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 .
The office action is in response to the application filed on April 05, 2024.
Claims 1-20 are pending and have been examined. Claims 1-20 are rejected.
Information Disclosure Statement
Acknowledgment is made of the information disclosure statements filed July 09, 2024, which comply with 37 CFR 1.97. As such, the information disclosure statements have been placed in the application file and the information referred to therein has been considered by the examiner.
Priority
Applicant's claim for the benefit of a prior-filed application under 35 U.S.C. 119(e) or under 35 U.S.C. 120, 121, 365(c), or 386(c) is acknowledged. The present application claims priority to U.S. Provisional application No. 63/457,618 (hereinafter "the '618 provisional application") filed on April 06, 2023.
The as-filed original specification of the '618 provisional application provides adequate support or enablement for all limitation elements of claims 1-20.
Therefore, the effective filing date for claims 1-20 of the instant application is the effective filing date of the provisional application, April 06, 2023.
Claim Interpretation
The following is a quotation of 35 U.S.C. 112(f):
(f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The following is a quotation of pre-AIA 35 U.S.C. 112, sixth paragraph:
An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof.
The claims in this application are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
As explained in MPEP § 2181, subsection I, claim limitations that meet the following three-prong test will be interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph:
(A) the claim limitation uses the term “means” or “step” or a term used as a substitute for “means” that is a generic placeholder (also called a nonce term or a non-structural term having no specific structural meaning) for performing the claimed function;
(B) the term “means” or “step” or the generic placeholder is modified by functional language, typically, but not always linked by the transition word “for” (e.g., “means for”) or another linking word or phrase, such as “configured to” or “so that”; and
(C) the term “means” or “step” or the generic placeholder is not modified by sufficient structure, material, or acts for performing the claimed function.
Absence of the word “means” (or “step”) in a claim creates a rebuttable presumption that the claim limitation is not to be treated in accordance with 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph. The presumption that the claim limitation is not interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is rebutted when the claim limitation recites function without reciting sufficient structure, material or acts to entirely perform the recited function.
Claim limitations in this application that use the word “means” (or “step”) are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action. Conversely, claim limitations in this application that do not use the word “means” (or “step”) are not being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, except as otherwise indicated in an Office action.
This application includes one or more claim limitations that do not use the word “means,” but are nonetheless being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, because the claim limitation(s) uses a generic placeholder that is coupled with functional language without reciting sufficient structure to perform the recited function and the generic placeholder is not preceded by a structural modifier.
Such claim limitation(s) are given their broadest reasonable interpretation using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art. The broadest reasonable interpretation of a claim element (also commonly referred to as a claim limitation) is limited by the description in the specification when 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, is invoked.
Claim 10 recites "digital twin input generator", “digital twin model runner”, “learning model runner”, “cost function determiner”, and “backpropagator”.
Claim 12 recites an “iterator”.
A review of the disclosure (see Fig. 9 and paragraphs [0012]-[0014] of the specification) shows that the following appears to be the corresponding structure described in the specification for the 35 U.S.C. 112(f) or 35 U.S.C. 112 (pre-AIA ), sixth paragraph limitation:
Because this/these claim limitation(s) is/are being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, it/they is/are being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof.
If applicant does not intend to have this/these limitation(s) interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph, applicant may: (1) amend the claim limitation(s) to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitation(s) recite(s) sufficient structure to perform the claimed function so as to avoid it/them being interpreted under 35 U.S.C. 112(f) or pre-AIA 35 U.S.C. 112, sixth paragraph.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
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 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 non-obviousness.
Claims 1, 2, 10, 11, 12, 13, 19, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et. Al, (Kalibre: Knowledge-based Neural Surrogate Model Calibration for Data Center Digital Twins, published November 18, 2020, hereinafter "Wang"), in view of Ruff et. Al, (Surrogate Neural Networks for Efficient Simulation-based Trajectory Planning Optimization, published March 30, 2023). The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claims 1, 10, and 19:
Wang teaches:
“A method of training a learning model using a digital twin model, the method comprising: a learning model interfacing with the digital twin model within an iterated digital twin system: for each iteration of the iterated digital twin system:” ([Wang pg. 209, Sec. 6] discloses that the calibrated CFD model is expressly identified as the digital twin, “The surrogate does not provide a full-fledged approximation of the CFD model. For instance, it does not model the temperatures at the locations without sensors, which are modeled by the CFD model in contrast. Thus, only the calibrated CFD model shall be used as a digital twin for the run-time temperature evaluation in improving energy efficiency and reducing operational risks.” [Wang pg. 201, Fig. 1] discloses the iterated digital twin system as a four-step loop between the CFD model (digital twin) and the surrogate (learning model), “Kalibre operates by iterating ① training a CFD surrogate model, ② searching optimal parameters through surrogate retraining, ③ configuring the found parameters back to the CFD model, ④ validating the new configuration.”)
“generating digital twin input-----” ([Wang pg. 203, sec. 3.2] discloses the acquiring/generating of the digital twin input (CFD model input), “Definition 1 (Input). The input data for solving a CFD model is a vector consisting of all modeling parameters. Formally, the input x =(Tc,V,P,α),whereTc = (Tc1,Tc2,...,Tcl),V = (V1,V2,.. .,Vl), P = (P1,P2,...,Pm), and α = (α1,α2,...,αm) are the vectors of CRAC setpoints, CRAC fan speeds, server powers, and server air flow rates, respectively.” [Wang pg. 206, Algorithm 1, Steps 7-8] further discloses the functional generator component operating on the α component of its digital twin’s input, “Search α by performing differential evolution; Search α by performing gradient descent on L2;”) Examiner’s Note: With respect to Claim 10: Under broadest reasonable interpretation the “digital twin input generator” is essentially the functional structure/procedure that performs the generating of the digital twin input.
