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 .
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 1-20 are rejected under 35 U.S.C 101 because the claimed invention is directed to a judicial exception without significantly more.
Claim 1.
STEP 1: Yes. The claim is directed to a “method” which is a process.
STEP 2A PRONE ONE:
The claim recites multiple mental processes.
analyzing, using the subsurface process data, previously generated subsurface models in order to select a subset of previously generated subsurface models;
This describes an observation, evaluation, judgment or opinion that can be done in the mind or with aid of pen and paper. In this case an evaluation of comparing data against candidate models and choosing the ones that match.
… analyzing, using the additional subsurface process data, reducing the previously generated subsurface models in the subset; and
This describes an observation, evaluation, judgment or opinion that can be done in the mind or with aid of pen and paper. In this case an evaluation of comparing data against candidate models and choosing the ones that match in order to reduce the subsets of models.
STEP 2A PRONG TWO: The claim does not integrate the exception into a practical application.
STEP 2B: The claim does not recite an inventive concept or significantly more than the exception.
accessing the subsurface process data indicative of at least one subsurface process, the subsurface process data being periodically generated;
MPEP 2106.05(g) – This is mere data gathering of subsurface process data.
iteratively, responsive to receiving additional subsurface process data …
MPEP 2106.05(g) – This is mere data gathering of subsurface process data.
using one or more of the previously generated subsurface models in the subset in order to perform the one or more subsurface operations in the subsurface.
MPEP 2106.05(f) – This is mere instructions to apply the exception to use the models to perform a subsurface operation, without providing the specific means of doing so.
Conclusion: Claim 1 is directed to multiple mental processes, not integrated into a practical application and lacks an inventive concept. Therefore, it is ineligible under 35 U.S.C 101.
Regarding Claims 2-6:
These claims merely add further mathematical concepts to the abstract idea. They recite that the previously generated subsurface models are manifested in a proxy model that generates outputs from the subsurface process data (claim 2), generating a training set and generating the proxy model using the training set (claim 3), generating the proxy model via machine learning (claim 4), solving differential equations (claim 5), and generating a forward proxy model and inverse proxy model that are consistent with one another (claim 6). Machine Learning is recited at a high level of generality, with no model, architecture, or training procedure, and amounts to mere instructions to apply the judicial exception using a generic tool. MPEP 2106.05(f). This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claims 7-10:
These claims merely narrow the abstract idea by specifying the relationships used to generate proxy models. They recite that the machine learning is constrained by physics-based rules (claim 7), that a physics simulator solving differential equations generates the training set (claim 8), that the physics-based rules correlate input parameters to output parameters (claim 9), and that the correlation is linear, logarithmic, or exponential (claim 10). The physics-based rules and their correlations are mathematical relationships, so the claims remain a mathematical / mental abstraction. The physics simulator is recited only as a generic tool that performs the recited calculation. MPEP 2106.05(f). This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claim 11:
This claim merely specifies that the machine learning comprises a deep learning model (claim 11). The deep learning model is recited at a high level of generality and amounts to mere instructions to apply the judicial exception using a generic machine learning model tool. MPEP 2106.05(f). This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claims 12-13:
These claims merely narrow the input data type to a production data (claim 12) or diagnostics data (claim 13) and specify that the inverse proxy model outputs geological parameters and parameters related to completions. This is pre-solution data gathering MPEP 2106.05(g) that links the abstract idea to a field of use MPEP 2106.05(h). This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Claim 14.
STEP 1: Yes. The claim is directed to a “method” which is a process.
STEP 2A PRONE ONE:
The claim recites multiple mathematical abstractions.
generating, using a physics simulator solving differential equations, a training set of forward models;
This is a mathematical abstraction that can be a calculation, relationship, or equation / formula. In this case calculations of solving differential equations to produce a training set of forward models.
generating, via machine learning using the training set of forward models, a forward proxy model;
This is a mathematical abstraction that can be a calculation, relationship, or equation / formula. In this case calculations of parameter fitting functions to the numerical training set that maps input parameters.
generating, via the machine learning using the training set of forward models, an inverse proxy model such that the inverse proxy model is consistent with the forward proxy model;
This is a mathematical abstraction that can be a calculation, relationship, or equation / formula. In this case mathematical relationships and calculations to make sure the inverse proxy model is consistent with the proxy model.
using the subsurface process data as input to the inverse proxy model in order to generate outputs comprising geological parameters and parameters related to completions that are indicative of potential inverse models; and
This is a mathematical abstraction that can be a calculation, relationship, or equation / formula. In this case mathematical calculation because the limitation recites evaluating the inverse proxy model on the received subsurface process data to compute values of the geological parameters and the parameters related to completions.
STEP 2A PRONG TWO: The claim does not integrate the exception into a practical application.
STEP 2B: The claim does not recite an inventive concept or significantly more than the exception.
receiving subsurface process data;
MPEP 2106.05(g) – This is mere data gathering of subsurface process data.
using one or more of the potential inverse models to perform the one or more subsurface operations in the subsurface.
MPEP 2106.05(g) - This is mere instructions to apply the exception to use the models to perform a subsurface operation, without providing the specific means of doing so.
Conclusion: Claim 14 is directed to multiple mathematical concepts, not integrated into a practical application and lacks an inventive concept. Therefore, it is ineligible under 35 U.S.C 101.
Regarding Claims 15-17:
These claims merely narrow the abstract idea by specifying the relationships used to generate the proxy models. They recite that the machine learning used to generate the forward proxy model and the inverse proxy model is constrained by physics based rules (claim 15), that the forward proxy model receives input parameters and generates output parameters and the physics based rules correlate input parameters to output parameters (claim 16), and that the correlation is linear, logarithmic, and exponential (claim 17). These remain mathematical concepts. Machine learning is recited at a high level of generality and amounts to mere instructions to apply it. This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claim 18:
This claim merely specifies that the machine learning comprises a deep learning model (claim 18). The deep learning model is recited at a high level of generality and amounts to mere instructions to apply the judicial exception using a generic machine learning tool. MPEP 2106.05(f). This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claim 19:
This claim merely adds a further mathematical calculation, generating a number of forward models using the forward proxy model, and specifies the number of forward models generated and the computational time for generating them relative to the physics simulator (claim 19). Generating forward models with the forward proxy model is evaluating a fitted mathematical function. The number and computational time limitations state only the intended result of substituting the forward proxy model for the physics simulator. The claim recites no particular technical means for achieving the result, so it amounts to mere instructions to apply the judicial exception. This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Regarding Claim 20:
This claim merely adds validating the potential inverse models using the forward proxy model and, if they are determined to be invalid, using the proxy model to determine the geological parameters and the parameters related to completions (claim 20). Validating is evaluating the forward proxy model and comparing the result against a tolerance, which is a mathematical calculation and an evaluation that can practically be performed in the human mind. Using the proxy model to determine the parameters is a further mathematical calculation performed with a generic tool. MPEP 2106.05(f). An improvement in the speed of the mathematical calculation itself is an improvement to the abstract idea, not to the functioning of a computer or a technology. This does not integrate the judicial exception into a practical application. The claims do not resolve the issues from the claims they depend upon.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
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.
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.
Claim 1 is rejected under 35 U.S.C 103 as being unpatentable over “Approximate Bayesian Computation for Probabilistic Decline-Curve Analysis in Unconventional Reservoirs” by PARYANI et al. [herein “PARYANI”] (2017) and in view of US 10678967 B2 by BABIN et al. [herein “BABIN”] (2020).
Regarding Claim 1, PARYANI teaches
A computer-implemented method for analyzing subsurface process data in order to perform one or more subsurface operations in a subsurface, the method comprising:
accessing the subsurface process data indicative of at least one subsurface process, the subsurface process data being periodically generated;
“In approximate Bayesian computation (ABC)with rejection sampling, the posterior distribution is approximated by substituting different values of the decline-equation parameters into the Arps’ DCA model and generating a large number of production-profile realizations. Summary statistics (mean, standard deviation) between the simulated and observed production data are then compared.”. (Pg. 478 Summary).
