while other layers (the decoder networks, hidden layers 350 leading to output layer 360) are not transferred and new final layers are trained, and that the transferred layers are drawn from a parallel arrangement of plural source models ((Yacoby), col. 7, ll. ~44–53: "The initial layers of the transfer NN 400 are the merged encoder networks 310 of the auto-encoder NNs 300, i.e., with merged input layers 320, followed by hidden layers 330 … leading to the internal 'code' layer (or 'bottleneck layer') 340"; "training the transfer NN creates a mapping from the bottleneck layer 340 to the output layer 410"; col. 3, ll. ~12–17).
(Yacoby) teaches something related to transferring different, selected layers from plural source machine learning models into the second machine learning model. However, (Yacoby) does not teach "wherein different layers from the first machine learning model and the third machine learning model are transferred to the second machine learning model," in that (Yacoby) does not teach that one of the plural source models from which layers are transferred is a third machine learning model trained for a third one or more structures.
In the same field of endeavor, (Honda) teaches obtaining metrology/training data for a third one or more structures and training a third machine learning model on that data, in addition to the first machine learning model, by teaching that "a plurality of ML models are used to predict the target(s) using the current training set data," including "new models created for this purpose," and that "a variety of different types of models … can be employed on the basis that an evaluation of all the different predictions of the various different models may provide a better overall prediction" ((Honda), ¶¶ [0046]–[0047]). (Honda) thereby teaches the existence of a third machine learning model trained for a third one or more structures, from which layers may be transferred together with layers from the first machine learning model.
(Yacoby) and (Honda) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based characterization and prediction of parameters of semiconductor structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the selective layer-transfer architecture of (Yacoby) — in which selected, different layers (the encoder networks) of plural trained source models are merged into the second model — with the plurality of models trained for additional structures of (Honda), such that different layers from the first machine learning model and a third machine learning model are transferred to the second machine learning model. The motivation to combine (Yacoby) and (Honda) is as recited by (Honda), which teaches that employing and evaluating a plurality of different models "may provide a better overall prediction" for the target features ((Honda), ¶ [0047]), such that transferring selected layers from both the first and a third model into (Yacoby)'s transfer NN would improve prediction accuracy and robustness across multiple structures, with a reasonable expectation of success because both references train machine learning models to predict parameters/features of semiconductor structures.
Regarding claim 8, the limitations of claim 1 are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching the method of claim 1. Claim 8 adds a limitation addressed by its sub-limitations below:
"obtaining a third set of metrology data for a third one or more structures"; and
"performing transfer learning from the first machine learning model and the second machine learning model to the third set of metrology data to produce a third machine learning model for predicting key parameters for the third one or more structures."
Regarding claim 8, (Yacoby) teaches:
"performing transfer learning from the … machine learning model and the … machine learning model to the third set of metrology data to produce a third machine learning model for predicting key parameters for the third one or more structures" ((Yacoby) teaches forming a transfer neural network whose initial layers are the merged, pretrained layers of plural prior source models, and training that transfer NN on a further set of scatterometric data and corresponding reference parameters to produce a model that estimates the wafer pattern parameters — (Yacoby), col. 7, ll. ~9–13: "The k² encoder networks 310 … are combined in parallel into an input stage of a new neural network, referred herein as a 'transfer NN'"; col. 3, ll. ~12–22: "training a transfer neural network (NN) having initial layers including a parallel arrangement of the … encoder neural networks … such that the transfer NN is trained to estimate new wafer pattern parameters from subsequently measured sets of scatterometric data").
(Yacoby) teaches something related to producing a further machine learning model by transferring the layers of plural prior trained models to a further set of scatterometric data and training the model to predict the pattern parameters. However, (Yacoby) does not teach "obtaining a third set of metrology data for a third one or more structures" and "performing transfer learning from the first machine learning model and the second machine learning model to the third set of metrology data to produce a third machine learning model for predicting key parameters for the third one or more structures," in that (Yacoby) does not teach a distinct third one or more structures, a corresponding third set of metrology data, or that the prior models whose layers are transferred to the further model are the first machine learning model and the second machine learning model.