“running the digital twin model using the digital twin input, producing digital twin output;” ([Wang pg. 205, sec. 4.2] discloses the generated digital twin input being utilized to run the digital twin model, “The training data is generated by solving the CFD model with collected system input (including CRAC temperature setpoints and fan speeds, server powers) and the initial α or calibrated α by step ③ of the previous iterations to yield the predicted temperatures at sensor locations. The detailed training data generation is described in §5.2.2. Note that each element of α should be within [αl,αu]. The system input, the α,and the predicted temperatures form a new training data sample that is added to the training dataset accumulated from the first iteration.” [Wang pg. 206, Algorithm 1] further discloses the digital twin model (CFD model) being run on the digital twin inputs to produce outputs, “4: Solve CFD model to obtain Ts;” [Wang pg. 203] further states the characteristics of the produced digital twin outputs, “Definition 2 (Output). The output of CFD is a steady-state temperature and air velocity distribution map. For CFD model calibration, we focus on a set of results within the map at the lo cations installed with temperature sensors, which is denoted by ~Ts = (˜Ts1, ˜Ts2,..., ˜Tsn).”) Examiner’s Note: With respect to Claim 10: Under broadest reasonable interpretation the “digital twin model runner” is essentially the functional step/procedure that invokes the digital twin on a given input and captures its output.
“using the digital twin output as learning model ground truth and learning model output in a learning model cost function;” ([Wang pg. 204, sec. 3.3, Equation 2] discloses a cost function pairing the digital twin’s (CFD model’s output ~Ts) against the learning model’s (surrogate’s output ^Ts), “W∗ ≜ argmin W ||~Ts(x) − ^Ts(W, x)||2 2 where W is a set of trainable weights of the surrogate and W∗ is the result of the surrogate training. Second, with W∗, the surrogate is re-optimized through re-training such that the discrepancy between the surrogate’s output and the measurement is minimized.” [Wang pg. 204, sec. 3.3] further discloses that the digital twin’s output functions as the target (ground truth) that the learning model is trained against, “the surrogate model is trained to be locally aligned with the CFD model by minimizing the discrepancy between the surrogate’s and the CFD’s outputs:”) Examiner’s Note: With respect to Claim 10: Under broadest reasonable interpretation the functional step/procedure/operation of Equation 2 is essentially the “cost function determiner” component that takes the digital twin’s outputs (~Ts) as the learning model ground truth and the learning model’s outputs (^Ts) as the comparison term and computes the cost (the squared discrepancy) between them.
“and using a cost derived from the learning model cost function to ----------- through the learning model as a part of improving values of parameters of the learning model.” ([Wang pg. 205, sec. 4.2] discloses the learning model’s (neural surrogate’s) parameters W being updated from the gradient of the cost function to improve outcomes/parameters, “the neural surrogate is updated to minimize the errors between its predicted temperatures and the CFD-predicted temperatures of the training samples. Thus, the weights of the neural surrogate are updated using the gradient of the least squares loss function of L1 = 1 n n k=1 (Tsk(W,x) − Tsk(x))2. As a result, the surrogate is trained to align with the CFD model.”) Examiner’s Note: With respect to Claim 10: Under broadest reasonable interpretation the “backpropagator” is essentially the functional step/procedure and weight-update mechanism.
Wang alone does not appear to explicitly disclose:
“------using an optimization process;”
“running the learning model using the digital twin input, producing learning model output;”
“---backpropagate---”
However, Ruff teaches:
“------using an optimization process;” ([pg. 3, sec. IV. Ruff] Algorithm 1, Step 4 discloses an optimization process (stochastic gradient descent) generating the candidate input that is later fed to the digital twin, “Using stochastic gradient descent to find the set xlocal that locally minimizes g(ˆf(x), y target).” [pg. 3, sec. IV. Ruff ] Algorithm 1, step 6 further teaches that the same optimizer-selected value is used to run the digital twin simulation, “Intelligently query f(x) with the selected xbest.”)
“running the learning model using the digital twin input, producing learning model output;” ([Ruff pg. 2, sec. III-B] discloses the learning model (surrogate ^f(x)) being run on the same input as the digital twin to produce learning model outputs, “a neural network is trained to be a surrogate model ˆf(x) for the real simulation. This surrogate model is able to predict n dimensional outputs based on real m-dimensional inputs ˆy = ˆf(x) where ˆy is the predicted outputs from the surrogate model.”) Examiner’s Note: ^f(x) is the learning model and ^y is its output computed from the identical x that also drives the digital twin f(x) which is a reference trajectory-based digital twin simulation model [pg. 3]. With respect to Claim 10: Under broadest reasonable interpretation the functional step/procedure of evaluating ^f(x) to produce ^y is essentially the “learning model runner” which is executing the trained surrogate on the shared input to generate its prediction.
“---backpropagate---” ([Ruff pg. 4, sec. III-C] discloses backpropagation improving the learning model’s (surrogate network’s) parameters from the cost, “the data must be standardized for training a neural network to be the best possible surrogate model to accurately represent the simulation…The weights of the network are tuned to reduce the MAE using a variant of stochastic gradient descent with backpropagated gradients. The model can be further improved by tuning the hyperparameters defining the neural network’s numbers of layers,”) Examiner’s Note: With respect to Claim 10: Under broadest reasonable interpretation the “backpropagator” is essentially the functional step/procedure and weight-update mechanism.