“The 10th, 50th, and the 90th percentiles of the cumulative production are computed at 10-month increments, and at each timestep, a distribution of the cumulative production is obtained.”. (Pg. 479 ABC methodology).
“For comparison purposes, both the simulations were run on an i5 processor with 2 cores and a clock speed of 2.4 GHz. The computation time to simulate 100,000 data sets with the ABC methodology was 1 minute 14 seconds …”. (Pg. 480 Results and Discussion).
“The results of this model will provide a tool for quick initial analysis that will help in long term planning of operations necessary for optimal well completions and field development, all performed in a fraction of the time currently used by reservoir simulators.”. (Pg. 484 Results and Discussion).
This shows a computer analyzing monthly production data from shale wells, which is data generated by the subsurface process of producing from the reservoir and reported on a recurring schedule, for the purpose of planning completions and field development.
analyzing, using the subsurface process data, previously generated subsurface models in order to select a subset of previously generated subsurface models;
“To illustrate further the procedure, suppose we generate 10,000 sets of DCA parameter values. We then store the parameter sets in a 10,000 3 matrix where each row corresponds to the parameter set to be used in a simulation run and each column corresponds to the Arps DCA parameters (b, Di, and qi). By substituting each row of this matrix into the Arps model, we generate 10,000 synthetic production-data sets.”. (Pg. 480 ABC methodology).
“When the ABC simulation is run, the rejection-sampling algorithm compares each row of the simulated-summary statistic matrix with the actual production-summary statistic. If the difference between the simulated and actual summary statistic is less than the 10% (threshold) of the actual summary statistic and this is true for all the columns in a given row, then, the algorithm accepts the simulated production data. All the simulations that do not meet this requirement are rejected.”. (Pg. 480 ABC methodology).
“
PNG
media_image1.png
452
1230
media_image1.png
Greyscale
”. (FIG. 1).
This shows a full set of candidate well models being generated before the comparison, then each one compared against the observed production data and kept or rejected on a threshold, so that what remains is a subset of the original set chosen by the production data.
PARYANI does not explicitly teach but BABIN teaches
iteratively, responsive to receiving additional subsurface process data, analyzing, using the additional subsurface process data, reducing the previously generated subsurface models in the subset; and
“The user input 1108 results in the update of the measurements log 1111 and updates the bookmark on the current stage of the development 1102. Then, the action selector 1104 again picks from the manual 700 the next action based on the current stage 1102 and the full history of measurements 1111.”. (0084).
“Such simulations necessarily are probabilistic and yield a relatively large set M for each measurement; multiple measurements can reduce the scale of the set M.”. (0041).
“… upon any positive result of the first sequence of subsequent wells, drilling a second sequence of subsequent wells at a second sequence of subsequent locations and trajectories that are common to a second subset of the plurality of development scenarios, wherein the second subset includes at least one development scenario that is selected from the first subset based at least on the result of the first sequence of subsequent wells, wherein the second sequence of subsequent wells …”. (0090).
“These “Pw” nodes are added several layers down the tree at the tree level that corresponds to the moment in time when the flow-related information from the production well is considered.”. (0082).
“The only difference is that the decision at “Pw” node is made based on the information from the previously drilled production well, rather than from a just drilled appraisal well.”. (0083).
“While operating, the method 1100 takes as preliminary input the pre-calculated development manual 700 and at each iteration the system asks for the following intermediate input 1108 …”. (0088).
This shows a repeating loop in which each new set of measurements, including production well flow information taken at later points in time, is added to the measurement history and used to limit the set of geological realization further than the previous round did, so that the subset carried forward shrinks with each round of data.
using one or more of the previously generated subsurface models in the subset in order to perform the one or more subsurface operations in the subsurface.
“Each leaf 402 corresponds to one development scenario. After building TT, we calculate performance metrics for possible geological models in each leaf according to the path from the root to the leaf taking into account the appraisal wells.”. (0053).
“Following from measurements 1110 on the second appraisal well, the potential set of geological realizations is limited. Considering that the projected performance metric for any allowed development scenario is above a threshold value (a positive result corresponding to the set of geological realizations consistent with the appraisal well measurements) the action selector 1104 then might instruct the operator to drill 1106 a second sequence of production wells at a second sequence of production locations and trajectories …”. (0085).
“When the implementation of a development scenario is completed, or when the projected performance metric for any allowed development scenario falls below a threshold value (a negative result corresponding to the subset of geological realizations consistent with the appraisal well measurements), the system informs the user to stop 1112.”. (0087).
“An aspect of the invention is that appraisal wells can be drilled interspersed with production wells and using the same drilling machinery. This approach saves on cost while permitting rapid adaptation of the reservoir development plan. Commonly found examples of reservoir development plan adaptations are adjustments of the number of production wells and of their locations (i.e., the adjustment of well density).”. (0045).
“… the order 1106 for the operator to drill a production well; the order 1109 for the operator to drill an appraisal well; or the order 1112 for the operator to drill no more wells for the reservoir.”. (0088).
This shows the performance metrics of the geological realizations that survive the narrowing being what decides whether and where the next production wells are drilled, so the drilling of those wells is carried out on the basis of the models remaining in the subset.
It would have been obvious before the effective filing data of the claimed invention to incorporate BABIN’s teaching of narrowing a set of geological realizations again as each new round of field measurements arrive and drilling production wells based on the surviving realizations with PARYANI’s method of screening a previously generated set of well models against periodically reported production data. The reason for doing so would have been to carry the screened set forward and use it to direct field development as production data continues to arrive. As expressed by BABIN, this provides “Rapid adaptation of resource field development plans in response to evolving information on geological parameters.”. (0012).
Claims 2-5 are rejected under 35 U.S.C 103 as being unpatentable over “Approximate Bayesian Computation for Probabilistic Decline-Curve Analysis in Unconventional Reservoirs” by PARYANI et al. [herein “PARYANI”] (2017), in view of US 10678967 B2 by BABIN et al. [herein “BABIN”] (2020), and further in view of WO 2017188858 A1 by SOFRONOV et al. [herein “SOFRONOV”] (2017).
Regarding Claim 2, PARYANI and BABIN do not explicitly teach but SOFRONOV teaches
The method of claim 1, wherein the previously generated subsurface models are manifested in a proxy model; and
wherein the proxy model is configured to receive the subsurface process data and the additional subsurface process data and to generate outputs indicative of the previously generated subsurface models.
“As an example, a ML production prediction component may utilize one or more training sets optionally with one or more of historical data and synthetic data, which may be generated by a reservoir modeling tool (e.g., the ECLIPSE® simulator, etc.) for building a model that can be called by a prompt reservoir performance predictor. In such an example, the predictor can act as a fast data-driven tool for forecasting reservoir behavior, for example, as to one or more different development scenarios.”. (0108).
“As an example, a ML inversion component may utilize one or more of historical production data, new geological data, and reservoir modeling results as a training set or training sets, for example, to generate a reservoir model adjuster.”. (0115).
“For example, field data can be received by an input component and stored to a database of the database component. Such field data may be utilized in an inversion process that aims to reconstruct subterranean, spatial properties of a reservoir. Such spatial properties can include properties utilized in simulation fluid flow in the reservoir by a reservoir simulation framework.”. (0175).
“In such an example, the method can include receiving, via the network interface, additional data and regenerating the trained machine learning algorithm based at least in part on the outcome and at least a portion of the additional data.”. (0179).
This shows a machine learning model built from the outputs of simulated reservoir models, standing in for those models, which takes in field production data and later additional data and returns reservoir properties that identify which of the modeled reservoirs the data corresponds to.
It would have been obvious before the effective filing date for the claimed invention to incorporate SOFRONOV’s teaching of a machine learning stand in built from simulated reservoir models that receives filed data and returns reservoir properties with the method of PARYANI-BABIN of screening and narrowing a previously generated set of reservoirs models against production data. The reason for doing so would have been to obtain the model selection result from the arriving producing data without re-running or looking up each simulated model. As expressed by SOFRONOV, “… the predictor can act as a fast data-driven tool for forecasting reservoir behavior …”. (0108).