In the same field of endeavor, (Honda) teaches "obtaining a third set of metrology data for a third one or more structures" and the use of plural prior models to produce a further model for predicting parameters of the structures. Specifically, (Honda) teaches identifying the targets of interest — "variables relating to specific features of the semiconductor device" ((Honda), ¶ [0046]) — and teaches that "a plurality of ML models are used to predict the target(s) using the current training set data," including "new models created for this purpose," wherein "an evaluation of all the different predictions of the various different models may provide a better overall prediction" ((Honda), ¶ [0047]). (Honda) thereby teaches obtaining metrology/training data for an additional (third) one or more structures and producing an additional (third) machine learning model for predicting the parameters of that additional structure, building upon plural prior models (the first and second machine learning models).
(Yacoby) and (Honda) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based characterization and prediction of parameters of semiconductor structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the multiple-source transfer-NN architecture of (Yacoby) — in which the layers of plural prior trained models are merged to produce a further model that predicts pattern parameters — with the plurality of models trained for additional structures of (Honda), such that transfer learning is performed from the first machine learning model and the second machine learning model to a third set of metrology data to produce a third machine learning model for predicting key parameters for a third one or more structures. The motivation to combine (Yacoby) and (Honda) is as recited by (Honda), which teaches that employing and evaluating a plurality of different models "may provide a better overall prediction" for the target features ((Honda), ¶ [0047]), such that producing a third model from the first and second models, as in (Yacoby)'s transfer-NN framework, would extend accurate parameter prediction to an additional structure while reducing the reference data and re-training otherwise required, with a reasonable expectation of success because both references train machine learning models to predict parameters/features of semiconductor structures.
Regarding claim 9, the limitations of claim 8 are rejected under the same rationale set forth above for claim 8, the combination of (Yacoby) and (Honda) teaching the method of claim 8.
Regarding the limitation added by claim 9, the claim adds:
"wherein different layers from the first machine learning model and the second machine learning model are transferred to the third machine learning model."
(Yacoby) teaches that particular, selected layers of the trained source models — specifically the encoder networks, comprising the input layer, the hidden layers, and the internal bottleneck (code) layer — are transferred into the further model (the transfer NN), while other layers (the decoder networks, hidden layers 350 leading to output layer 360) are not transferred and new final layers are trained, and that the transferred layers are drawn from a parallel arrangement of plural source models ((Yacoby), col. 7, ll. ~44–53: "The initial layers of the transfer NN 400 are the merged encoder networks 310 of the auto-encoder NNs 300, i.e., with merged input layers 320, followed by hidden layers 330 … leading to the internal 'code' layer (or 'bottleneck layer') 340"; "training the transfer NN creates a mapping from the bottleneck layer 340 to the output layer 410"; col. 3, ll. ~12–17).
(Yacoby) teaches something related to transferring different, selected layers from plural source machine learning models into a further machine learning model. However, (Yacoby) does not teach "wherein different layers from the first machine learning model and the second machine learning model are transferred to the third machine learning model," in that (Yacoby) does not teach a distinct third machine learning model for a third one or more structures into which the selected layers of the first and second machine learning models are transferred.
In the same field of endeavor, (Honda) teaches obtaining metrology/training data for a third one or more structures and producing a third machine learning model for that structure, building upon plural prior models, by teaching that "a plurality of ML models are used to predict the target(s) using the current training set data," including "new models created for this purpose," and that "an evaluation of all the different predictions of the various different models may provide a better overall prediction" ((Honda), ¶¶ [0046]–[0047]). (Honda) thereby teaches the third machine learning model into which selected layers from the first and second machine learning models may be transferred.