Wang and Ruff are analogous art and in the same field of invention because both references pertain to training a neural network learning model against a physics-based digital twin, using the digital twin’s own output as ground truth within an iterated training loop, to obtain a model that runs faster than the digital twin itself. While Wang teaches training a knowledge-based surrogate against a CFD digital twin through a 4-step calibration cycle, Ruff teaches a gradient-descent optimization process that generates the full candidate input used to query a digital twin simulation each iteration specifically to minimize the number of costly digital twin evaluations needed. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (offloading the search from the expensive digital twin to the surrogate) with the teachings of Ruff (optimizing over the complete input space rather than a subset of it) in order to arrive at a faster running learning model in place of the digital twin. One of ordinary skill in the art would be motivated to do so because by integrating Ruff's framework into the methods of Wang one would be able to note that a system as such, "offers a novel simulation-based optimization approach that eliminates the need for hand-tuning input parameters and significantly reduces the overall computation time…improves desired simulation outcome by 74%, {[pg. 1] of Ruff}."
Therefore, Claims 1, 10, and 19 are rejected.
With respect to Claims 2, 11, and 20:
The combination of Wang-Ruff teaches:
"The method of claim 1, further comprising: running the learning model using new input, producing learning model output; running the digital twin model using the new input, producing digital twin output; and comparing the learning model output and the digital twin output to determine learning model divergence from the digital twin model." ([Ruff pg. 3, Algorithm 1] Step 5 discloses selection of new input via the learning model which also generates new outputs, “Select the input xbest that produces minimum loss according to objective function.” [Ruff pg. 3, Algorithm 1] Step 6 further discloses that same new input being run through the digital twin to produce new digital twin outputs, “Intelligently query f(x) with the selected xbest.” [Ruff pg. 4, fig. 3] also teaches digital twin and learning model divergence comparisons being plotted, “Predicted outputs from surrogate neural network versus actual outputs from simulation. Ideally, these outputs would be perfectly correlated but the neural network is able to mostly accurately predict all three outputs with a few outliers.”) Examiner’s Note: x_best is new relative to the accumulated training sets.
Therefore, Claims 2, 11, and 20 are rejected.
With respect to Claim 12:
The combination of Wang-Ruff teaches:
"The system of claim 10, further comprising an iterator, such that for each iteration, input of the digital twin model runner is used as input into the learning model." ([Wang pg. 205, Step 1] discloses the component that repeatedly captures the digital twin runner’s input being reused as the learning model’s training input during each iteration, “The training data is generated by solving the CFD model with collected system input (including CRAC temperature setpoints and fan speeds, server powers) and the initial α or calibrated α by step ③ of the previous iterations to yield the predicted temperatures at sensor locations… The system input, the α,and the predicted temperatures form a new training data sample…” [Wang pg. 206, Algorithm 1, Step 3] further discloses the repeating control structure that drives this component through successive passes, “for i = 1 : Max iteration do.”) Examiner’s Note: Under broadest reasonable interpretation the iterator component behaves consistently with the disclosure’s own inventive concept of a single control structure that repeatedly synchronizes the digital twin runner’s input with the learning model’s training input across the training loop.
Therefore, Claim 12 is rejected.
The combination of Wang-Ruff teaches:
"The system of claim 12, wherein the digital twin model has input nodes and output nodes, the learning model has input nodes and output nodes, wherein there is an equal number of digital 4/5 twin input nodes and learning model input nodes, and wherein there is an equal number of digital twin output nodes and learning model output nodes." ([Ruff pg. 2, sec. III-B] discloses the digital twin’s input and output dimensionality, “For all simulations, there must be m-dimensional inputs (x ∈ Rm) into the simulation f(x) and n-dimensional out puts (y ∈ Rn). Furthermore, there are m-dimensional minimum and maximum bounds (xmin, xmax ∈ Rm) for each input as well as a n-dimensional desired set of outputs (ytarget ∈ Rn).” [Ruff pg. 2, sec. III-B] further discloses that the learning model shares these identical dimensions, “a neural network is trained to be a surrogate model ˆf(x) for the real simulation. This surrogate model is able to predict n dimensional outputs based on real m-dimensional inputs…”) Examiner’s Note: input/output dimension equivalence match denoted for both digital twin (simulation f) and the learning model (surrogate ^f).)
Therefore, Claim 13 is rejected.
Claims 3 and 4 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, in view of Ruff, further in view of Vardhan et. Al (Deep Active Learning for regression using ε-weighted Hybrid Query strategy, published July 10, 2022, hereinafter "Vardhan”). The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claim 3:
The combination of Wang-Ruff alone does not appear to explicitly disclose:
“The method of claim 2, wherein when the learning model divergence from the digital twin model is greater than a predetermined value then performing at least one iteration of the iterated digital twin system.”