Regarding Claim 3, PARYANI and BABIN do not explicitly teach but SOFRONOV teaches
The method of claim 2, wherein the proxy model is generated by:
generating a training set of subsurface models; and
generating the proxy model using the training set of subsurface models.
“As an example, a learning tool can be associated with a training set (TS), for example, accessible via a data storage. As an example, TSs may be used for building one or more corresponding prediction models. As an example, a TS may be subdivided on historical data including initial and live information (e.g., from a smart database), and on synthetic data (e.g., obtained from modeling and stored in smart database).”. (0107).
“As an example, a ML production prediction component may utilize one or more training sets optionally with one or more of historical data and synthetic data, which may be generated by a reservoir modeling tool (e.g., the ECLIPSE® simulator, etc.) for building a model that can be called by a prompt reservoir performance predictor.”. (0108).
“As an example, machine learning (ML) optimization can allow for prediction of an optimal strategy for reservoir treatment with regard to particular production characteristics (e.g., cumulative oil production within a given time period). As an example, historical data and synthetic data may be utilized in one or more training sets (TSs) in one or more machine learning (ML) blocks. As an example, synthetic data can include numerous automatically generated scenarios of field development with a wide range of potential technology impacts (e.g., reservoir treatment scenarios) and corresponding output from a prompt production prediction model.”. (0110).
This shows a set of reservoir simulation runs being produced first to serve as the training set, and the prediction model then being built from that set, so the stand in model is derived from the simulated models rather than from the field data alone.
Regarding Claim 4, PARYANI and BABIN do not explicitly teach but SOFRONOV teaches
The method of claim 3, wherein the training set of subsurface models are generated via a non-machine learning methodology; and
wherein the proxy model is generated via machine learning using the training set of subsurface models.
“As an example, a learning tool can be associated with a training set (TS), for example, accessible via a data storage. As an example, TSs may be used for building one or more corresponding prediction models.”. (0107).
“As an example, a ML production prediction component may utilize one or more training sets optionally with one or more of historical data and synthetic data, which may be generated by a reservoir modeling tool (e.g., the ECLIPSE® simulator, etc.) for building a model that can be called by a prompt reservoir performance predictor.”. (0108).
“As an example, a predictor can include one or more relatively simplistic physical models of a well or wells (see, e.g., Kaviani et al., How accurate are Capacitance Model connectivity estimates?, Journal of Petroleum Science and Engineering, Volume 122, October 2014, Pages 439-452; and Artun, Characterizing interwell connectivity in waterflooded reservoirs using data- driven and reduced-physics models: a comparative study, Neural Computing and Applications, pp. 1-15, January 2016). In such an example, a reservoir performance predictor may utilize a reservoir model for training and output of a set of coefficients for one or more systems of physics-based equations (e.g., capacity resistance model -like approaches). As an example, machine learning can include deep learning (see, e.g., Deng et al., "Deep Learning: Methods and Applications", Foundations and Trends in Signal Processing 7: 3-4, 2014), for example, with one or more multi-layered artificial neural networks.”. (0109).
This shows the training set coming from a reservoir simulator, which is a physics-based tool rather than a learned one, and the prediction model then being built from that set by a machine learning component, so the two steps of the claim are performed by different kinds of methods in the same order.
Regarding Claim 5, PARYANI and BABIN do not explicitly teach but SOFRONOV teaches
The method of claim 4, wherein the non-machine learning methodology comprises solving differential equations.
“A simulator can be a computerized system (e.g., a computing system) that can execute instructions using one or more processors to solve a system of equations that describe physical phenomena subject to various constraints. The system of equations may be spatial defined (e.g., numerically discretized) according to a spatial model …”. (0027).
“The ECLIPSE® simulator includes numerical solvers that may provide simulation results such as, for example, results that may predict dynamic behavior for one or more types of reservoirs …”. (0033).
“… quantities such as temperature, pressure and porosity distributions within the sediments may be modeled, for example, by solving partial differential equations (PDEs) using one or more numerical techniques.”. (0048).
This shows that the reservoir modeling tool named as the source of the training data works by numerically solving the differential equations of the physical system.
Claim 6 is rejected under 35 U.S.C 103 as being unpatentable over “Approximate Bayesian Computation for Probabilistic Decline-Curve Analysis in Unconventional Reservoirs” by PARYANI et al. [herein “PARYANI”] (2017), in view of US 10678967 B2 by BABIN et al. [herein “BABIN”] (2020), in view of WO 2017188858 A1 by SOFRONOV et al. [herein “SOFRONOV”] (2017), and further in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019).
Regarding Claim 6, PARYANI, BABIN, and SOFRONOV do not explicitly teach but NARAYANASWAMY teaches
The method of claim 4, wherein a forward proxy model is generated via the machine learning using the training set of subsurface models;
wherein, after generating the forward proxy model, an inverse proxy model is generated via the machine learning using the training set of subsurface models and to be consistent with the forward proxy model; and
“In practice, the original simulator is often too expensive to support the necessary optimization and instead an approximate forward model, also called a surrogate, is fitted to a given simulation ensemble [1]. This creates two coupled problems: First, fitting a forward model F : x→ ythat approximates the original simulation by mapping the inputs x to the outputs y; and Second, deriving an inverse model G : y → x. Even for complex inputs and outputs, modern machine learning techniques have proven effective for building forward models [1].”. (Pg. 1 Introduction).
“As illustrated in Fig. 1(a), this approach utilizes two independent fully-connected networks representing the forward model F and the inverse model G respectively that are tied together through a cyclical consistency constraint … The cyclical constraint ensures that the residuals from the inversion process, r = x − µ, are controlled by the sensitivities of the forward model F.”. (Pg. 2 Section 2.1).
This shows a forward network and an inverse network both learned from the same set of simulator runs, with the inverse fitted so that its output passed back through the forward network reproduces the data, which is what makes the inverse consistent with the forward.
It would have been obvious before the effective filing date of the claimed invention to incorporate NARAYANASWAMY’s teaching of a forward surrogate and an inverse model learned from the same simulation set under a consistency constraint with PARYANI-BABIN-SOFRONOV of screening and narrowing a previously generated set of reservoir models against production data through a machine learned stand in built from reservoir modeling results. The reason for doing so would have been to obtain a reliable inversion from a limited number of simulator runs. As expressed by NARAYANASWAMY, “… self-consistency is highly beneficial, and both training strategies produce well calibrated inverse models.”. (Pg. 1 Abstract).
PARYANI, BABIN, and NARAYANASWAMY do not explicitly teach but SOFRONOV teaches
wherein the inverse proxy model is configured to receive the subsurface process data and the additional subsurface process data and to generate outputs comprising geological parameters and parameters related to completions.
“Such field data may be utilized in an inversion process that aims to reconstruct subterranean, spatial properties of a reservoir. Such spatial properties can include properties utilized in simulation fluid flow in the reservoir by a reservoir simulation framework.”. (0175).
“In such an example, the method can include receiving, via the network interface, additional data and regenerating the trained machine learning algorithm based at least in part on the outcome and at least a portion of the additional data.”. (0179).
This shows the inversion component taking field data, and then further data as it arrives, and returning the spatial properties if the reservoir that the flow simulation depends on.
Claims 7-9 and 11-13 are rejected under 35 U.S.C 103 as being unpatentable over “Approximate Bayesian Computation for Probabilistic Decline-Curve Analysis in Unconventional Reservoirs” by PARYANI et al. [herein “PARYANI”] (2017), in view of US 10678967 B2 by BABIN et al. [herein “BABIN”] (2020), in view of WO 2017188858 A1 by SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), and in further view of “Deep-Learning based Inverse Modeling Approaches: A Subsurface Flow Example” by WANG et al. [herein “WANG”] (2020).
Regarding Claim 7, PARYANI, BABIN, SOFRONOV, and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 6, wherein the machine learning to generate the forward proxy model is constrained by physics-based rules; and
“In the TgNN surrogate, physical/engineering constraints and other domain knowledge are incorporated as prior knowledge into the neural network training process.”. (Pg. 11 Section 4.2).