(Yacoby) and (Honda) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based characterization and prediction of parameters of semiconductor structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the selective layer-transfer architecture of (Yacoby) — in which selected, different layers (the encoder networks) of plural trained source models are merged into a further model — with the plurality of models trained for additional structures of (Honda), such that different layers from the first machine learning model and the second machine learning model are transferred to the third machine learning model. The motivation to combine (Yacoby) and (Honda) is as recited by (Honda), which teaches that employing and evaluating a plurality of different models "may provide a better overall prediction" for the target features ((Honda), ¶ [0047]), such that transferring selected layers from both the first and second models into a third model, as in (Yacoby)'s transfer-NN framework, would extend accurate parameter prediction to an additional structure while reducing the reference data and re-training otherwise required, with a reasonable expectation of success because both references train machine learning models to predict parameters/features of semiconductor structures.
Claim 15 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 6.
Claim 16 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 7.
Claim 17 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 8.
Claim 18 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 9.
Claims 19, 20, 22, 24-28, 30, and 32- 34 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2, in view of Pandev et al. (Pandev), US 2021/0109453 A1.
Regarding claim 19, the limitations of "obtaining a first set of metrology data for a first one or more structures" and "obtaining a second set of metrology data for a second one or more structures" are rejected under the same rationale set forth above for claim 1, (Yacoby) teaching these limitations ((Yacoby), col. 5, ll. ~62–64 and FIG. 2, step 214; col. 3, ll. ~9–12).
Regarding the limitations added by claim 19, (Yacoby) teaches:
"training a machine learning model with selected metrology data for predicting key parameters for the second one or more structures" ((Yacoby), col. 7, ll. ~9–35: the transfer NN is "trained to estimate new wafer pattern parameters from subsequently measured sets of scatterometric data"; col. 7, ll. ~56–60: output parameters, e.g., "height, width, and pitch of a given wafer stack").
(Yacoby) teaches something related to a feature extractor, in that (Yacoby) teaches encoder neural networks each having an input layer, hidden layers, and an internal bottleneck (code) layer that encode the scatterometric sub-vectors into a latent representation ((Yacoby), col. 6, ll. ~60–67 and FIG. 3A: "An encoder network 310 of the auto-encoder NN includes an input layer 320 and hidden layers 330, leading to an internal 'code' or 'latent representation' layer 340"). However, (Yacoby) does not teach "selecting metrology data from the first set of metrology data and the second set of metrology data using a feature extractor."
In the same field of endeavor, (Pandev) teaches this limitation. Specifically, (Pandev) teaches obtaining first metrology data for a first one or more structures (a device area) and second metrology data for a second one or more structures (a metrology target containing structures distinct from the device area), and using both sets to train a metrology-data prediction model that performs signal-domain adaptation between the two domains ((Pandev), ¶ [0023]: "first metrology data 302" and "second metrology data 304"; ¶ [0028]: "Using the first metrology data 302 and the second metrology data 304" to train "a metrology-data prediction model 306," which "is a neural network"). (Pandev) further teaches that the signal-domain adaptation selects and converts metrology data across the two domains so that features common to the first and second metrology data are used by the prediction model ((Pandev), ¶ [0019]: "signal-domain adaptation is used to convert the metrology" data across domains; ¶ [0030]: the second metrology data 304 serves as ground truth against which the model's prediction from the first metrology data is matched). (Pandev) thereby teaches selecting metrology data from the first and second sets using a feature extractor that performs domain adaptation.
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the encoder-based (feature-extractor) machine-learning metrology of (Yacoby) with the signal-domain-adaptation feature selection across first and second metrology data of (Pandev), such that metrology data is selected from the first set and the second set using a feature extractor and a machine learning model is trained with the selected data to predict the key parameters of the second one or more structures. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that selecting and adapting metrology data across the first and second domains enables accurate prediction of metrology data for a target structure whose distribution differs from the source ((Pandev), ¶¶ [0028]–[0030]), thereby reducing the need to separately measure or re-train for the target structure, with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
Regarding claim 20, the limitations of claim 19 are rejected under the same rationale set forth above for claim 19, the combination of (Yacoby) and (Pandev) teaching the method of claim 19.