However, Vardhan teaches:
“The method of claim 2, wherein when the learning model divergence from the digital twin model is greater than a predetermined value then performing at least one iteration of the iterated digital twin system.” ([pg. 2, sec. 1] discloses an active learning loop in which a model is trained, used to select new points, evaluated against the simulator (digital twin model), and added back for retraining, “Train a machine learning model (h) using Dtrain for constraining the sampling procedure in next iteration… Select new samples in next iteration based on score (s), and evaluate these samples to create new additional training data (Dnew train)… Iterate 2-5 until budget of training sample (B = |Dtrain|).” [pg. 3, sec. 2] further discloses a binary indicator marking whether the student’s (learning model’s) prediction diverges too far from the simulator’s (digital twin’s) output, “we use already evaluated historical training and test data from prior iterations and label it with a binary indicator (C), where C : X →{0,1}, where C = 1, indicates student network is able to predict on the sample within an acceptable level of accuracy otherwise C = 0.” [pg. 5, sec. 4] further teaches the concrete fixed threshold, “Our accuracy metrics is fraction of test predictions which are out of acceptable accuracy(+/- 5% error) i.e., ∑T i=1I(|yi−y∗ i |≤0.05∗|y∗ i |) . Here yi is a prediction on ith test sample having ground truth y∗ i and T is the total number of samples in the test data.” [pg. 3, sec. 2] further discloses that the points predicted to exceed this divergence threshold are specifically what gets selected for re-evaluation, “it is possible to identify at each iteration, what are the samples x ∈ DS on which student network has a high likelihood to fail.” [pg. 4, Algorithm 1] teaches the fixed cutoff used to make this selection, “Xf ⊆Xt :Xf =Xt ∀Xt, where f(Xt)≥0.5 iltered high failure probability samples Xf.” [pg. 3, sec. 2] further discloses that this cycle repeats/iterates, “We continue refining the student and teacher networks on further evaluated samples at each iteration until we exhaust our budget for iterations.”) Examiner’s Note: C=0 is the condition where the student’s (learning model’s) output diverges from the digital twin’s output by more than a fixed predetermined tolerance (5% error). The points predicted to meet that condition are selected to be sent to the simulator (digital twin) for re-evaluation and the resulting data feeds back into another training pass (iteration).
Wang-Ruff-Vardhan are analogous art and in the same field of invention because all three references pertain to analyzing the precision of artificial intelligence models to measure their computational reliance, output trustworthiness, and accuracy prospects. While Wang teaches a cost function pairing a trained surrogate’s output directly against real sensor ground truth, Ruff teaches a general iteration control structure in which threshold-tested criteria governs whether a surrogate training loop continues or halts. Similarly, Vardhan teaches a concrete fixed tolerance rule that converts a prediction versus target cost into a binary determination of whether a trained network’s output is acceptable. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (ground truth cost computation) with the teachings of Ruff (re-querying across successive passes) and further with the teachings of Vardhan (fixed percentage error tolerance conversion) in order to build an intelligent training system that automatically limits digital twin consultation further cutting computational overhead while ensuring deployed models meet verification and accuracy standards. One of ordinary skill in the art would be motivated to do so because by integrating Ruff and Vardhan’s frameworks into the methods of Wang one would be able to recognize that a system as such enables, "the improvement in accuracy of surrogate using epsilon-greedy HQS at a higher rate than randomized sampling as the epsilon-greedy HQS can find strategic samples, which random sampling cannot. The accuracy increment was 4.14 %(86.47% to 90.61%) for epsilon-greedy HQS in comparison to 1.99%(83.76% to 85.75%) in the case of random sampling, {[Pg. 6] of Vardhan}."
Therefore, Claim 3 is rejected.
With respect to Claim 4:
The combination of Wang-Ruff-Vardhan teaches:
“The method of claim 3, wherein the iterated digital twin system further comprises a digital twin ground truth, and further comprising:” ([Wang pg. 201] discloses an iterated digital twin system and a digital twin ground truth, “To evolve a CFD model into its digital twin form, it is important to instrument the model with sufficiently complete configuration of the physical infrastructure. A model with incomplete configuration may diverge from the ground truth.” [Wang pg. 203] further discloses a digital twin ground truth distinct from the digital twin’s own output (Definition 2 output ~T), “Definition 3 (Measurement). The measurement is a vector of real temperature values recorded by the physical sensors, which is denoted by Ts = (Ts1,Ts2,. . .,Tsn).”) Examiner’s Note: as per disclosure’s own claim language [fig. 8 description], this ground truth differs from the previously addressed “learning model ground truth”.
“using the cost derived from the learning model output and the digital twin ground truth-----” ([Wang pg. 204, Equation 3] discloses a cost pairing the learning model’s (surrogate model) output directly against this ground truth, “α∗ ≜ argmin α ||Ts(W∗,x) − Ts||2 2. (3) Third, the α∗ is configured into the CFD model. Finally, the CFD is validated based on sensor measurements. If the surrogate approaches to the CFD model, the α∗ after the convergence of the four-step iterations will approach to the one given by Eq. (1). As discussed in §1, to address the challenges of the surrogate’s complexity versus the needed volume of CFD-generated training data, we build a knowledge-based neural surrogate…”)
“-----to determine when a learning model stopping state has been reached;” ([Ruff pg. 6, sec. IV-H] discloses that a threshold-based comparison against a target quantity functions as the determination governing whether an iterated training loop continues or reaches a stopping state, “Until the stopping criteria is met, this algorithm will iteratively train a new neural network, optimize the objective function, and intelligently query the simulation. The stopping criteria for this application checks for three conditions every iteration: 1) If the goal criteria for the outputs have been met 2) If the algorithm converged on a solution outside of the goal criteria 3) If the maximum number of iterations been met.”) Examiner’s Note: A fixed criterion applied to a prediction-vs-target cost producing an accept/continue decision (stopping state determination).
and when the learning model stopping state has been reached, then running the learning model without running the digital twin model.” ([Wang pg. 206, Algorithm 1, lines 10-11] teaches a conditional check comparing the digital twin’s current accuracy against the best result recorded so far, “if 1 n n i |Tsi −Tsi| < ϵ then ϵ ← 1 n n i |Tsi −Tsi|; α∗ ← α;” [Wang pg. 201, sec. 1] discloses that once the learning model is trained well enough for that iteration it operates on its own for the parameter search step without the digital twin, “Kalibre avoids directly solving the CFD model for configuration search with the help of a trainable neural net.”) Examiner’s Note: Claims 15 and 16 combined amount to the repeated content of Claim 4.