“In order to achieve theory-guided training, the DNN can then be substituted into the governing equation and boundary/initial conditions in Eq. (11), and the loss functions for physics-violation can be expressed as follows:”. (Pg. 12 Eq. 17-19).
“By incorporating physical laws, the TgNN surrogate can be constructed with limited labeled data, or even in a label-free manner.”. (Pg. 14).
This shows the surrogate network being trained against a loss that penalizes any departure from the governing flow equation and its boundary and initial conditions, in addition to mismatch with the simulation data, so the physical law limits what the model can learn.
wherein the machine learning to generate the inverse proxy model is constrained by the physics-based rules.
“The first three terms in the loss function ( , ) k h L[Symbol font/0x71] [Symbol font/0x71] constitute the physics-based loss, and the last two terms represent the data-driven loss or the data mismatch (Tartakovsky et al., 2020).”. (Pg. 18).
“Although the loss function is similar to that for TgNN surrogate construction, i.e., Eq. (20), the following two differences exist: first, the PINN model is constructed for a specific conductivity field, and thus no random variables exist in the neural network inputs; and second, two different neural networks are defined in the PINN model with two sets of network parameters ( ,k h [Symbol font/0x71] [Symbol font/0x71] ) needing to be trained simultaneously.”. (Pg. 18).
“The physics-based loss is also incorporated, as shown in the first three terms of Eq. (33). 10,000 collocation points are employed to enforce the physics constraints.”. (Pg. 36).
This shows the networks that recover the conductivity field from measurements being trained under the same governing equation, boundary conditions and initial condition terms used for the forward surrogate, with the paper stating that the two losses differ only in their inputs and number of networks and not in the physics applied.
It would have been obvious before the effective filing date of the claimed invention to incorporate WANG’s teaching of training both a forward and inverse network under the same governing equation with the method of PARYANI-BABIN-SOFRONOV-NARAYANASWAMY of generating forward and inverse proxy models from a training set of simulated reservoir models. The reason for doing so would have been to build reliable proxy models from a limited number of simulators runs. As expressed by WANG, “By incorporating physical laws and other constraints, the TgNN surrogate can be constructed with limited simulation runs and accelerate the inversion process significantly.”. (Abstract).
Regarding Claim 8, PARYANI, BABIN, SOFRONOV, and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 7, wherein a physics simulator, solving differential equations, generates a training set of subsurface forward models;
wherein the machine learning, constrained by the physics-based rules, generates the forward proxy model using the training set of subsurface forward models; and
“Several forward model simulations should be performed to provide training data for [Symbol font/0x71] 12 surrogate modeling, and the labeled data points can be represented as [Symbol font/0x28] [Symbol font/0x29] [Symbol font/0x7B] [Symbol font/0x7D] 1 , , , N i i i i i t h [Symbol font/0x3D] x ξ , where denotes the total number of labeled training data.”. (Pg. 11-12).
“30 hydraulic conductivity fields are generated with KLE, and numerical simulations are then performed to obtain training data.”. (Pg. 23).
“For the TgNN surrogate-based methods, the TgNN surrogate is trained by matching the available simulation data and honoring physical/engineering principles at selected collocation points.”. (Pg. 45-46).
This shows a set of flow simulations, each numerically solving the governing flow equation for a different sample conductivity field, being used as the training data, and the forward surrogate being trained to match those simulations while adhering to the physical principles.
wherein the machine learning, constrained by the physics-based rules,
“The first three terms in the loss function ( , ) k h L[Symbol font/0x71] [Symbol font/0x71] constitute the physics-based loss, and the last two terms represent the data-driven loss or the data mismatch (Tartakovsky et al., 2020).”. (Pg. 18).
“The physics-based loss is also incorporated, as shown in the first three terms of Eq. (33). 10,000 collocation points are employed to enforce the physics constraints. 2,000 epochs of training are performed with learning rate of 0.001.”. (Pg. 36).
This shows the inverse networks training being held to the same physical based loss terms.
PARYANI, BABIN, SOFRONOV, and WANG do not explicitly teach but NARAYANASWAMY teaches
generates the inverse proxy model using the training set of subsurface forward models.
“Given an experiment design (e.g. Latin Hypercube) for the input parameters x ∈ RD, we build a forward model F : x → z, where z ∈ Rd is the latent space inferred using a pre-trained encoder E : y → z. Here, y ∈ RT×k corresponds to a multivariate time-series output from the simulator comprised of k quantities and T time-steps. Correspondingly, the inverse G : z → x is used to recover the input parameters from a sample in the latent space. Our dataset is comprised of tuples {(xi,zi)},i = 1···N, where N denotes the total number of samples.”. (Pg. 2).
“As illustrated in Fig. 1(a), this approach utilizes two independent fully-connected networks representing the forward model F and the inverse model G respectively that are tied together through a cyclical consistency constraint.”. (Pg. 2).
“In this section, we evaluate the proposed approaches on a reservoir model simulation dataset comprised of N = 2000 samples …”. (Pg. 3).
This shows the inverse network being trained on the same set of reservoir simulator runs that is used to train the forward network.
Regarding Claim 9, PARYANI, BABIN, SOFRONOV, and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 8, wherein the forward proxy model receives input parameters and generates output parameters; and
“Considering the governing equation in Eq. (11), in order to approximate the quantity of interest ( , ; ) h t[Symbol font/0x77] x , a Deep Neural Network (DNN) is defined as follows (with ( ; ) K [Symbol font/0x77]x parameterized by ξ):”. (Pg. 11 Eq. 15).
“Therefore, the location, time, and stochastic parameters comprise the inputs of the neural network, i.e., ( , , ) t x ξ , as shown in Figure 1.”. (Pg. 11).
This shows the surrogate taking in the parameters that describe the conductivity field, along with location and time, and returning the hydraulic head, so it maps input parameters to an output parameter.
wherein the physics-based rules correlate one or more input parameters to one or more output parameters.
“… where N denotes a general nonlinear differential operator; [Symbol font/0x77] denotes the random variables in the probability space [Symbol font/0x57] ; D denotes the physical domain; T denotes the time span; ( ; ) K [Symbol font/0x77]x denotes the space-dependent coefficients, which can be seen as a random field; ( , ; ) h t[Symbol font/0x77] x denotes the quantity of interest or the solution of the problem …”. (Pg. 10).
“In order to achieve theory-guided training, the DNN can then be substituted into the governing equation and boundary/initial conditions in Eq. (11), and the loss functions for physics-violation can be expressed as follows:”. (Pg. 12 17-19).
“In this work, transient saturated subsurface flow problems are considered, which have a general form of governing equation: ( , ) ( ( ) ( , )) 0 s h t S K h t t [Symbol font/0xB6] [Symbol font/0x2D][Symbol font/0xD1][Symbol font/0xD7] [Symbol font/0xD1] [Symbol font/0x3D] [Symbol font/0xB6] x x x (34) where s S denotes the specific storage; ( , ) h tx denotes the hydraulic head; and ( ) Kx denotes the hydraulic conductivity.”. (Pg. 21).
This shows the physics-based rule applied during training is the governing flow equation, which links the conductivity field that the network receives as its input parameters to the hydraulic head the network outputs, so that the head the network predicts must change with the conductivity it is given in the way the equation necessitates.
Regarding Claim 11, PARYANI, BABIN, SOFRONOV, and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 9, wherein the machine learning comprises a deep learning model.
“The first category is deep-learning surrogate-based inversion methods, in which the Theory-guided Neural Network (TgNN) is constructed as a deep-learning surrogate for problems with uncertain model parameters.”. (Abstract).
“Considering the governing equation in Eq. (11), in order to approximate the quantity of interest ( , ; ) h t[Symbol font/0x77] x , a Deep Neural Network (DNN) is defined as follows (with ( ; ) K [Symbol font/0x77]x parameterized by ξ):”. (Pg. 11 Eq. 15).
“The TgNN surrogate is first constructed. A seven-hidden-layer fully-connected neural network with 50 neurons per layer is chosen for TgNN surrogate construction.”. (Pg. 23).
This shows machine learning used to build the physics constrained surrogate (proxy) being a fully connected neural network with seven hidden layers that the paper describes as a deep learning surrogate (i.e. proxy model).