Regarding the limitation added by claim 20, the claim adds:
"further comprising minimizing domain differences between the first set of metrology data and the second set of metrology data."
(Yacoby) teaches something related to reducing differences between sets of metrology data through joint training, in that (Yacoby) teaches training the transfer NN by minimizing a mean squared error loss with respect to the reference parameters corresponding to the sets of scatterometric data ((Yacoby), col. 7, ll. ~31–35: "the loss function that the transfer NN is trained to minimize is a mean squared error (MSE) loss function"). However, (Yacoby) does not teach "minimizing domain differences between the first set of metrology data and the second set of metrology data."
In the same field of endeavor, (Pandev) teaches "minimizing domain differences between the first set of metrology data and the second set of metrology data." Specifically, (Pandev) teaches that signal-domain adaptation is performed to convert and align the metrology data across the first and second domains, training the prediction model so that metrology data predicted from the first metrology data matches the second metrology data, thereby minimizing the difference between the two domains ((Pandev), ¶ [0019]: "signal-domain adaptation is used to convert the metrology" data across domains; ¶ [0028]: "Using the first metrology data 302 and the second metrology data 304" to train the prediction model; ¶ [0030]: the second metrology data 304 "serves as" the ground truth that the prediction from the first metrology data is trained to match). (Pandev) thereby teaches minimizing domain differences between the first set of metrology data and the second set of metrology data.
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the machine-learning metrology of (Yacoby) with the signal-domain adaptation of (Pandev) so as to minimize domain differences between the first set of metrology data and the second set of metrology data. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that minimizing the difference between the source and target metrology domains enables accurate prediction of metrology data for a target structure whose distribution differs from the source ((Pandev), ¶¶ [0028]–[0030]), thereby reducing the need to separately measure or re-train for the target structure, with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
Regarding claim 22, the limitations of claim 19 are rejected under the same rationale set forth above for claim 19, the combination of (Yacoby) and (Pandev) teaching the method of claim 19.
Regarding the limitation added by claim 22, the claim adds:
"further comprising minimizing domain differences by co-training based on the first set of metrology data and the second set of metrology data."
(Yacoby) teaches something related to minimizing differences between sets of metrology data through joint training, in that (Yacoby) teaches training a single neural network on the combined feature input derived from the sets of scatterometric data by minimizing a mean squared error loss with respect to the corresponding reference parameters ((Yacoby), col. 7, ll. ~14–35: the feature input to the transfer NN is formed from the sub-vectors of the sets of scatterometric data, and "the loss function that the transfer NN is trained to minimize is a mean squared error (MSE) loss function"). However, (Yacoby) does not teach "minimizing domain differences by co-training based on the first set of metrology data and the second set of metrology data."
In the same field of endeavor, (Pandev) teaches "minimizing domain differences by co-training based on the first set of metrology data and the second set of metrology data." Specifically, (Pandev) teaches jointly using both the first metrology data and the second metrology data to train a single metrology-data prediction model, such that the model is trained on both domains together (co-training) to align the source and target metrology domains and minimize the difference between them ((Pandev), ¶ [0028]: "Using the first metrology data 302 and the second metrology data 304" to train "a metrology-data prediction model 306"; ¶ [0030]: the second metrology data 304 "serves as" the ground truth against which the model's prediction from the first metrology data is matched, so that training on both sets together drives the two domains into agreement). (Pandev) thereby teaches minimizing domain differences by co-training based on the first set of metrology data and the second set of metrology data.
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the joint-training machine-learning metrology of (Yacoby) with the co-training-based signal-domain adaptation of (Pandev), such that domain differences are minimized by co-training based on the first set of metrology data and the second set of metrology data. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that co-training on the first and second metrology data to align the source and target domains enables accurate prediction of metrology data for a target structure whose distribution differs from the source ((Pandev), ¶¶ [0028]–[0030]), thereby reducing the need to separately measure or re-train for the target structure, with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
Regarding claim 24, the limitations of claim 19 are rejected under the same rationale set forth above for claim 19, the combination of (Yacoby) and (Pandev) teaching the method of claim 19.