Therefore, Claim 4 is rejected.
Claims 5, 14, 15, 16, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, in view of Ruff, further in view of Bunning et. Al (Input Convex Neural Networks for Building MPC, published November 26, 2020, hereinafter "Bunning”). The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claims 5 and 14:
The combination of Wang-Ruff alone does not appear to explicitly disclose:
“The method of claim 1, wherein the learning model represents a digital twin input space more convexly than the digital twin model.”
However, Bunning teaches:
“The method of claim 1, wherein the learning model represents a digital twin input space more convexly than the digital twin model.” ([pg. 2, sec. 1] discloses a specific concrete design-time constraint applied to a network/model’s structure/input space to force it to be provably convex, “present restrictions on the structure and weights of feed-forward ANN to build Fully Input Convex Neural Networks (FICNN), where the model output is convex with respect to all model inputs in single step predictions…We do this for both FICNN and PICNN by further constraining the network weights and by adding activation functions in appropriate places.” [pg. 3, sec. 2.1-2.2] further discloses the exact form of that constraint, “To find optimal solutions to problem (1) fast and reliably, it is beneficial if the overall optimization problem is convex in the decision variables u and x. This implies restrictions on the cost function J, dynamics function f, and constraint functions X and U. A common way for ensuring convexity is to select J to be a convex function…The network output zL is convex with respect to the input y if all W(z) 1:L−1 are non-negative and all activation functions gi are convex and non-decreasing.”) Examiner’s Note: This claim inherits its digital twin and learning model structure from Claim 1. As per the disclosure’s own specification [0093, Fig. 10], stating that the convexity difference between the learning model and the digital twin, “may be enforced by design through, for example, constraining the number of inputs into the learning model as compared to the digital twin, applying a smoothing techniques, such as regularization, during the learning model's training, and so on.” Thus, a structural weight constraint and activation function restriction is applied to the learning model deliberately to guarantee convexity that an unconstrained version of the same network would not have, further representing the input space more convexly.
Wang-Ruff-Bunning are analogous art and in the same field of invention because all three references pertain to analyzing the precision of artificial intelligence models to measure their computational reliance, output trustworthiness, and accuracy prospects. While Wang teaches replacing a computationally expensive CFD model with a lighter surrogate specifically to make configuration search tractable, Ruff teaches selecting inputs via gradient based optimization to drive an efficient search process. Similarly, Bunning teaches restricting a network’s weights to be non-negative and its activation functions to be convex and non-decreasing. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (neural surrogate retraining to find optimal parameters) with the teachings of Ruff (minimize the number of simulation executions required) further with the teachings of Bunning (achieve a convex input-output relationship for multi-step ahead predictions) in order to give the learning model a single reliably reachable solution over the same input space the digital twin represents. One of ordinary skill in the art would be motivated to do so because by integrating Ruff and Bunning’s frameworks into the methods of Wang one would be able to recognize that a system as such, "further demonstrates that such networks have high prediction accuracy in many domains, such as multi-label prediction, image completion, and reinforcement learning problems, {[Pg. 2] of Bunning}."
Therefore, Claims 5 and 14 are rejected.
With respect to Claim 15:
The combination of Wang-Ruff-Bunning teaches:
“The system of claim 14, wherein the iterated digital twin system further comprises a digital twin ground truth, and further comprising” ([Wang pg. 201] discloses an iterated digital twin system and a digital twin ground truth, “To evolve a CFD model into its digital twin form, it is important to instrument the model with sufficiently complete configuration of the physical infrastructure. A model with incomplete configuration may diverge from the ground truth.” [Wang pg. 203] further discloses a digital twin ground truth distinct from the digital twin’s own output (Definition 2 output ~T), “Definition 3 (Measurement). The measurement is a vector of real temperature values recorded by the physical sensors, which is denoted by Ts = (Ts1,Ts2,. . .,Tsn).”) Examiner’s Note: as per disclosure’s own claim language [fig. 8 description], this ground truth differs from the previously addressed “learning model ground truth”.
“using the cost derived from the learning model output and the digital twin ground truth” ([Wang pg. 204, Equation 3] discloses a cost pairing the learning model’s (surrogate model) output directly against this ground truth, “α∗ ≜ argmin α ||Ts(W∗,x) − Ts||2 2. (3) Third, the α∗ is configured into the CFD model. Finally, the CFD is validated based on sensor measurements. If the surrogate approaches to the CFD model, the α∗ after the convergence of the four-step iterations will approach to the one given by Eq. (1). As discussed in §1, to address the challenges of the surrogate’s complexity versus the needed volume of CFD-generated training data, we build a knowledge-based neural surrogate…”)
to determine when a learning model stopping state has been reached.” ([Ruff pg. 6, sec. IV-H] discloses that a threshold-based comparison against a target quantity functions as the determination governing whether an iterated training loop continues or reaches a stopping state, “Until the stopping criteria is met, this algorithm will iteratively train a new neural network, optimize the objective function, and intelligently query the simulation. The stopping criteria for this application checks for three conditions every iteration: 1) If the goal criteria for the outputs have been met 2) If the algorithm converged on a solution outside of the goal criteria 3) If the maximum number of iterations been met.”) Examiner’s Note: A fixed criterion applied to a prediction-vs-target cost producing an accept/continue decision (stopping state determination).