Regarding Claim 12, BABIN, SOFRONOV, NARAYANASWAMY, and WANG do not explicitly teach but PARYANI teaches
The method of claim 8, wherein the subsurface process data comprises production data; and
“Summary statistics (mean, standard deviation) between the simulated and observed production data are then compared.”. (Pg. 478).
“Chosen for analysis were 57 gas wells from Johnson County (Barnett Shale) and 21 oil wells from Karnes County (Eagle Ford Shale) with production history of 90 and 45 months, respectively.”. (Pg. 478).
This shows the subsurface process data being the recorded production history of producing shale wells.
the parameters related to completions …
“Reserves estimates and production forecasts in hydraulically fractured shale wells have considerable uncertainty. One major source of uncertainty in the forecasts arises from the interaction between the induced hydraulic and natural fractures. This results in some wells with a variable stimulated reservoir volume (SRV). The variable SRV causes the well to deplete in different stages, which increases the uncertainty. The first depletion stage from the highly conductive fractures explains the early high production rates. This is followed by the production from the moderate- and finally low-conductivity fractures, respectively. In other words, the more-conductive SRV depletes first (Suliman et al. 2013). Other sources of uncertainty include lack of long-term production history, complex flow geometry, and variability in the completion properties (Agrawal et al. 2012).”. (Pg. 478).
“The DCA parameters, hj, will depend on the DCA model under consideration. To apply Bayesian inference to the deterministic DCA models, the parameters are linked to the Bayes theorem by assuming that the prior distribution is the distribution of DCA parameters before any production data have been observed. This can be a uniform distribution that is based on practical ranges of the parameters. The likelihood function is the conditional probability of the available production history given the DCA parameters. The posterior distribution is the distribution of the DCA parameters after all the available production history has been taken into account.”. (Pg. 478).
“Because of a lack of publicly available completions data, we analyzed wells that were completed in the same year with an underlying assumption that these wells will be completed in similar manner.”. (Pg. 482).
This shows the parameters recovered from each well’s production history being its initial rate, decline rate and decline exponent. The reference attributes the behavior these describes to the hydraulic fractures and stimulated volume created in the well and to variability in its completion properties, and it estimates them for wells grouped by how they were completed.
PARYANI, BABIN, SOFRONOV, and WANG do not explicitly teach but NARAYANASWAMY teaches
wherein the inverse proxy model inputs the production data and outputs the geological parameters and
“In this paper, we consider the problem of history matching in reservoir models, an inverse problem aimed at estimating and calibrating the parameters of a simulator to best match the empirical observations.”. (Pg. 1).
“Here, y ∈ RT×k corresponds to a multivariate time-series output from the simulator comprised of k quantities and T time-steps. Correspondingly, the inverse G : z → x is used to recover the input parameters from a sample in the latent space.”. (Pg. 2).
This shows a trained inverse network that takes the multi quantity time series observed from a reservoir and returns the reservoir parameters that produced it, so the recorded field response is what the inverse model receives as its input.
… that are indicative of potential inverse models.
“Since this estimation process is non-unique, determining a single history matched model that explains the observed measurements is an ill-posed problem [2]. Consequently, it is critical to estimate all possible inverse estimations to improve the reliability of a model [3]. In other words, rather than providing a single response for a given target observation yh one must find the posterior density on all possible parameter settings, i.e. p(x|y = yh).”. (Pg. 1-2).
“Note that, the inverse G is designed to produce both mean (µ) and variance (ν) estimates, and is optimized with the heteroscedastic loss (Gaussian likelihood assumption).”. (Pg. 2).
This shows the inverse model returning a mean and a spread for each parameter instead of a single answer, so its outputs describe a range of candidate parameter sets, each a possible model of the reservoir consistent with the observed data.
Regarding Claim 13, PARYANI, BABIN, NARAYANASWAMY, and WANG do not explicitly teach but SOFRONOV teaches
The method of claim 8, wherein the subsurface process data comprises diagnostics data; and
“As an example, a tool may be positioned to acquire information in a portion of a borehole. Analysis of such information may reveal vugs, dissolution planes (e.g., dissolution along bedding planes), stress-related features, dip events, etc. As an example, a tool may acquire information that may help to characterize a fractured reservoir, optionally where fractures may be natural and/or artificial (e.g., hydraulic fractures).”. (0054).
“In the example of Fig. 14, sensors (S) are located, for example, to monitor various parameters during operations. The sensors (S) may measure, for example, pressure, temperature, flowrate, composition, and other parameters of the reservoir, wellbore, gathering network, process facilities and/or other portions of an operation.”. (00134).
“Managing the cycle of water production (e.g., separation downhole or at surface, disposal, etc.) can involve a wide range of oilfield services, which can include data acquisition, diagnostics, production logging, water analysis, reservoir modeling (e.g., to characterize flow), and technologies to address water-related issues.”. (00141).
This shows the data collected from the field including downhole tool measurements that characterize the natural and hydraulic fractures around a well and sensor measurements of reservoir and wellbore pressure taken during operations, which are measurements made to diagnose the condition of the subsurface and the well.
and outputs the geological parameters and the parameters related to completions
“Such field data may be utilized in an inversion process that aims to reconstruct subterranean, spatial properties of a reservoir. Such spatial properties can include properties utilized in simulation fluid flow in the reservoir by a reservoir simulation framework. Such spatial properties can be assigned to a reservoir model such that the reservoir simulation framework simulates fluid flow based at least in part on such spatial properties.”. (00175).
“As an example, data received via the TECHLOG® framework may be utilized by one or more inversion algorithms, which may aim to generate spatially distributed properties of a subterranean environment.”. (0043).
“For example, consider a well in a shale formation that may include natural fractures, artificial fractures (e.g., hydraulic fractures) or a combination of natural and artificial fractures.”. (0041).
“Such information may assist with completions, stimulation treatment, etc. As an example, information acquired by a tool may be analyzed using a framework such as the aforementioned TECHLOG® framework (Schlumberger Limited, Houston, Texas). “. (0054).
This shows the inversion reconstructing the properties of a shale reservoir whose well intersects natural and hydraulic fractures, so its outputs describe both the rock and the stimulated fractures that assist with completions.
PARYANI, BABIN, SOFRONOV, and WANG do not explicitly teach but NARAYANASWAMY teaches
wherein the inverse proxy model inputs the diagnostics data and outputs the geological parameters
“In this paper, we consider the problem of history matching in reservoir models, an inverse problem aimed at estimating and calibrating the parameters of a simulator to best match the empirical observations.”. (Pg. 1).
“Here, y ∈ RT×k corresponds to a multivariate time-series output from the simulator comprised of k quantities and T time-steps. Correspondingly, the inverse G : z → x is used to recover the input parameters from a sample in the latent space.”. (Pg. 2).
This shows a trained inverse network that takes in the time series of measurements observed from a reservoir and returns the reservoir parameters that produced that response, so the recorded field measurements are what the inverse model receives as its input.
… that are indicative of potential inverse models.
“Since this estimation process is non-unique, determining a single history matched model that explains the observed measurements is an ill-posed problem [2]. Consequently, it is critical to estimate all possible inverse estimations to improve the reliability of a model [3]. In other words, rather than providing a single response for a given target observation yh one must find the posterior density on all possible parameter settings, i.e. p(x|y = yh).”. (Pg. 1-2).
“Note that, the inverse G is designed to produce both mean (µ) and variance (ν) estimates, and is optimized with the heteroscedastic loss (Gaussian likelihood assumption).”. (Pg. 2).
This shows the inverse model returning a mean and a spread for each parameter instead of a single answer, so its outputs describe a range of candidate parameter sets, each a possible model of the reservoir consistent with the observed data.
Claim 10 is rejected under 35 U.S.C 103 as being unpatentable over “Approximate Bayesian Computation for Probabilistic Decline-Curve Analysis in Unconventional Reservoirs” by PARYANI et al. [herein “PARYANI”] (2017), in view of US 10678967 B2 by BABIN et al. [herein “BABIN”] (2020), in view of WO 2017188858 A1 by SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), in view of “Deep-Learning based Inverse Modeling Approaches: A Subsurface Flow Example” by WANG et al. [herein “WANG”] (2020), and in further view of US 20130096900 A1 by USADI et al [herein “USADI”] (2013).