Regarding the limitation added by claim 24, the claim adds:
"wherein one set of metrology data comprising either the first set of metrology data or the second set of metrology data is at least partially labeled"; and
"and a remaining set of metrology data is labeled, unlabeled, or a combination thereof."
(Yacoby) teaches something related to labeling and unlabeled training of the sets of metrology data, in that (Yacoby) teaches training the auto-encoder neural networks on sets of scatterometric data "in a self-supervised manner" (i.e., without labels) ((Yacoby), col. 6, ll. ~30–37), and separately teaches that sets of scatterometric data are paired with reference parameters used as the target output, i.e., labels, for training ((Yacoby), col. 5, ll. ~30–37: "Reference parameters 44 may be used as target output for ML training"). However, (Yacoby) does not teach "wherein one set of metrology data comprising either the first set of metrology data or the second set of metrology data is at least partially labeled and a remaining set of metrology data is labeled, unlabeled, or a combination thereof."
In the same field of endeavor, (Pandev) teaches this limitation. Specifically, (Pandev) teaches a signal-domain-adaptation framework in which one set of metrology data is at least partially labeled and serves as the ground truth, while the remaining set need not be fully labeled, in that the second metrology data serves as the labeled ground truth against which the model's prediction from the first metrology data is trained ((Pandev), ¶ [0030]: the second metrology data 304 "serves as" the ground truth that the prediction model is trained to match), and the first metrology data is used as the input from which metrology data is predicted ((Pandev), ¶ [0028]: "Using the first metrology data 302 and the second metrology data 304" to train the prediction model). (Pandev) thereby teaches that one set of metrology data (the second set) is at least partially labeled and a remaining set of metrology data (the first set) is labeled, unlabeled, or a combination thereof.
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-supervised-and-labeled machine-learning metrology of (Yacoby) with the partially-labeled signal-domain adaptation of (Pandev), such that one set of metrology data is at least partially labeled and a remaining set of metrology data is labeled, unlabeled, or a combination thereof. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that using one at-least-partially-labeled set as the ground truth while allowing the remaining set to be unlabeled enables accurate prediction of metrology data for a target structure without requiring full labeling of both domains ((Pandev), ¶¶ [0028]–[0030]), thereby reducing the labeling and measurement burden, with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
Regarding Claim 25, the limitations of claim 19 are rejected under the same rationale set forth above for claim 19. The limitation added by claim 25 — "wherein the first one or more structures and the second one or more structures are different types of structures or are a same type of structures produced using a same or different processes" — is taught by (Yacoby) under the same rationale set forth above for the corresponding limitation of claim 5.
Regarding claim 26, the limitations of claim 19 are rejected under the same rationale set forth above for claim 19, the combination of (Yacoby) and (Pandev) teaching the method of claim 19.
Regarding the limitations added by claim 26:
(Yacoby) teaches something related to obtaining additional sets of metrology data and to selecting metrology data across plural sets using a feature extractor, in that (Yacoby) teaches receiving multiple sets of scatterometric data and encoding the sub-vectors of those sets through the encoder networks (feature extractor) whose layers are merged in parallel for training ((Yacoby), col. 5, ll. ~62–64 and FIG. 2, step 214; col. 6, ll. ~60–67 and FIG. 3A; col. 7, ll. ~9–13). However, (Yacoby) does not teach the limitations added by claim 26. In the same field of endeavor, (Pandev) teaches these limitations, broken down into their sub-limitations below.
With respect to "obtaining a third set of metrology data for a third one or more structures," (Pandev) teaches obtaining third metrology data for an instance of a structure that is distinct from the structures associated with the first and second metrology data ((Pandev), ¶ [0032]: "Third metrology data 308 is obtained (216) for an instance of the device area on a first semiconductor die that is distinct" from the prior die). (Pandev) thereby teaches obtaining a third set of metrology data for a third one or more structures.