Therefore, Claim 15 is rejected.
With respect to Claim 16:
The combination of Wang-Ruff-Bunning teaches:
“The system of claim 15, further comprising when the learning model stopping state has been reached, then running the learning model without running the digital twin model.” ([Wang pg. 206, Algorithm 1, lines 10-11] teaches a conditional check comparing the digital twin’s current accuracy against the best result recorded so far, “if 1 n n i |Tsi −Tsi| < ϵ then ϵ ← 1 n n i |Tsi −Tsi|; α∗ ← α;” [Wang pg. 201, sec. 1] discloses that once the learning model is trained well enough for that iteration it operates on its own for the parameter search step without the digital twin, “Kalibre avoids directly solving the CFD model for configuration search with the help of a trainable neural net.”) Examiner’s Note: Claims 15 and 16 combined amount to the repeated content of Claim 4.
Therefore, Claim 16 is rejected.
With respect to Claim 17:
The combination of Wang-Ruff-Bunning teaches:
"The system of claim 16, running the learning model using new input, producing learning model output; running the digital twin model using the new input, producing digital twin output; and comparing the learning model output and the digital twin output to determine learning model divergence from the digital twin model." ([Ruff pg. 3, Algorithm 1] Step 5 discloses selection of new input via the learning model which also generates new outputs, “Select the input xbest that produces minimum loss according to objective function.” [Ruff pg. 3, Algorithm 1] Step 6 further discloses that same new input being run through the digital twin to produce new digital twin outputs, “Intelligently query f(x) with the selected xbest.” [Ruff pg. 4, fig. 3] also teaches digital twin and learning model divergence comparisons being plotted, “Predicted outputs from surrogate neural network versus actual outputs from simulation. Ideally, these outputs would be perfectly correlated but the neural network is able to mostly accurately predict all three outputs with a few outliers.”) Examiner’s Note: x_best is new relative to the accumulated training sets.
Therefore, Claim 17 is rejected.
Claim 18 is rejected under 35 U.S.C. 103 as being unpatentable over Wang, in view of Ruff, further in view of Bunning, further in view of Vardhan. The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claim 18:
The combination of Wang-Ruff-Bunning alone does not appear to explicitly disclose:
“The system of claim 17, wherein when the learning model divergence is greater than a predetermined amount, then running the learning model within the iterated digital twin system.”
However, Vardhan teaches:
“The system of claim 17, wherein when the learning model divergence is greater than a predetermined amount, then running the learning model within the iterated digital twin system.” ([pg. 2, sec. 1] discloses an active learning loop in which a model is trained, used to select new points, evaluated against the simulator (digital twin model), and added back for retraining, “Train a machine learning model (h) using Dtrain for constraining the sampling procedure in next iteration… Select new samples in next iteration based on score (s), and evaluate these samples to create new additional training data (Dnew train)… Iterate 2-5 until budget of training sample (B = |Dtrain|).” [pg. 3, sec. 2] further discloses a binary indicator marking whether the student’s (learning model’s) prediction diverges too far from the simulator’s (digital twin’s) output, “we use already evaluated historical training and test data from prior iterations and label it with a binary indicator (C), where C : X →{0,1}, where C = 1, indicates student network is able to predict on the sample within an acceptable level of accuracy otherwise C = 0.” [pg. 5, sec. 4] further teaches the concrete fixed threshold, “Our accuracy metrics is fraction of test predictions which are out of acceptable accuracy(+/- 5% error) i.e., ∑T i=1I(|yi−y∗ i |≤0.05∗|y∗ i |) . Here yi is a prediction on ith test sample having ground truth y∗ i and T is the total number of samples in the test data.” [pg. 3, sec. 2] further discloses that the points predicted to exceed this divergence threshold are specifically what gets selected for re-evaluation, “it is possible to identify at each iteration, what are the samples x ∈ DS on which student network has a high likelihood to fail.” [pg. 4, Algorithm 1] teaches the fixed cutoff used to make this selection, “Xf ⊆Xt :Xf =Xt ∀Xt, where f(Xt)≥0.5 iltered high failure probability samples Xf.” [pg. 3, sec. 2] further discloses that this cycle repeats/iterates, “We continue refining the student and teacher networks on further evaluated samples at each iteration until we exhaust our budget for iterations.”) Examiner’s Note: C=0 is the condition where the student’s (learning model’s) output diverges from the digital twin’s output by more than a fixed predetermined tolerance (5% error). The points predicted to meet that condition are selected to be sent to the simulator (digital twin) for re-evaluation and the resulting data feeds back into another training pass (iteration).
Wang-Ruff-Bunning-Vardhan are analogous art and in the same field of invention because all four references pertain to constructing a learning model that stays persistently relative to a digital twin’s predictive maintenance reliability. While Wang teaches allowing the surrogate to learn from far fewer training samples than a generic network would require, Ruff teaches a sampling strategy that ranks candidate inputs by their expected information value before committing to a costly simulation run. Similarly, Bunning teaches a network design that mathematically guarantees a single global optimum exists and can be found through gradient descent rather than leaving the optimizer exposed to unpredictable local minima, while Vardhan teaches a trained network dedicated to forecasting where the primary network’s errors are likely to concentrate so that a scarce evaluation budget is spent on the region’s most likely to be wrong rather than spread evenly. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (sample efficient architecture) with the teachings of Ruff (value ranked input selection) with the teachings of Bunning (guaranteed solvable optimization structure) further with the teachings of Vardhan (failure-forecasting retraining trigger) in order to bridge the gap between how well a learning model performs and how well the digital twin it approximates actually behaves. One of ordinary skill in the art would be motivated to do so because by integrating Ruff, Bunning, and Vardhan's frameworks into the methods of Wang one would be able to recognize that, "Kalibre’s improvement of up to 0.5°C is significant in CFD modeling due to the sharply increased difficulty in improving accuracy when the errors are already low (i.e., at around 1°C). Kalibre sheds lights on the calibration of other compute-intensive models to pursue high accuracy in approximating complex physical processes, {[pg. 209] of Wang}."
Therefore, Claim 18 is rejected.
Claims 6 and 7 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, in view of Ruff, further in view of Bunning, further in view of Chen et. Al (Optimal Control Via Neural Networks: A Convex Approach, published February 26, 2019, hereinafter "Chen”). The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claim 6:
The combination of Wang-Ruff-Bunning alone does not appear to explicitly disclose:
“The method of claim 5, wherein the learning model is a recurrent neural network.”
However, Chen teaches:
“The method of claim 5, wherein the learning model is a recurrent neural network.” ([pg. 11-12, sec. B] teaches the digital twin generating the training data the learning model is trained against, “At time t, we assume the building’s running profile xt := [st,ut] is available, where st denotes building system states, including outside temperature, room temperature measurements, zone occupancies and etc. ut denotes a collection of control actions such as room temperature set points and appliance schedule…we simulate the building running through the year and record (xt, Pt) with a resolution of 10 minutes.” [pg. 11, sec. B] further discloses the learning model (the recurrent network trained on that digital twin’s output], “we firstly learn an ICRNN model f(·) of the building dynamics, which is trained to minimize the error between Pt and f (xt−nw ,...,xt), while nw denotes the memory window of recurrent neural networks.” [pg. 5, sec. B] teaches that this learning model’s architecture is recurrent, “we propose to use recurrent neural networks (instead of feed-forward neural networks)…Fig. 2(b) shows the proposed input convex recurrent neural networks (ICRNN) structure.”) Examiner’s Note: As per Chen [pg. 12], EnergyPlus is the physics based digital twin generating ground-truth building energy data (Pt) and the ICRNN is trained via gradient descent to minimize the error between its own prediction and that digital twin’s output.
Wang-Ruff-Bunning-Chen are analogous art and in the same field of invention because all four references contribute a distinct capability towards accurately capturing and reliably optimizing a system’s behavior using a trained network in place of a physical model. While Wang teaches a knowledge-based network architecture whose internal blocks mirror physical heat-transfer relationships, Ruff teaches input-selection strategies that query the digital twin only on the most informative candidate inputs rather than sampling broadly. Similarly, Bunning teaches constraining a network’s weights and activation functions, so its optimization landscape is provably convex, while Chen teaches a recurrent network architecture using ReLu activation functions mathematically proven to preserver that same convexity guarantees across multi-step time-dependent predictions. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (sub-system decomposed architecture) with the teachings of Ruff (gradient-guided querying strategy) with the teachings of Bunning (layer-wise convexity proof) further with the teachings of Chen (time-extended recurrent structure) in order to assist a learning model’s representation of sequential digital twin behaviors through the utilization of recurrent architectures which are robustly standard for enhancing time-dependent system dynamics. One of ordinary skill in the art would be motivated to do so because by integrating Ruff, Bunning, and Chen's frameworks into the methods of Wang one would be able to recognize that a system as such can, "bridge the gap between model accuracy and control tractability faced by neural networks, by explicitly constructing networks that are convex with respect to their inputs. We show that these input convex networks can be trained to obtain accurate models of complex physical systems, {[pg. 1] of Chen}."
Therefore, Claim 6 is rejected.
With respect to Claim 7:
The combination of Wang-Ruff-Bunning alone does not appear to explicitly disclose:
“The method of claim 6, wherein the recurrent neural network uses a ReLu as its activation function.”
However, Chen teaches:
“The method of claim 6, wherein the recurrent neural network uses a ReLu as its activation function.” ([pg. 5, sec. B, Proposition 2] refers to Fig. 2(b) which addresses the proposed input convex recurrent neural network (ICRNN) architectures, “The network shown in Fig. 2(b) is a convex function from inputs to output if all weightsU,V,W,D1,D2,D3 are non-negative, and all activation functions are convex and nondecreasing (e.g. ReLU).” [pg. 8, sec. B] further reinforces the recurrent neural network learning model utilizing ReLu as its activation function, “A key reason that we propose to use ICNN (or ICRNN) to fit a function rather than directly finding a maximum of affine functions is that the former is a much more efficient parameterization than the latter. As stated in Theorem 2, a maximum of K affine functions can be represented by an ICNN with K layers, where each layer only requires a single ReLU activation function.”)
Therefore, Claim 7 is rejected.
Claims 8 and 9 are rejected under 35 U.S.C. 103 as being unpatentable over Wang, in view of Ruff, further in view of Bunning, further in view of Chen, further in view of Yoon et. Al (Zero-shot Transfer Learning within a Heterogeneous Graph via Knowledge Transfer Networks, published October 12, 2022, hereinafter "Yoon”). The specified dates are before the effective filing date of this application, i.e., April 05, 2024, as well as the provisional application date, April 06, 2023, where it applies.
With respect to Claim 8:
The combination of Wang-Ruff-Bunning-Chen alone does not appear to explicitly disclose:
“The method of claim 7, wherein the digital twin model is a heterogenous neural network.”
However, Yoon teaches:
“The method of claim 7, wherein the digital twin model is a heterogenous neural network.” ([pg. 4, sec. 3.2, Equation 4] discloses that each node’s transformation function is governed by parameters specific to that node’s type not shared uniformly across the heterogenous digital twin network, “Transform function maps the message into a type-specific latent space: Transform(l)(j)=α(W(l) τ(j) ·Aggregate(l)(E(j))) (4) where W(l) t are the specific transformation parameters for each node type t ∈ T and each of L HMPNN layers.” [pg. 4, sec. 4.1] further reinforces that nodes of different types are computed using entirely separate (heterogenous) parameter sets within the same overall network, “we see that h(l) 1 and h(l) 2, which are features of different types, are extracted using disjoint sets of model parameters at l-th layer. In a 2-layer HMPNN, this creates unique gradient backpropagation paths between the two node types...Therefore they project node features into different latent spaces and have different update equations during training.” [pg. 23, sec. A.10] further teaches a heterogenous neural network architecture, “All HGNN models we describe above have layer-wise parameters. As all HGNN models have parameters specialized in either node/edge/meta-path types, they all have distinct feature extractors for each node types…” ) Examiner’s Note: This claim inherits its digital twin structure from Claim 1 and utilizes the HGNN architecture to model/mirror the real-world data it represents.
Wang-Ruff-Bunning-Chen-Yoon are analogous art and in the same field of invention because all five references pertain to developing a surrogate trained against a physical model refined to be reliably optimizable and extended to handle sequential behaviors. While Wang teaches training a neural surrogate against a CFD digital twin’s own output, with the surrogate’s output nodes matched to the digital twin’s sensor locations, Ruff teaches matching a surrogate’s input and output dimensions node-for-node to the digital twin simulation it approximates. Similarly, Bunning teaches constraining a network’s weights and activation functions, so its optimization problem remains convex and solvable to a single global optimum, while Chen teaches extending that same convexity guarantees to a recurrent architecture so time dependent predictions retain the same reliable optimization properties as single-step ones. Additionally, Yoon teaches a graph neural network architecture in which each node type carries its own specialized transformation parameters rather than one function shared uniformly across all nodes in the network. It would have been obvious to a person having ordinary skill in the art (PHOSITA) before the effective filing date of the claimed invention to implement the base reference of Wang (sub-system decomposed architecture) with the teachings of Ruff (gradient-guided querying strategy) with the teachings of Bunning (layer-wise convexity proof) further with the teachings of Chen (time-extended recurrent structure) further with the teachings of Yoon (per-node-type specialization) in order to reduce both the computational cost of running the expensive digital twin and the training cost of the learning model that replaces it while preserving a structure that doesn’t waste capacity forcing dissimilar components through one shared function. One of ordinary skill in the art would be motivated to do so because by integrating Ruff, Bunning, Chen's, and Yoon’s frameworks into the methods of Wang one would be able to recognize that a system as such, "handily improves HGNN performances up to 960% for zero-labeled node types across 6 different HGNN models and outperforms many challenging baselines up to 73% higher in MRR. Future work in the area includes filtering noisy edges between source and target domains and making KTN more robust and less dependent on structure of given noisy heterogeneous graphs., {[pg. 10] of Yoon}."
Therefore, Claim 8 is rejected.
With respect to Claim 9:
The combination of Wang-Ruff-Bunning-Chen-Yoon teaches:
“The method of claim 8, wherein the digital twin model has input nodes and output nodes, the learning model has input nodes and output nodes, and wherein there is an equal number of digital twin input nodes and learning model input nodes; and wherein there is an equal number of digital twin output nodes and learning model output nodes.” ([Ruff pg. 2, sec. III-B] discloses the digital twin’s input and output dimensionality, “For all simulations, there must be m-dimensional inputs (x ∈ Rm) into the simulation f(x) and n-dimensional out puts (y ∈ Rn). Furthermore, there are m-dimensional minimum and maximum bounds (xmin, xmax ∈ Rm) for each input as well as a n-dimensional desired set of outputs (ytarget ∈ Rn).” [Ruff pg. 2, sec. III-B] further discloses that the learning model shares these identical dimensions, “a neural network is trained to be a surrogate model ˆf(x) for the real simulation. This surrogate model is able to predict n dimensional outputs based on real m-dimensional inputs…”) Examiner’s Note: input/output dimension equivalence match denoted for both digital twin (simulation f) and the learning model (surrogate ^f).
Therefore, Claim 9 is rejected.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure: Cella et. Al (US20230214925A1: TRANSACTION PLATFORMS WHERE SYSTEMS INCLUDE SETS OF OTHER SYSTEMS, Filed March 07, 2023).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NOOR F CHEEMA whose telephone number is (571)272-9642. The examiner can normally be reached Monday-Friday 7:30am-5:00pm alternative Fridays off.
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, Mariela Reyes can be reached at (571) 270-1006. 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.
/N.F.C./ Examiner, Art Unit 2142 /HAIMEI JIANG/Primary Examiner, Art Unit 2142