Regarding Claim 10, PARYANI, BABIN, SOFRONOV, NARAYANASWAMY, and WANG do not explicitly teach but USADI teaches
The method of claim 9, wherein the correlation of the physics-based rules is any one of linear, logarithmic, or exponential.
“Most reservoir simulation models assume linear potential flow. Darcy's law may be used to describe the linear relationship between potential gradients and flow velocity.”. (0006).
“"Permeability" is the capacity of a rock to transmit fluids through the interconnected pore spaces of the rock. Permeability may be measured using Darcy's Law: Q=(k .DELTA.P A)/(.mu.L), wherein Q=flow rate (cm.sup.3/s), .DELTA.P=pressure drop (atm) across a cylinder having a length L (cm) and a cross-sectional area A (cm.sup.2), .mu.=fluid viscosity (cp), and k=permeability (Darcy).”. (0053).
“In some exemplary embodiments, the surrogate solution may be a formulation of Darcy's law, and supervised machine learning may be used to generate a coarse scale approximation of the phase permeability of a coarse grid cell.”. (0061).
This shows Darcy’s law, which is described as a linear relationship reservoir models used for flow, being the law the machine learned flow surrogate is formulated on.
It would have been obvious before the effective filing date of the claimed invention to incorporate USADI’s teaching of Darcy’s law as a linear relationship between permeability, pressure gradient and flow with the method of PARYANI-BABIN-SOFRONOV- NARAYANASWAMY-WANG of training forward and inverse proxy models under a governing flow equation. The reason for doing so would have been to use, as the physics-based rule for a reservoir proxy, that reservoir simulation models themselves rely on. As expressed by USADI, “Most reservoir simulation models assume linear potential flow.”. (0006).
Claim 14 is rejected under 35 U.S.C 103 as being unpatentable over SOFRONOV et al. [herein “SOFRONOV”] (2017) and in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019).
Regarding Claim 14, SOFORNOV teaches
A computer-implemented method for generating an inverse proxy model in order to perform one or more subsurface operations in a subsurface, the method comprising:
generating, using a physics simulator solving differential equations, a training set of forward models;
“A system can include a processor; memory accessibly by the processor; and processor-executable instructions stored in the memory and executable to instruct the system to receive information associated with a reservoir in a geologic environment; predict one or more conditions based at least in part on the information; and output information …”. (0170).
“As an example, a ML inversion component may utilize one or more of historical production data, new geological data, and reservoir modeling results as a training set or training sets, for example, to generate a reservoir model adjuster.”. (00115).
“As an example, a ML production prediction component may utilize one or more training sets optionally with one or more of historical data and synthetic data, which may be generated by a reservoir modeling tool (e.g., the ECLIPSE® simulator, etc.)”. (00108).
“A simulator can be a computerized system (e.g., a computing system) that can execute instructions using one or more processors to solve a system of equations that describe physical phenomena subject to various constraints. The system of equations may be spatial defined (e.g., numerically discretized) according to a spatial model …”. (0027).
This shows a processor-based system that builds a machine learned inversion component and draws its training data from a reservoir simulator that solves the discretized equations describing the reservoir physical behavior.
generating, via machine learning using the training set of forward models, a forward proxy model;
“… for building a model that can be called by a prompt reservoir performance predictor. In such an example, the predictor can act as a fast data-driven tool for forecasting reservoir behavior, for example, as to one or more different development scenarios.”. (00108).
“As an example, a predictor can include one or more relatively simplistic physical models of a well or wells (see, e.g., Kaviani et al., How accurate are Capacitance Model connectivity estimates?, Journal of Petroleum Science and Engineering, Volume 122, October 2014, Pages 439-452; and Artun, Characterizing interwell connectivity in waterflooded reservoirs using data- driven and reduced-physics models: a comparative study, Neural Computing and Applications, pp. 1-15, January 2016). In such an example, a reservoir performance predictor may utilize a reservoir model for training and output of a set of coefficients for one or more systems of physics-based equations (e.g., capacity resistance model -like approaches). As an example, machine learning can include deep learning (see, e.g., Deng et al., "Deep Learning: Methods and Applications", Foundations and Trends in Signal Processing 7: 3-4, 2014), for example, with one or more multi-layered artificial neural networks.”. (00109).
This shows a machine learned predictor built from the simulator’s outputs that forecasts reservoir behavior quickly in place of running the simulator.
receiving subsurface process data;
“As an example, live data (e.g., available after initiation of a CRS exploitation) can include streaming production and/or sensor data, technology application data and non-technical information (e.g., market conditions, climate impact, etc.).”. (00100).
“For example, field data can be received by an input component and stored to a database of the database component.”. (00175).
This shows the system receiving production and sensor measurements from the field as they are acquired.
using one or more of the potential inverse models to perform the one or more subsurface operations in the subsurface.
“In the example of Fig. 10, the system 1000 can output an adjusted reservoir model. As shown, such an adjusted reservoir model can be based at least in part on machine learning-based prediction. As an example, an adjusted reservoir model may be utilized in a reservoir simulator (e.g., a tool) to generate simulation results, which may, in turn, be utilized for purposes of machine learning (e.g., by the algorithm component 1030) and further prediction (e.g., by the predictor component 1060).”. (00125).
“As shown in Fig. 11, the system 1 100 can output a field development plan (FDP), which can include one or more specified actions that can be implemented to develop a field. As shown, the FDP can be based at least in part on the prompt production prediction of the predictor component 1106. As an example, implementation of one or more FDP actions can result in changes to physical structures, fluids, etc.”. (00127).
“As an example, output information can include a field development plan for production of hydrocarbons via one or more pieces of field equipment at the field site.”. (00174).
This shows the adjusted reservoir model produced by the inversion being run to generate the predictions from which the system issues development plans and equipment control actions that are carried out in the field.
in order to generate outputs comprising geological parameters and parameters related to completions
“Such field data may be utilized in an inversion process that aims to reconstruct subterranean, spatial properties of a reservoir. Such spatial properties can include properties utilized in simulation fluid flow in the reservoir by a reservoir simulation framework.”. (00175).
“For example, consider a well in a shale formation that may include natural fractures, artificial fractures (e.g., hydraulic fractures) or a combination of natural and artificial fractures.”. (0041).
“As an example, a tool may be positioned to acquire information in a portion of a borehole. Analysis of such information may reveal vugs, dissolution planes (e.g., dissolution along bedding planes), stress-related features, dip events, etc. As an example, a tool may acquire information that may help to characterize a fractured reservoir, optionally where fractures may be natural and/or artificial (e.g., hydraulic fractures). Such information may assist with completions, stimulation treatment, etc.”. (0054).
This shows the inversion reconstructing the properties of a shale reservoir whose well intersects natural and hydraulic fractures, so its outputs describe both the rock and the stimulated fractures that assist with completions.
SOFRONOV does not explicitly teach but NARAYANASWAMY teaches
generating, via the machine learning using the training set of forward models, an inverse proxy model such that the inverse proxy model is consistent with the forward proxy model;
“This creates two coupled problems: First, fitting a forward model F : x→ ythat approximates the original simulation by mapping the inputs x to the outputs y; and Second, deriving an inverse model G : y → x.”. (Pg. 1).
“As illustrated in Fig. 1(a), this approach utilizes two independent fully-connected networks representing the forward model F and the inverse model G respectively that are tied together through a cyclical consistency constraint.”. (Pg. 2).
“The cyclical constraint ensures that the residuals from the inversion process, r = x − µ, are controlled by the sensitivities of the forward model F.”. (Pg. 2).
This shows an inverse network learned from the same simulation set as the forward network and trained so that its output, passed back through the forward network, reproduces the data, which ties the inverse model to the forward model.
using the subsurface process data as input to the inverse proxy model
“In this paper, we consider the problem of history matching in reservoir models, an inverse problem aimed at estimating and calibrating the parameters of a simulator to best match the empirical observations.”. (Pg. 1).
“Here, y ∈ RT×k corresponds to a multivariate time-series output from the simulator comprised of k quantities and T time-steps. Correspondingly, the inverse G : z → x is used to recover the input parameters from a sample in the latent space.”. (Pg. 2).
This shows a trained inverse network that takes in the time series of measurements observed from a reservoir and returns the reservoir parameters that produced that response.
that are indicative of potential inverse models; and
“Since this estimation process is non-unique, determining a single history matched model that explains the observed measurements is an ill-posed problem [2]. Consequently, it is critical to estimate all possible inverse estimations to improve the reliability of a model [3]. In other words, rather than providing a single response for a given target observation yh one must find the posterior density on all possible parameter settings, i.e. p(x|y = yh).”. (Pg. 1-2).
“Note that, the inverse G is designed to produce both mean (µ) and variance (ν) estimates, and is optimized with the heteroscedastic loss (Gaussian likelihood assumption).”. (Pg. 2).
The shows the inverse model returning a mean and a spread of each parameter instead of a single answer, so its outputs describe a range of candidate parameter sets, each a possible model of the reservoir consistent with the observed data.
It would have been obvious before the effective filing date of the claimed invention to incorporate NARAYANASWAMY’s teaching of learning an inverse model from the same simulation set as a forward surrogate under a consistency constraint with SOFRONOV’s method of building machine learned prediction and inversion components from reservoir simulation results and using the adjusted reservoir model to direct field operations. The reason for doing so would have been to obtain a reliable inverse model from a limited number of simulator runs. As expressed by NARAYANASWAMY’s “Our results show that self-consistency is highly beneficial, and both training strategies produce well calibrated inverse models.”. (Abstract).
Claims 15-16 and 18 are rejected under 35 U.S.C 103 as being unpatentable over SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), and in further view of “Deep-Learning based Inverse Modeling Approaches: A Subsurface Flow Example” by WANG et al [herein “WANG”] (2020).
Regarding Claim 15, SOFRONOV and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 14, wherein the machine learning to generate the forward proxy model is constrained by physics-based rules; and
“In the TgNN surrogate, physical/engineering constraints and other domain knowledge are incorporated as prior knowledge into the neural network training process.”. (Pg. 11 Section 4.2).
“In order to achieve theory-guided training, the DNN can then be substituted into the governing equation and boundary/initial conditions in Eq. (11), and the loss functions for physics-violation can be expressed as follows:”. (Pg. 12 Eq. 17-19).
“By incorporating physical laws, the TgNN surrogate can be constructed with limited labeled data, or even in a label-free manner.”. (Pg. 14).
This shows the surrogate network being trained against a loss that penalizes any departure from the governing flow equation and its boundary and initial conditions, in addition to mismatch with the simulation data, so the physical law limits what the model can learn.
wherein the machine learning to generate the inverse proxy model is constrained by the physics-based rules.
“The first three terms in the loss function ( , ) k h L[Symbol font/0x71] [Symbol font/0x71] constitute the physics-based loss, and the last two terms represent the data-driven loss or the data mismatch (Tartakovsky et al., 2020).”. (Pg. 18).
“Although the loss function is similar to that for TgNN surrogate construction, i.e., Eq. (20), the following two differences exist: first, the PINN model is constructed for a specific conductivity field, and thus no random variables exist in the neural network inputs; and second, two different neural networks are defined in the PINN model with two sets of network parameters ( ,k h [Symbol font/0x71] [Symbol font/0x71] ) needing to be trained simultaneously.”. (Pg. 18).
“The physics-based loss is also incorporated, as shown in the first three terms of Eq. (33). 10,000 collocation points are employed to enforce the physics constraints.”. (Pg. 36).
This shows the networks that recover the conductivity field from measurements being trained under the same governing equation, boundary conditions and initial condition terms used for the forward surrogate, with the paper stating that the two losses differ only in their inputs and number of networks and not in the physics applied.
It would have been obvious before the effective filing date of the claimed invention to incorporate WANG’s teaching of training both a forward and inverse network under the same governing equation with the method of SOFRONOV-NARAYANASWAMY of generating forward and inverse proxy models from a training set of simulated reservoir models. The reason for doing so would have been to build reliable proxy models from a limited number of simulators runs. As expressed by WANG, “By incorporating physical laws and other constraints, the TgNN surrogate can be constructed with limited simulation runs and accelerate the inversion process significantly.”. (Abstract).
Regarding Claim 16, SOFRONOV and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 15, wherein the forward proxy model receives input parameters and generates output parameters; and
“Considering the governing equation in Eq. (11), in order to approximate the quantity of interest ( , ; ) h t[Symbol font/0x77] x , a Deep Neural Network (DNN) is defined as follows (with ( ; ) K [Symbol font/0x77]x parameterized by ξ):”. (Pg. 11 Eq. 15).
“Therefore, the location, time, and stochastic parameters comprise the inputs of the neural network, i.e., ( , , ) t x ξ , as shown in Figure 1.”. (Pg. 11).
This shows the surrogate taking in the parameters that describe the conductivity field, along with location and time, and returning the hydraulic head, so it maps input parameters to an output parameter.
wherein the physics-based rules correlate one or more input parameters to one or more output parameters.
“… where N denotes a general nonlinear differential operator; [Symbol font/0x77] denotes the random variables in the probability space [Symbol font/0x57] ; D denotes the physical domain; T denotes the time span; ( ; ) K [Symbol font/0x77]x denotes the space-dependent coefficients, which can be seen as a random field; ( , ; ) h t[Symbol font/0x77] x denotes the quantity of interest or the solution of the problem …”. (Pg. 10).
“In order to achieve theory-guided training, the DNN can then be substituted into the governing equation and boundary/initial conditions in Eq. (11), and the loss functions for physics-violation can be expressed as follows:”. (Pg. 12 17-19).
“In this work, transient saturated subsurface flow problems are considered, which have a general form of governing equation: ( , ) ( ( ) ( , )) 0 s h t S K h t t [Symbol font/0xB6] [Symbol font/0x2D][Symbol font/0xD1][Symbol font/0xD7] [Symbol font/0xD1] [Symbol font/0x3D] [Symbol font/0xB6] x x x (34) where s S denotes the specific storage; ( , ) h tx denotes the hydraulic head; and ( ) Kx denotes the hydraulic conductivity.”. (Pg. 21).
This shows the physics-based rule applied during training is the governing flow equation, which links the conductivity filed that the network receives as its input parameters to the hydraulic head the network outputs, so that the head the network predicts must change with the conductivity it is given in the way the equation necessitates.
Regarding Claim 18, SOFRONOV and NARAYANASWAMY do not explicitly teach but WANG teaches
The method of claim 16, wherein the machine learning comprises a deep learning model.
“The first category is deep-learning surrogate-based inversion methods, in which the Theory-guided Neural Network (TgNN) is constructed as a deep-learning surrogate for problems with uncertain model parameters.”. (Abstract).
“Considering the governing equation in Eq. (11), in order to approximate the quantity of interest ( , ; ) h t[Symbol font/0x77] x , a Deep Neural Network (DNN) is defined as follows (with ( ; ) K [Symbol font/0x77]x parameterized by ξ):”. (Pg. 11 Eq. 15).
“The TgNN surrogate is first constructed. A seven-hidden-layer fully-connected neural network with 50 neurons per layer is chosen for TgNN surrogate construction.”. (Pg. 23).
This shows machine learning used to build the physics constrained surrogate (proxy) being a fully connected neural network with seven hidden layers that the paper describes as a deep learning surrogate (i.e. proxy model).
Claim 17 is rejected under 35 U.S.C 103 as being unpatentable over SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), in view of “Deep-Learning based Inverse Modeling Approaches: A Subsurface Flow Example” by WANG et al [herein “WANG”] (2020), and in further view of US 20130096900 A1 by USADI et al. [herein “USADI”] (2013).
Regarding Claim 17, SOFRONOV, NARAYANASWAMY, and WANG does not explicitly teach but USADI teaches
The method of claim 16, wherein the correlation of the physics-based rules is any one of linear, logarithmic, or exponential.