With respect to "selecting metrology data from the third set of metrology data with the first set of metrology data and the second set of metrology data using the feature extractor," (Pandev) teaches applying the trained metrology-data prediction model — which embodies the signal-domain feature selection learned from the first and second metrology data — to the third metrology data to predict metrology data for the target ((Pandev), ¶ [0033]: "Using the trained machine-learning model," fourth metrology data is predicted for the metrology target based on the third metrology data; ¶ [0028]: the model is trained "Using the first metrology data 302 and the second metrology data 304"). Because the trained model embodies the feature extractor that selects and adapts metrology data across the first and second sets, applying it to the third set selects metrology data from the third set together with the first and second sets using the feature extractor. (Pandev) thereby teaches selecting metrology data from the third set of metrology data with the first set of metrology data and the second set of metrology data using the feature extractor.
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the feature-extractor-based machine-learning metrology of (Yacoby) with the third-metrology-data feature selection of (Pandev), such that metrology data is selected from the third set together with the first and second sets using the feature extractor. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that applying the trained feature-selecting model to additional metrology data enables accurate prediction of metrology data for additional structures whose distribution differs from the source without requiring separate measurement or re-training for each structure ((Pandev), ¶¶ [0032]–[0033]), with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
Claim 27 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 19. Claim 27 recites a computer system comprising at least one processor configured to perform the steps of claim 19 ((Yacoby), col. 3, ll. ~36–44 and claim 8).
Claim 28 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 20.
Claim 30 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 22.
Claim 32 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 24.
Claim 33 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 25.
Claim 34 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 26.
Claims 21 and 29 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2, in view of Pandev et al. (Pandev), US 2021/0109453 A1, and further in view of Ganin et al. (Ganin), "Domain-Adversarial Training of Neural Networks," Journal of Machine Learning Research 17 (published 26 May 2016), pp. 1–35
Regarding claim 21, the limitations of claim 20 are rejected under the same rationale set forth above for claim 20, the combination of (Yacoby) and (Pandev) teaching the method of claim 20.
Regarding the limitation added by claim 21, the claim adds:
"wherein the domain differences are minimized using a domain classifier"; and
"via a gradient reversal layer."
(Yacoby) teaches something related to minimizing differences between sets of metrology data through joint training, in that (Yacoby) teaches training the transfer NN by minimizing a mean squared error loss with respect to the reference parameters corresponding to the sets of scatterometric data ((Yacoby), col. 7, ll. ~31–35: "the loss function that the transfer NN is trained to minimize is a mean squared error (MSE) loss function"). However, (Yacoby) does not teach "wherein the domain differences are minimized using a domain classifier."
In the same field of endeavor, (Pandev) teaches "wherein the domain differences are minimized using a domain classifier." Specifically, (Pandev) teaches performing the signal-domain adaptation using a Cycle Generative Adversarial Network (Cycle GAN) having a generator and a discriminator, wherein the discriminator determines whether metrology data corresponds to the metrology target domain, i.e., a domain classifier that drives the minimization of the difference between the source and target metrology domains ((Pandev), ¶ [0031]: "Cycle GAN involves two models, a generator and a discriminator"; "the discriminator determines whether metrology data for the … metrology target" domain, and "the discriminator receives the second metrology data 304").
(Yacoby) and (Pandev) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based semiconductor metrology and prediction of structure parameters from metrology data. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the machine-learning metrology of (Yacoby) with the domain-classifier-based signal-domain adaptation of (Pandev), such that the domain differences between the first set of metrology data and the second set of metrology data are minimized using a domain classifier. The motivation to combine (Yacoby) and (Pandev) is as recited by (Pandev), which teaches that using the discriminator (domain classifier) to align the source and target metrology domains enables accurate prediction of metrology data for a target structure whose distribution differs from the source ((Pandev), ¶¶ [0028]–[0031]), thereby reducing the need to separately measure or re-train for the target structure, with a reasonable expectation of success because both references train machine learning models to predict parameters of semiconductor structures from metrology data.
The combination of (Yacoby) and (Pandev), however, does not teach "via a gradient reversal layer."
In the same field of endeavor, (Ganin) teaches "via a gradient reversal layer." Specifically, (Ganin) teaches a domain-adaptation neural network in which a feature extractor, a label predictor, and a domain classifier are trained jointly, and in which the domain differences are minimized by augmenting the network with "a new gradient reversal layer" that, during backpropagation, reverses the gradient from the domain classifier so as to maximize the domain-classifier loss and thereby promote domain-invariant features ((Ganin), p. 1, Abstract: the approach is implemented "by augmenting it with few standard layers and a new gradient reversal layer"; p. 2, § 1: "the only non-standard component of the proposed architecture is a rather trivial gradient reversal layer that leaves the input unchanged during forward propagation and reverses the gradient … during the backpropagation"). (Ganin) thereby teaches minimizing the domain differences using a domain classifier via a gradient reversal layer.
(Yacoby), (Pandev), and (Ganin) are analogous to the claimed invention as all are from the same field of endeavor of machine learning using feature extraction and domain adaptation to train models that make predictions across differing data domains. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the machine-learning metrology of (Yacoby) and the domain-classifier-based signal-domain adaptation of (Pandev) with the gradient reversal layer of (Ganin), such that the domain differences between the first set of metrology data and the second set of metrology data are minimized using a domain classifier via a gradient reversal layer. The motivation to combine (Yacoby), (Pandev), and (Ganin) is as recited by (Ganin), which teaches that the gradient reversal layer provides a simple and standard-backpropagation-trainable mechanism to make the learned features domain-invariant and thereby improve prediction performance on a target domain that differs from the source domain ((Ganin), p. 1, Abstract; p. 2, § 1), such that incorporating the gradient reversal layer into the domain-adaptation metrology of (Yacoby) and (Pandev) would improve the accuracy of parameter prediction for the second structures, with a reasonable expectation of success because each reference trains neural-network models using feature extraction and domain adaptation.
Claim 29 is rejected under 35 U.S.C. 103 under the same rationale set forth above for claim 21.
Claims 23 and 31 are rejected under 35 U.S.C. 103 as being unpatentable over Yacoby et al. (Yacoby), US 11,747,740 B2, in view of Pandev et al. (Pandev), US 2021/0109453 A1, and further in view of Bhaskar et al. (Bhaskar), US 2017/0193400 A1, cited in the IDS filed 7/17/2024.
Regarding claim 23, the limitations of claim 9 are rejected under the same rationale set forth above for claim 9, (Yacoby) teaching the method of claim 9.
Regarding the limitations added by claim 23, (Yacoby) teaches:
"experimental metrology data generated from the first one or more structures" ((Yacoby), col. 5, ll. ~62–64 and FIG. 2, step 214: "receiving multiple sets of scatterometric data, measured from respective wafer patterns"; col. 5, ll. ~14–18: "The scatterometry data 32 generated by the metrology system 30"); and
"experimental metrology data generated from the second one or more structures" ((Yacoby), col. 3, ll. ~9–12: "multiple corresponding second sets of scatterometric data, measured from multiple respective wafer patterns").
(Yacoby) teaches something related to generating estimated scatterometry data from an optical model of a structure having known parameters ((Yacoby), col. 2, ll. ~8–17: optical models "can … be applied to generate, from a set of known pattern parameters, an estimate of scatterometry data that would be measured during spectrographic testing"). However, (Yacoby) does not teach "synthetic metrology data generated from one or more models of the first one or more structures … and synthetic metrology data generated from one or more models of the second one or more structures."