“Most reservoir simulation models assume linear potential flow. Darcy's law may be used to describe the linear relationship between potential gradients and flow velocity.”. (0006).
“"Permeability" is the capacity of a rock to transmit fluids through the interconnected pore spaces of the rock. Permeability may be measured using Darcy's Law: Q=(k .DELTA.P A)/(.mu.L), wherein Q=flow rate (cm.sup.3/s), .DELTA.P=pressure drop (atm) across a cylinder having a length L (cm) and a cross-sectional area A (cm.sup.2), .mu.=fluid viscosity (cp), and k=permeability (Darcy).”. (0053).
“In some exemplary embodiments, the surrogate solution may be a formulation of Darcy's law, and supervised machine learning may be used to generate a coarse scale approximation of the phase permeability of a coarse grid cell.”. (0061).
This shows Darcy’s law, which is described as a linear relationship reservoir models used for flow, being the law the machine learned flow surrogate is formulated on.
It would have been obvious before the effective filing date of the claimed invention to incorporate USADI’s teaching of Darcy’s law as a linear relationship between permeability, pressure gradient and flow with the method of SOFRONOV-NARAYANASWAMY-WANG of training forward and inverse proxy models under a governing flow equation. The reason for doing so would have been to use, as the physics-based rule for a reservoir proxy, that reservoir simulation models themselves rely on. As expressed by USADI, “Most reservoir simulation models assume linear potential flow.”. (0006).
Claim 19 is rejected under 35 U.S.C 103 as being unpatentable over SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), and in further view of “Hydraulic Fracture Design with a Proxy Model for Unconventional Shale Gas Reservoir with Considering Feasibility Study” by KIM et al. [herein “KIM”] (2019).
Regarding Claim 19, NARAYANASWAMY does not explicitly teach but SOFRONOV teaches
The method of claim 14, further comprising generating, using the forward proxy model, a number of forward models;
“As an example, synthetic data can include numerous automatically generated scenarios of field development with a wide range of potential technology impacts (e.g., reservoir treatment scenarios) and corresponding output from a prompt production prediction model.”. (00110).
“In such an example, the predictor can act as a fast data-driven tool for forecasting reservoir behavior, for example, as to one or more different development scenarios.”. (00108).
This shows the machine learning forward proxy producing a predicted output for each of numerous development scenarios. Each of those outputs is a generated forward model.
NARAYANASWAMY and SOFRONOV do not explicitly teach but KIM teaches
wherein the number of forward models generated by the forward proxy model is at least one order of magnitude greater than a number of the training set of forward models generated by the physics simulator and used for machine learning to generate the forward proxy model; and
“Samples for the proxy model were 100 training data sets by LHS based on the properties shown in Table 1.”. (Pg. 7).
“Figure 7 is the cross plot between the commercial simulation and the proxy model. It is the average of the productions from the 100 training data.”. (Pg. 7).
“The stage of HF was four (Figure 10), and the 14,641 designs of the hydraulic fracturing were examined assuming that the HF length varied from 50 to 550 ft for each stage (11 x 11 x 11 x 11) and HF conductivity was 30 md ft.”. (Pg. 9).
This shows a proxy fitted to 100 simulator training runs and then used to evaluate 14,641 fracture designs. That is about 146 times the number of training runs, which is more than two orders of magnitude.
wherein a computational time for the forward proxy model to generate one of the forward models is at least an order of magnitude less than the computational time for the physics simulator to generate one forward model in the training set of forward models.
“For comparing the computation time of the commercial software and the proxy model, it took 475 s to simulate one case by the commercial software (CMG), but it took just 0.000341 s for the proposed proxy model on a desktop with Intel Core 2 Duo CPU 3.”. (Pg. 10).
“It was about 1.4 million times faster than the full simulator compared.”. (Pg. 10).
This shows the proxy taking less time per case for the same reservoir model. That shows magnitude less computational time for each forward model.
It would have been obvious before the effective filing date of the claimed invention to incorporate KIM’s teaching of fitting a proxy on a small set of simulator runs and then evaluating far more cases with it, at a small fraction of the per case simulator time, with SOFRONOV-NARAYANASWAMY’s method of generating a machine learning forward proxy model from a simulator training set and using it to forecast many development scenarios. The reason for doing so would have been to cover a large design space that would take too long to run through the physics simulator alone, by spending simulator time only on the training runs. As expressed by KIM, “As shale gas production can be quickly evaluated in the various hydraulic fracture designs, the proposed proxy model solves the time-consuming problem of shale gas reservoir simulation.”. (Pg. 11).
Claim 20 is rejected under 35 U.S.C 103 as being unpatentable over SOFRONOV et al. [herein “SOFRONOV”] (2017), in view “Designing Deep Inverse Models for History Matching in Reservoir Simulations” by NARAYANASWAMY et al. [herein “NARAYANASWAMY”] (2019), and in further view of “Hydraulic Fracture Design with a Proxy Model for Unconventional Shale Gas Reservoir with Considering Feasibility Study” by KIM et al. [herein “KIM”] (2019).
Regarding Claim 20, SOFRONOV and NARAYANASWAMY do not explicitly teach but SUN teaches
The method of claim 14, further comprising:
validating, using the forward proxy model, the potential inverse models; and
“Moreover, history matching is an inverse solution process exhibiting strong non-uniqueness solution characteristics.”. (0004).
“The proxy model is used to predict the field response data based on the temporary optimal solution and the timeline of the validation data.”. (0028).
“The predicted field response data is compared with the field response data in the validation data set to evaluate the error function of the validation as described in Step 5.”. (0030).
This shows a candidate set of subsurface parameters from an inversion being run back through the machine learning forward proxy, with the predicted response checked against held out field data. That is validating a potential inverse model with the forward proxy model.
responsive to determining that the potential inverse models are invalid, using the proxy model in order to determine the geological parameters and the parameters related to completions.
“If the matching quality is unsatisfactory, the step 4 is re-executed, and the optimizer is re-initialized to find another temporary solution (i.e., a temporary petrophysical property parameter combination).”. (0030).
“The workflow uses a heuristic global optimization algorithm (e.g., particle swarm optimization and genetic algorithm) to minimize the history matching error that compares the model prediction with the tuning historical data. During the optimization, the proxy model trained in the first portion of the workflow will be called to evaluate the error function.”. (0033).
This shows that when the validation check fails, the forward proxy is used again, through reinitialized optimizer, to determine a new set of subsurface parameters in place of the rejected ones.
It would have been obvious before the effective filing date of the claimed invention to incorporate SUN’s teaching of checking each inverted parameter set through the machine learning forward proxy against held out data and re determining the parameters with the forward proxy when the check fails, with SOFRONOV-NARAYANSWAMY’s method of generating potential inverse models from subsurface process data using an inverse proxy model that is consistent with a forward proxy model. The reason for doing so would have been to deal with the non-uniqueness of the inverse problems by only candidate inverse models that reproduce the observed data are kept, and a failed candidate is replaced rather than used for subsurface operations. As expressed by, “Compared with a current history matching technology, the workflow of the present disclosure shows the advantage of effectively finding the non-unique solution. More importantly, since each solution is obtained and validated by tuning and validating different combinations of the data, the solution represents versatile physical understandings of an underground environment through a mathematical model.”. (0034).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
US9372943B2 teaches systems methods, and computer program products that offer advanced techniques to simulate dynamic systems (e.g., hydrocarbon reservoirs) by using one or more of the following methodologies: visualizing a multi-dimensional parameter space with a smaller number of dimensions (e.g., two or three dimensions) and conditioning a parameter space with response data to narrow the parameter space to a plurality of possibilities.
US7584081B2 teaches a method, system and apparatus for real time reservoir model updating using an ensemble Kalman filter.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NARCISO EDUARDO MONTES whose telephone number is (571)272-5773. The examiner can normally be reached Mon-Fri 8-5.
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, REHANA PERVEEN, can be reached at (571) 272-3676. 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.E.M./Examiner, Art Unit 2189
/REHANA PERVEEN/Supervisory Patent Examiner, Art Unit 2189
Prosecution Insights