In the same field of endeavor, (Bhaskar) teaches this limitation. (Bhaskar) teaches a machine-learning-based model "configured for performing simulation(s) for the specimens" ((Bhaskar), Abstract), and teaches generating the training input by "empirical simulation of real defect events on wafers and reticles using DOEs" ((Bhaskar), col. 16, ll. ~14–17) and by models that generate the instance information used for training ((Bhaskar), col. 15–16), thereby teaching synthetic metrology data generated from one or more models of the structures. Combined with the experimental metrology data taught by (Yacoby), (Bhaskar) also satisfies the "or a combination thereof" alternative recited for each set.
(Yacoby) and (Bhaskar) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based optical metrology and characterization of semiconductor wafer structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-supervised-and-transfer-learning OCD-metrology method of (Yacoby) with the model-generated (synthetic) training data of (Bhaskar), such that the first and second sets of metrology data each comprise synthetic metrology data generated from one or more models of the respective structures, experimental metrology data, or a combination thereof. The motivation to combine (Yacoby) and (Bhaskar) is as recited by (Bhaskar), which teaches that generating training information by simulation allows the model to be adequately trained where actual specimen data is limited or where variation/defect instances are rare, thereby reducing the amount of measured data required ((Bhaskar), Abstract; col. 16–17), with a reasonable expectation of success because both references train machine learning models on metrology data of semiconductor wafer structures.
Regarding claim 31, the limitations of claim 27 are rejected under the same rationale set forth above for claim 27, (Yacoby) teaching the claim 27.
Regarding the limitations added by claim 31, (Yacoby) teaches:
"experimental metrology data generated from the first one or more structures" ((Yacoby), col. 5, ll. ~62–64 and FIG. 2, step 214: "receiving multiple sets of scatterometric data, measured from respective wafer patterns"; col. 5, ll. ~14–18: "The scatterometry data 32 generated by the metrology system 30"); and
"experimental metrology data generated from the second one or more structures" ((Yacoby), col. 3, ll. ~9–12: "multiple corresponding second sets of scatterometric data, measured from multiple respective wafer patterns").
(Yacoby) teaches something related to generating estimated scatterometry data from an optical model of a structure having known parameters ((Yacoby), col. 2, ll. ~8–17: optical models "can … be applied to generate, from a set of known pattern parameters, an estimate of scatterometry data that would be measured during spectrographic testing"). However, (Yacoby) does not teach "synthetic metrology data generated from one or more models of the first one or more structures … and synthetic metrology data generated from one or more models of the second one or more structures."
In the same field of endeavor, (Bhaskar) teaches this limitation. (Bhaskar) teaches a machine-learning-based model "configured for performing simulation(s) for the specimens" ((Bhaskar), Abstract), and teaches generating the training input by "empirical simulation of real defect events on wafers and reticles using DOEs" ((Bhaskar), col. 16, ll. ~14–17) and by models that generate the instance information used for training ((Bhaskar), col. 15–16), thereby teaching synthetic metrology data generated from one or more models of the structures. Combined with the experimental metrology data taught by (Yacoby), (Bhaskar) also satisfies the "or a combination thereof" alternative recited for each set.
(Yacoby) and (Bhaskar) are analogous to the claimed invention as both are from the same field of endeavor of machine-learning-based optical metrology and characterization of semiconductor wafer structures. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the self-supervised-and-transfer-learning OCD-metrology method of (Yacoby) with the model-generated (synthetic) training data of (Bhaskar), such that the first and second sets of metrology data each comprise synthetic metrology data generated from one or more models of the respective structures, experimental metrology data, or a combination thereof. The motivation to combine (Yacoby) and (Bhaskar) is as recited by (Bhaskar), which teaches that generating training information by simulation allows the model to be adequately trained where actual specimen data is limited or where variation/defect instances are rare, thereby reducing the amount of measured data required ((Bhaskar), Abstract; col. 16–17), with a reasonable expectation of success because both references train machine learning models on metrology data of semiconductor wafer structures.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HUNG VAN LE whose telephone number is (571)270-0164. The examiner can normally be reached 8 a.m. - 5 p.m..
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, Cesar Paula can be reached at (571) 272-4128. 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.
/HUNG VAN LE/Examiner, Art Unit 2145
/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights