DETAILED ACTION
This action is in response to the amendments filed on Feb. 17th, 2026. A summary of this action:
Claims 1-23 have been presented for examination.
Claims 1 and 14 are objected to because of informalities
Claim 7-8 and 19-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite
Claim 22 and 23 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement
Claim(s) 1-4, 6-9, 14-22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/
Claim(s) 5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Pettersen, Øystein. "Basics of reservoir simulation with the eclipse reservoir simulator." Lecture Notes. University of Bergen, Norway 114 (2006): 81. URL: mj-oystein(dot)no/index_htm_files/ResSimNotes(dot)pdf
Claim(s) 10-13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Miller, Brad, et al. "Back to the future: Malware detection with temporally consistent labels." Under submission (2015).
Claim(s) 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Diaz, J. C., W. R. Jines, and T. Steihaug. "On a convergence criterion for linear (inner) iterative solvers for reservoir simulation." SPE Reservoir Simulation Conference. SPE, 1985
This action is Final
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments/Amendments
Regarding the objections
Withdrawn in part in view of amendments, maintained in part.
Regarding the § 101 Rejection
Withdrawn in view of amendments and remarks (remarks at prong 1, in view of instant disclosure ¶¶ 38, 42, 102, and 104, as a practical application of the abstract idea), as well as the updated guidance regarding Ex Parte Desjardins (specifically, contrast alleged improvement with the improvement noted in Ex Parte Desjardins, wherein the instant improvement is to simulator technology without claiming the math itself of the simulations) and Example 39.
Regarding the § 102/103 Rejection
Maintained, updated below as necessitated by amendment.
With respect to the remarks, see the rejection below for how the newly amended limitation are taught by the relied upon combination of prior art.
Claim Interpretation
Claim 5 recites “with a relaxed time-step strategy” – the Examiner notes that such strategy’s are well-known in the art, and as such POSTIA would have readily recognized that this was referring to a relaxation convergence method, as commonly found in simulators, including the assignee’s own Eclipse reservoir simulator. E.g. see the relied upon Pettersen reference, § 17 which discusses this.
The Examiner is noting this because in view of ¶¶ 40 and 98, this limitation may appear to be a subjective limitation (potentially indefinite under § 112(b)) for the term “relaxed”, but POSITA would have readily known what scope the claims were directed to, for it is a well-known feature in the prior art, long in use. MPEP § 2111.01(I and III) to further clarify.
Furthermore, the claims contain the term “heuristic” – this is interpreted in view of ¶ 69: “simulator's existing heuristic algorithms” and ¶ 99: “the simulator's underlying heuristic algorithms to” and ¶¶ 2-3 incl.: “There are many heuristic techniques of selecting time-step size used in various simulation models”.
Claim Objections
Claims 1 and 14 are objected to because of the following informalities:
Independent claims 1 and 14 recite “via a simulator” – which may cause ambiguity with respect to § 112(f). To clarify, this ambiguity is furthered by the claim setting forth the limitation to performing a simulation is via a simulator, wherein the processor is distinctly claimed from this and only used for some of the other limitations – the Examiner suggests, in view of ¶ 72, to more clearly reflect this is simply performing/executing a simulation using the selected step-size values to obtain the results, or similar such verbiage (e.g. inputting the time-step sizes to perform a simulation to obtain a result) that does not create § 112(f) ambiguity, as there is no structure clearly linked to the simulator - ¶¶ 81-82 and fig. 10.
Independent claim 1 also objected to because only some of the limitations require the use of the processor. Examiner suggests for conciseness amending the claims to recite the processor in the preamble instead, or a computer, e.g. “A computer-implemented method”. See MPEP § 2111 for In re Prater to clarify on this objection.
Appropriate correction is required.
Claim Rejections - 35 USC § 112(b)
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 7-8 and 19-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention. The dependent claims inherit the deficiencies of the claims they depend upon.
MPEP § 2173.05(b)(IV): “A claim term that requires the exercise of subjective judgment without restriction may render the claim indefinite. In re Musgrave, 431 F.2d 882, 893, 167 USPQ 280, 289 (CCPA 1970). Claim scope cannot depend solely on the unrestrained, subjective opinion of a particular individual purported to be practicing the invention. Datamize LLC v. Plumtree Software, Inc., 417 F.3d 1342, 1350, 75 USPQ2d 1801, 1807 (Fed. Cir. 2005));”
Representative claim 7:
The method of claim 1, wherein generating the first set of data comprises determining whether the first size meets a predetermined criteria defining an optimal size.
Representative claim 8:
The method of claim 7, wherein generating the first set of data comprises devising the training set using the optimal size.
These are indefinite, because there is no objective criteria for ascertain what is an “optimal size” and what is not. The “criteria” is merely stating any standard that is pre-determined, i.e. by the subjective opinion of POSITA. See ¶¶ 34-42 and original claims as filed.
Examiner suggests cancelling these claims.
To clarify, no objective standard is clearly set forth to ascertain what is required to be an “optimal size” for the time step size, and what is not “optimal”.
Claim Rejections - 35 USC § 112(a)
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claim 22 and 23 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention. The dependent claims inherit the deficiencies of the claims they depend upon.
See MPEP 2163(II)(A): "For example, in Hyatt v. Dudas, 492 F.3d 1365, 1371, 83 USPQ2d 1373, 1376-1377 (Fed. Cir. 2007), the examiner made a prima facie case by clearly and specifically explaining why applicant’s specification did not support the particular claimed combination of elements, even though applicant’s specification listed each and every element in the claimed combination. The court found the "examiner was explicit that while each element may be individually described in the specification, the deficiency was lack of adequate description of their combination" and, thus, "[t]he burden was then properly shifted to [inventor] to cite to the examiner where adequate written description could be found or to make an amendment to address the deficiency.""
Also, see MPEP 2163(I) for Lockwood v. Amer. Airlines, Inc., 107 F.3d 1565, 1572, 41 USPQ2d 1961, 1966 (Fed. Cir. 1997).
Claim 22: The method of claim 21, wherein the mathematical parameters further comprise error estimates and a number of iterations.
The term error estimates appear in ¶¶ 5 and 101. Similar with number of iterations. To clarify, ¶ 101: “concepts such as error estimates, convergence theorems, number of iterations, etc.”
The mathematical parameters that are consumed as listed in ¶¶ 30-31 (“…The considered parameters include the previous time-step size, the magnitude of solution updates and other measures of the characteristics of the solution (such as CFL number), the convergence conditions, the behavior of both non-linear and linear solvers, well events, the type of fluid, and recovery methods used….”), and these do not include error estimates or number of iterations.
As such, the particular combination recited in claim 22 is not sufficiently described.
Claim 23:
The method of claim 1, wherein the measuring the convergence condition of the simulation of the reservoir workflow process based on the ML model having the first time-step comprises measuring a non-linear convergence condition.
This is not sufficiently described. While the specification discloses that the equations to be solved may be non-linear (¶ 38) and a non-linear solver (¶ 31), it does not sufficiently describe that the convergence condition is itself non-linear.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-4, 6-9, 14-22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/
Regarding Claim 1
Rowan teaches:
A method for modeling a reservoir comprising: (Rowan, abstract)
receiving, using one or more computing device processors, a reservoir model associated with a reservoir workflow process; modifying, using the one or more computing device processors, the reservoir model associated with the reservoir workflow process using an optimum time-step strategy; extracting, using the one or more computing device processors, features from the reservoir model along with first time-step sizes; generating, using the one or more computing device processors, a first set of data for devising a training set using the first time-step sizes; collecting, using the one or more computing device processors, a selected amount of the first set of data for the training set;… triggering a using the training set;
generating, using the one or more computing device processors, a machine learning (ML)
model having the first time-step using the training set; performing a simulation of the reservoir workflow process via a simulator based on the ML model; measuring a convergence condition of the simulation during the performance of the simulation;
Rowan, abstract and ¶¶ 1-2, then see fig. 2.1 in particular, then see ¶ 26-29, then ¶ 45: “In Block 201, a set of historical parameter values of a runtime parameter and [0046] a set of historical core datasets are obtained. In particular, the set of historical parameter values and the set of historical core datasets relate to a simulation of the field. In one or more embodiments, each historical core dataset is obtained from a successful model computation of the simulation at a particular simulation time point. Further, a corresponding historical parameter value is used for a subsequent successful model computation at the next simulation time point immediately following the particular simulation time point. Specifically, each historical parameter value results in a simulation convergence during the simulation.” Then see Rowan, ¶¶ 47-48
then see ¶ 54: “In Block 207, the model computation is performed using the predicted parameter value of the runtime parameter and iterates till simulation convergence is actually achieved at the current simulation time point. Different types of runtime parameters may be used. For example, in one or more embodiments, the runtime parameter is the simulation time step used to compute the current simulation time point from the most recent simulation time point. Specifically, the most recent simulation time point is incremented by the predicted target value to determine a predicted simulation time point for the initial iteration of the modeling computation” – to further clarify, ¶ 66: “As described above, FIG. 3.1 shows the simulation time line (310) where a runtime parameter value (e.g., time step ti+I) is determined upon a successful model computation at the time point ti, to continue simulation for the next time point ti+J. The runtime parameter value (e.g., size of the time step ti) is related to the complexity of the model computation and the computing resources (e.g., computing time) for the model computation.” And ¶ 67: “Selecting time step size based on the current and recent historical state of the simulation (e.g., simulated reservoir properties) is not well suited to predict a target time step size for sudden changes in the simulation input. These changes may occur when wells that are connected to the reservoir come online and/or are closed. Moreover, injection of fluids or other local physical effects may temporally increase the complexity of the flow equations. A target size of the time step is crucial to compute the solution in reasonable time…An example of selecting a target time step under sudden increases in computing complexity is described below” – i.e. this receives a reservoir simulation (abstract and ¶¶ 1-2 clarifies) along with data associated with it, including the “size of the time step”, and performs an iterative process (see fig. 2.1 including the loop, as clarified on below) so as to modify the later simulations using a machine learning model that predicts “size of the time step”, wherein at each iteration the “current core dataset” including the step size is obtained from the “successful model computation iteration” (fig. 2.1, and accompanying description cited to below, e.g. ¶¶ 38-39, 72, 76, etc. clarifies on this) along with other data which forms a “fingerprint” of the simulation (e.g. ¶¶ 69-76) which is added to the training set at each iteration (¶¶ 75-76)
note in particular # 204 as discussed in ¶ 29: “In Block 204, a determination is made as to whether the simulation is completed. If the determination is positive, i.e., the simulation is completed, the method proceeds to Block 208. If the determination is negative, i.e., the simulation is not yet completed, the method proceeds to Block 205” in view of ¶ 55: “In one or more embodiments, the runtime parameter value is adjusted during the model computation iterations leading to the actual simulation convergence” – i.e. its adjusting the runtime parameter value (the “size of the time step” as discussed above) until convergence is measured to have been completed - ¶ 38 to clarify: “At each time point t;, model computation is performed iteratively, till a convergence condition is met [measured at each iteration, and when it is met the simulation is complete], to solve the equations that model the fluid flow in the porous media. In particular, each time point ti corresponds to a successful model computation that is completed with one or more iterations”
to clarify, see Rowan ¶ 68: “The input value set (311) of the simulation specifies an open well injection of fluid at time point t3, which increases the complexity of model computation and results in excessive iterations of unsuccessful model computation. After each iteration of unsuccessful model computation, the subsequent runtime parameter value (e.g., time step size) is reduced for the next iteration until the model computation is finally successful. Due to the excessive number of failed iterations, the runtime parameter value used for the final successful iteration (e.g., subsequent time step t4) is substantially smaller than the runtime parameter value used for the previous time point (e.g., the previous time step t3).” Then ¶ 69: “…In particular, each core dataset corresponding to the time point ti is represented as a tuple { {X, Di}, Yi} or { {X, D}, y}h, where X, D, and y represent a continuous data vector, discrete data vector, and runtime parameter value, respectively. These core datasets are collected and stored in a fingerprint F = { { {X1, D1 }, y1 }, ... ,{ {Xn, Dn}, Yn} }, which is included in a training set to generate the aforementioned machine learning model…In addition, y i includes a runtime parameter value used in the final successful model computation iteration at time point ti+J.” – and ¶ 72 to further clarify
also, ¶ 73: “FIG. 3.2 shows a schematic diagram of improving simulation by capturing core datasets into a training set for the machine learning model. In particular, the simulation time line A (321), simulation time line B (322), and simulation time line C (323) are similar to the simulation time line (310) depicted in FIG. 3 .1 above…In addition, separate simulation studies may be performed for different reservoirs. Although the realization 1 and realization 2 are shown as contributing to the training set (320), the fingerprints of more than two realizations for multiple simulation studies may be included in the training set (320) for generating the machine learning model…”
as to the machine learning model, see ¶¶ 74-75, including: “For example, a tree ensemble classifier may be used.” Which POSITA would have recognized was a decision tree ensemble classifier
also, ¶ 76: “In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases, to the extent that the algorithm to may replace existing time step selection heuristics within the modeling engine to achieve target runtime performance in the general simulation case.” – i.e. Rowan updates the training set every iteration
to further clarify, ¶ 65: “As described above, the runtime parameter engine is an intelligent selector for the runtime parameter value that builds a fingerprint of a simulation based on multiple realizations. The runtime parameter engine captures specific data of each successful model computation in each realization to store in the fingerprint. Each realization improves the accuracy of the selected target runtime parameter value as more information about the complexity of the fluid flow equations and boundary effects (such as wells) becomes available” – see the instant disclosure, ¶¶ 41 and 71 which discuss a “fingerprint of the model” for relevance
updating, during the performance of the simulation based on the ML model having the first time-step, the ML model to have a second time-step using the training set, wherein the second time-step has a second size different than the first size; updating, before a convergence of the simulation performed based on the ML model having the first time-step, the simulation to be based on the updated ML model having the second time-step; measuring the convergence condition of the updated simulation; (Rowan, as cited above, incl. fig. 2.1 for its iterative loop, to clarify, see ¶¶ 38-39: “As an example, the simulation time line (310) represents a simulation of fluid flow in porous media starting from the initial simulation time to till the ending simulation time tn. At each time point t;, model computation is performed iteratively, till a convergence condition is met, to solve the equations that model the fluid flow in the porous media. In particular, each time point ti corresponds to a successful model computation that is completed with one or more iterations. Specifically, the successful model computation is based on meeting the convergence condition, which is referred to as a simulation convergence… Upon successful model computation at the time point ti, a runtime parameter value (e.g., time step ti+ 1) is determined to continue simulation for the next time point ti+ 1.” – and ¶ 45: “In one or more embodiments, each historical core dataset is obtained from a successful model computation of the simulation at a particular simulation time point. Further, a corresponding historical parameter value is used for a subsequent successful model computation at the next simulation time point immediately following the particular simulation time point. Specifically, each historical parameter value results in a simulation convergence during the simulation.” And ¶ 53: “In Block 206, a predicted parameter value of the runtime parameter is generated for achieving the simulation convergence during the current simulation. In particular, the predicted parameter value is a target value to achieve the simulation convergence for a current simulation time point immediately follows the most recent simulation time point from which the current core dataset is obtained. In one or more embodiments, the predicted parameter value of the runtime parameter is generated using the machine learning model and based on the current core dataset” – and ¶ 54: “In Block 207, the model computation is performed using the predicted parameter value of the runtime parameter and iterates till simulation convergence is actually achieved at the current simulation time point” to ¶ 55: “In one or more embodiments, the runtime parameter value is adjusted during the model computation iterations leading to the actual simulation convergence. In other words, the final runtime parameter value used for the successful model computation iteration may differ from the predicted parameter value of the runtime parameter. For example, the actual simulation time point of the successful model computation iteration may differ from the predicted simulation time point for the initial iteration of the modeling computation… The method then returns to Block 202 where the final runtime parameter value used for the successful model computation iteration and the current core dataset are added to the training set.”; and see ¶ 75: “This machine learning model (324) is updated each time new core dataset is added to the training set (320).” – i.e. as shown in fig. 2.1, and as described herein (Examiner noting ¶ 66 as well as discussed above: “The runtime parameter value (e.g., size of the time step ti) is related to the complexity of the model computation and the computing resources (e.g., computing time) for the model computation.”) the ML predicts a time step size for each simulation time point as the simulation iterates through, i.e. on the nth simulation time point with fig. 2.1, the ML model predicted a “size of the time step” as the “runtime parameter” (e.g. ¶ 50: “In another example scenario where at least one successful model computation has been completed for the simulation, the current core dataset is obtained from the latest successful model computation at the most recent simulation time point.”)and uses this to perform a simulation based on the predicted “size of the time step”, and then once that time step is successful (but during the performance of the overall simulation with its numerous time points, i.e. before the full simulation of all time points is complete) the training data set is updated with the results of the nth time step, the ML model updated from this (e.g. ¶ 75 and elsewhere above), and then for the nth+1 time point the ML model predicts a new time step for the nth+1th time point based on the updated model and updated training step [second time step for second time point being different then the first]
and for the updating before a convergence - see ¶ 55 as cited above: “…In other words, the final runtime parameter value used for the successful model computation iteration may differ from the predicted parameter value of the runtime parameter. For example, the actual simulation time point of the successful model computation iteration may differ from the predicted simulation time point for the initial iteration of the modeling computation. In other examples, the actual solver tolerance and/or variable change threshold used in the successful model computation iteration may differ from the predicted target values used in the initial iteration of the modeling computation.” – i.e. block 207 [i.e. the simulation is performed initially with the predicted time step, and updated before convergence, hence “the final runtime parameter value used for the successful model computation iteration may differ from the predicted parameter value”]
Rowan does not explicitly teach the follow features, but these would have been obvious when Rowan was taken in view of Wong:
generating a confidence level based on the second size of the second time-step; selecting, using the one or more computing device processors, the ML model having the second time-step based on the confidence level; continue performing the simulation using the ML model having the second time-step; (Rowan, as cited above wherein this was predicting time-step sizes using ML for a simulation between simulation time points, and in particular note ¶ 76: “In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases, to the extent that the algorithm to may replace existing time step selection heuristics within the modeling engine to achieve target runtime performance in the general simulation case” and ¶ 74: “Accordingly, the target time step is identified [selected] from the historical time steps of these closely matched historical core datasets. For example, the target time step … may be randomly selected from these historical time steps. In another example, the target time step… may be an average, median, geometric mean, or other statistical representation of these historical time steps.” – i.e. it selects the new step sizes from the data in the training set which is “closely matched”, but does not teach doing this with a confidence level – and ¶ 55 as well as discussed above
in view of Wong, abstract and see fig. 2B, in particular # 201-205, which is a similar invention in the field of simulation, also note col. 10, ¶ 2: “a comparatively large number of Newton steps (e.g., non-quadratic behavior), an increase in step time cut back”
and note in # 205: “the optimal simulation settings associated with an optimal simulation settings confidence value, the confidence value representative of how likely the predicted simulation settings feature vector, when applied as simulation settings for a future simulation of the first simulation type, will result in a simulation output that achieves a desired optimization” – as clarified in col. 9 lines 20-30: “The term "confidence value" refers to a programmatically generated number representative of a confidence in an accuracy of a trained machine learning model output. For example, if a machine learning model output is associated 25 with a higher confidence value in comparison to other outputs, it is more likely that the predicted output is accurate. Outputs from a machine learning model associated with a lower confidence value in comparison to other outputs may be modified according to modification rules, assigned a lower weight value, or discarded.” – it would have been obvious to have selected the highest confidence prediction when comparing the ML model output to other outputs (e.g. from the first step sizes and their associated simulation in Rowan) because it would have been “more likely that the predicted output is accurate”
determining, using the one or more computing device processors, whether results from the simulator require updating the training set (Rowan, as cited above which always updates the training set, in view of Wong, col. 10, lines 24-30: “The term "approval signal" refers to an electronic signal representative of an indication by a client device that a simulation output is an acceptable result. An acceptable result may be stored in the simulation repository for use in future training datasets.”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan on a system using ML to determine optimal runtime parameters for simulation with the teachings from Wong on a similar system using ML to determine optimal simulation settings, wherein the predictions are compared with a confidence value. The motivation to combine would have been that “For example, if a machine learning model output is associated with a higher confidence value in comparison to other outputs, it is more likely that the predicted output is accurate” (Wong, as cited above).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan on a system using ML to determine optimal runtime parameters for simulation with the teachings from Wong on a similar system using ML to determine optimal simulation settings, wherein it only stores simulation results into the training set that the user approves of. The motivation to combine would have been that this would have ensured that the user would be able to control which datasets are in the training set, thus ensuring all the datasets were “acceptable result[s]” to the user (Wong, as cited above).
While Rowan in view of Wong does not explicitly teach the following feature, Rowan, in view of Wong and in further view of Brownlee teaches:
determining, using the one or more computing device processors, whether the selected amount of the first set of data reaches a predetermined level; in response to the selected amount of the first set of data reaching the predetermined level (Rowan, as was taken in view of Wong above, for the training of Rowan
as taken in view of Brownlee, first page: “How much data do I need” as a “common question”, followed by the “Why…” section for the questions including “Do you have too little data? Consider confirming that you indeed have too little data. Consider collecting more data, or using data augmentation methods to artificially increase your sample size.” As well as “In practice, I answer this question myself using learning curves (see below), using resampling methods on small datasets (e.g. k-fold cross validation and the bootstrap), and by adding confidence intervals to final results.” - then # 6: “I would suggest performing your own study with your available data and a single well-performing algorithm, such as random forest.” (for relevance, Rowan, ¶ 75: “For example, a tree ensemble classifier may be used.”), followed by: “Plotting the result as a line plot with training dataset size on the x-axis and model skill on the y-axis will give you an idea of how the size of the data affects the skill of the model on your specific problem. This graph is called a learning curve. From this graph, you may be able to project the amount of data that is required to develop a skillful model, or perhaps how little data you actually need before hitting an inflection point of diminishing returns. I highly recommend this approach in general in order to develop robust models in the context of a well-rounded understanding of the problem.”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan, as was modified above on a system using ML for time step size prediction for simulation with the teachings from Brownlee on methods of determining how much data is needed for training a ML model. The motivation to combine would have been that “I highly recommend this approach in general in order to develop robust models in the context of a well-rounded understanding of the problem”.
Regarding Claim 2
Rowan teaches:
The method of claim 1, wherein receiving the reservoir model for the reservoir workflow process comprises information for creating the reservoir model. (Rowan, as cited above for the data in the fingerprint)
Regarding Claim 3
Rowan teaches:
The method of claim 1, wherein modifying the reservoir model associated with the reservoir workflow process comprises inputting time-step information. (Rowan, as cited above for the data in the fingerprint)
Regarding Claim 4
Rowan teaches:
The method of claim 1, wherein extracting the features from the reservoir model comprises receiving the first size from one or more heuristic options. (¶ 76: “In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases, to the extent that the algorithm to may replace existing time step selection heuristics within the modeling engine to achieve target runtime performance in the general simulation case.” – i.e. in earlier iterations, the time step sizes are from heuristic options
Regarding Claim 6
Rowan teaches:
The method of claim 1, wherein generating the first set of data comprises accessing direct physical quantities and derived mathematical properties of the reservoir.. (Rowan, the fingerprint as discussed above, e.g. ¶ 29: “In one or more embodiments, the training set (234) includes simulation fingerprints (e.g., simulation fingerprint A (235), simulation fingerprint B (236), simulation fingerprint C (237), etc.). Each simulation fingerprint represents characteristics of a realization and includes one or more core datasets and corresponding runtime parameter values… In particular, the core dataset A (235-1) may include a portion of simulation input/output data, of the realization A (233-1), at a particular simulation time point” – see ¶¶ 26-28 to clarify, and ¶¶ 69-72, and ¶ 67: “Selecting time step size based on the current and recent historical state of the simulation (e.g., simulated reservoir properties) is not well suited to predict a target time step size for sudden changes in the simulation input. These changes
may occur when wells that are connected to the reservoir come online and/or are closed.” And ¶ 2: “As an example, modeling may solve a complex set of non-linear partial differential equations that model the fluid flow in porous media over a sequence of simulation time points. The act of applying the computer model to solve the equations and generate resultant attribute values of the reservoir, geologic basin, petroleum system, etc. over the sequence of simulation time points is referred to as a simulation.” – i.e. ¶ 70 provides a list of examples of derived math properties, and ¶ 71 provides a list which includes a plurality of examples of direct physical quantities
Regarding Claim 7
Rowan teaches:
The method of claim 1, wherein generating the first set of data comprises determining whether the first size meets a predetermined criteria defining an optimal size.. (Rowan, as cited above, at ¶ 68: “The input value set (311) of the simulation specifies an open well injection of fluid at time point t3, which increases the complexity of model computation and results in excessive iterations of unsuccessful model computation. After each iteration of unsuccessful model computation, the subsequent runtime parameter value (e.g., time step size) is reduced for the next iteration until the model computation is finally successful. Due to the excessive number of failed iterations, the runtime parameter value used for the final successful iteration (e.g., subsequent time step t4) is substantially smaller than the runtime parameter value used for the previous time point (e.g., the previous time step t3).” – wherein only the results from successful [i.e. optimal] simulations were stored
To clarify, fig. 2.1 for the “convergence” as an example of the critera
Regarding Claim 8
Rowan teaches:
The method of claim 7, wherein generating the first set of data comprises devising the training set using the optimal size.. (Rowan, as cited above)
Regarding Claim 9
Rowan teaches:
The method of claim 7, wherein generating the first set of data comprises removing the first size that does not meet the criteria.. (Rowan, as cited above, ¶ 68, i.e. the failed simulations/iteration time step size values were removed that did not result in convergence – see fig. 2.1 to clarify)
Regarding Claim 14.
Rejected under a similar rationale as claim 1 above, wherein the combination of prior art relied upon above teaches:
A system for modeling a reservoir, the system comprising one or more computing device processors; and one or more computing device memories, coupled to the one or more computing device processors, the one or more computing device memories storing instructions executed by the one or more computing device processors, wherein the instructions are configured to: (Rowan, fig. 4.1)
Regarding Claim 15.
Rejected under similar rationale as claim 2 above.
Regarding Claim 16.
Rejected under similar rationale as claim 3 above.
Regarding Claim 17.
Rejected under similar rationale as claim 4 above.
Regarding Claim 18.
Rejected under similar rationale as claim 6 above.
Regarding Claim 19.
Rejected under similar rationale as claims 7-9 above.
Regarding Claim 20.
Rejected under similar rationale as claims 7-9 above.
Regarding Claim 21.
Rowan teaches:
The method of claim 1, further comprising measuring, via the ML model having the first time-step, a performance of the simulation including consuming data from a physical state of the reservoir and deriving mathematical parameters, wherein the mathematical parameters comprise the convergence condition. (Rowan, as cited above, incl. ¶¶ 38-39: “As an example, the simulation time line (310) represents a simulation of fluid flow in porous media starting from the initial simulation time to till the ending simulation time tn. At each time point t;, model computation is performed iteratively, till a convergence condition is met, to solve the equations that model the fluid flow in the porous media. In particular, each time point ti corresponds to a successful model computation that is completed with one or more iterations. Specifically, the successful model computation is based on meeting the convergence condition, which is referred to as a simulation convergence… Upon successful model computation at the time point ti, a runtime parameter value (e.g., time step ti+ 1) is determined to continue simulation for the next time point ti+ 1. The simulation is continued in the manner described above till the ending simulation time tn.”
And Rowan, the fingerprint as discussed above, e.g. ¶ 29: “In one or more embodiments, the training set (234) includes simulation fingerprints (e.g., simulation fingerprint A (235), simulation fingerprint B (236), simulation fingerprint C (237), etc.). Each simulation fingerprint represents characteristics of a realization and includes one or more core datasets and corresponding runtime parameter values… In particular, the core dataset A (235-1) may include a portion of simulation input/output data, of the realization A (233-1), at a particular simulation time point” – see ¶¶ 26-28 to clarify, and ¶ 65: “The runtime parameter engine captures specific data of each successful model computation in each realization to store in the fingerprint. Each realization improves the accuracy of the selected target runtime parameter value as more information about the complexity of the fluid flow equations and boundary effects (such as wells) becomes available.” and ¶¶ 69-72, and ¶ 67: “Selecting time step size based on the current and recent historical state of the simulation (e.g., simulated reservoir properties) is not well suited to predict a target time step size for sudden changes in the simulation input. These changes may occur when wells that are connected to the reservoir come online and/or are closed.” And ¶ 2: “As an example, modeling may solve a complex set of non-linear partial differential equations that model the fluid flow in porous media over a sequence of simulation time points. The act of applying the computer model to solve the equations and generate resultant attribute values of the reservoir, geologic basin, petroleum system, etc. over the sequence of simulation time points is referred to as a simulation.” – i.e. ¶ 70: “Examples of continuous data elements of the continuous data vector X include, but are not limited to, reservoir pressures and temperatures, fluid composition, well oil and gas production/injection rates, mobility of reservoir fluids [examples of physical states], mass and energy balance errors, etc….”, and ¶ 71: “Examples of discrete data elements of the discrete data vector Di include, but are not limited to, a number of nonlinear and/or linear solver iterations occurred during the model computation at time point ti, a number of open wells included in the model computation at time point ti, and additional modeling attributes of the model computation at time point ti. The additional modeling attributes may specify a number of active and inactive grid cells, a number of fluid phases modeled, a number of components used in a fluid model (e.g., a three component oil/water/gas system for modeling black oil or n-component hydrocarbon and water mixtures), whether injection of a specific fluid is included in the model computation, whether heat transfer and temperature dependence is included in the model computation, etc.”
To clarify this includes math parameters comprising the convergence condition, see ¶ 14: “For example, the historical parameter values and historical core datasets are used for a first simulation of the field where each historical parameter value results in a simulation convergence during the first simulation.” And ¶ 45, i.e. the simulation conditions at convergence are included in the training set
Regarding Claim 22.
Rowan teaches:
The method of claim 21, wherein the mathematical parameters further comprise error estimates and a number of iterations. (Rowan, as cited above for claim 21, in particular note ¶ 71 for the number of iterations; and in ¶ 70: “Examples of continuous data elements of the continuous data vector X include, but are not limited to, reservoir pressures and temperatures, fluid composition, well oil and gas production/injection rates, mobility of reservoir fluids, mass and energy balance errors, etc. that are generated during the model computation at time point ti.”)
Claim(s) 5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Pettersen, Øystein. "Basics of reservoir simulation with the eclipse reservoir simulator." Lecture Notes. University of Bergen, Norway 114 (2006): 81. URL: mj-oystein(dot)no/index_htm_files/ResSimNotes(dot)pdf
Regarding Claim 5
While Rowan, in view of Wong and Brownlee above, does not explicitly teach the following, this would have been obvious when said combination of art was taken in further view of Petersen:
The method of claim 1, wherein generating the first set of data comprises running a simulation model with a relaxed time-step strategy. (Rowan, as cited above including ¶ 76, i.e. the first sets of data are based on simulations with “existing time step selection heuristics”
Taken in view of Petersen, Section 17, starting on page 95, then see subsection “Accelerators – the point SOR method: “Equation (58) still has much room for improvement. Imagine that the update performed on a single iteration by Equation (58) changes the solution vector in the correct manner (“pushes it in the right direction”), but e.g. not enough. So if we could “push it a little further” the system would obviously (?) converge in fewer iterations. Such arguments are the foundation for many acceleration techniques. We first look at one called point SOR, where SOR means “successive over-relaxation... In conclusion, point SOR can be very efficient if a good relaxation parameter is found… We will not discuss the other SOR methods here, but go directly to the relaxation method used in Eclipse.” Followed by the section “Conjugate Gradients -Orthomin”, starting on page 96, including: “The relaxation parameter wk is tried optimized, but note that in contrast to the SOR methods, wk here is updated at each iteration… This is however a theoretical limit, and much too large to be of interest for typical problems. Experience has shown that the method converges in maximum a few tens of iterations. As a rule of thumb, if orthomin hasn’t converged in about 50 iterations, it is recommended to reduce the time step… But we shouldn’t stress these problems too much – by experience the method generally has excellent performance.” (pages 96-98)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan, as discussed above, on a system for reservoir simulation combined with ML wherein the initial data sets are results from reservoir simulation based on “existing time step selection heuristics” with the teachings from Pettersen on methods of reducing the number of iterations in a reservoir solver including by using of relaxation. The motivation to combine would have been that “But we shouldn’t stress these problems too much – by experience the method generally has excellent performance” (Petersen, as cited above, page 98)
Claim(s) 10-13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Miller, Brad, et al. "Back to the future: Malware detection with temporally consistent labels." Under submission (2015).
Regarding Claim 10.
Rejected under a similar rationale as claim 1 above, wherein while said combination of prior art does not explicitly teach the following, it would have been obvious when said combination was taken in further view of Miller:
determining whether the confidence level is below a threshold; and in response to the confidence level being below the threshold, updating, using the one or more computing device processors, the training set. (Rowan, including ¶ 76 as cited above: “In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases, to the extent that the algorithm may replace existing time step selection heuristics within the modeling engine to achieve target
runtime performance in the general simulation case”
as was taken in view of Wong above, see fig. 2B, in particular # 201-205, which is a similar invention in the field of simulation, also note col. 10, ¶ 2: “a comparatively large number of Newton steps (e.g., non-quadratic behavior), an increase in step time cut back…” and note in # 205: “the optimal simulation settings associated with an optimal simulation settings confidence value, the confidence value representative of how likely the predicted simulation settings feature vector, when applied as simulation settings for a future simulation of the first simulation type, will result in a simulation output that achieves a desired optimization” – as clarified in col. 9 lines 20-30: “The term "confidence value" refers to a programmatically generated number representative of a confidence in an accuracy of a trained machine learning model output. For example, if a machine learning model output is associated 25 with a higher confidence value in comparison to other outputs, it is more likely that the predicted output is accurate. Outputs from a machine learning model associated with a lower confidence value in comparison to other outputs may be modified according to modification rules, assigned a lower weight value, or discarded.”
And Wong, col. 10, lines 24-30: “The term "approval signal" refers to an electronic signal representative of an indication by a client device that a simulation output is an acceptable result. An acceptable result may be stored in the simulation repository for use in future training datasets.” – see Wong col. 12 lines 10-20 to clarify: “configure the apparatus to transmit 305A/305B for display at a client device the simulation output feature vector... upon receiving 306 an approval signal, the simulation output feature vector is added 307 to the simulation repository 102.”
As taken in further view of Miller, fig. 1 including the caption: “The detection pipeline employs the current model to detect malware, and the training pipeline produces the next model for use in the detection pipeline. During each retraining period, the training pipeline reviews all available training data and selects binaries for submission to the human labeling expert. Binaries labeled by the expert are combined with training data labeled using the current model and anti-virus scan results to serve as the training data for the next model.” Then see § 3: “When a binary arrives, the detection pipeline extracts the features, applies the current model to classify the binary as malicious or benign, and the training pipeline stores the binary in a database along with all other binaries seen to-date. During each retraining period, binaries not detected by scanners on VirusTotal are considered for submission to the human labeling expert. Binaries confidently detected by the current model are included in training data with a malicious label, and the remaining purportedly benign binaries are submitted to the human labeling expert as the query budget allows. The remaining un-submitted binaries are included in the training data as benign. At the end of the retraining period, the next model produced in the training pipeline replaces the current model and the process repeats” and § 3.4 ¶¶ 1-2 incl.: “Furthermore, since we trust our current detection model to a certain extent, we assign a malicious label to any binary which detection score exceeds a confidence threshold M. We call this heuristic auto-relabeling. If both of these heuristics fails to produce a (malicious) label, and if no known expert label is yet available for the binary, we submit the binary to the query strategy.”
Miller is considered as an analogous art as 1) it is in the field of machine learning, and implementations of machine learning, and 2) it is reasonably pertinent to the problem faced by the instant inventor of determining the confidence/reliability (¶ 99) of an ML model.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan, as was taken in combination above, on a system that constantly re-trained an ML model and updated the data set (Rowan, ¶¶ 75-77) with the teachings from Miller on detector of concept “drift” (page 3, col. 1, ¶ 2) so as to detect “changes over time” (abstract) in predictions of an ML model.
The motivation to combine would have been that “Schwenk et al. compare detection performance when iteratively retraining using the system’s own predictions as training labels with detection performance when training on labels generated after all data collection. By focusing on a system which retrains purely on the system’s own predictions, there is no means of incorporating or analyzing the effect of temporally available knowledge in the anti-malware community. Separate from Schwenk et al., several other works have identified the problem of time lag in sample ground truth [10,24]… This apparent performance degradation occurs because in reality, the system can only rely on past data to predict present and future samples” (Miller, § 2, subsection “Temporally Consistent Labels” - to clarify, § 8 ¶ 1: “We explain why machine learning systems perform very well in research settings and yet fail to perform reasonably in production settings by demonstrating the critical temporal factors of labeling, training, and evaluation that affect evaluation accuracy in real-world settings. We also provide guidelines for the proper temporal use of labeling during training and evaluation and show that the use of statistical machine learning in malware detection is not just promising, but can produce high quality, competitive results in a setting that more closely reflects realistic conditions.”
Regarding Claim 11.
Rejected under similar rationale as claim 1 above for the similar limitation.
Regarding Claim 12.
Rowan teaches:
The method of claim 10, wherein updating the training set comprises generating a second set of data. (Rowan, ¶¶ 74-76 as cited above, including: “This machine learning model (324) is updated each time new core dataset is added to the training set (320)… In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases...”, as was taken in combination above including with Miller)
Regarding Claim 13.
Rowan teaches:
The method of claim 12, wherein updating the training set comprises generating a second training set by appending the training set and the second set of data. (Rowan, ¶¶ 74-76 as cited above, including: “This machine learning model (324) is updated each time new core dataset is added to the training set (320)… In general, as more realizations of different simulations studies are run and new fingerprint data is added to the training set (320), the ability of the algorithm to predict a target time step based on knowledge of previous successful simulations increases...” , as was taken in combination above including with Miller)
Claim(s) 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Rowan et al., WO 2017/217957 in view of Wong et al., US 11,348,017 and in further view of Brownlee, “How Much Training Data is Required for Machine Learning”, May 23rd, 2019, article on Machine Learning Mastery, URL: machinelearningmastery(dot)com/much-training-data-required-machine-learning/ and in further view of Diaz, J. C., W. R. Jines, and T. Steihaug. "On a convergence criterion for linear (inner) iterative solvers for reservoir simulation." SPE Reservoir Simulation Conference. SPE, 1985
Regarding Claim 23.
While Rowan as modified above does not explicitly teach the following, it would have been obvious when Rowan, as modified above, was taken in further view of Diaz
The method of claim 1, wherein the measuring the convergence condition of the simulation of the reservoir workflow process based on the ML model having the first time-step comprises measuring a non-linear convergence condition. (Rowan, as cited above, for the measuring of the convergence of the simulation for each time point
Taken in further view of Diaz, abstract, then see page 2 subsection “SMITS STOPPING RULES” for an example of a non-linear converge Rence condition for reservoir simulation (see ¶¶ 1-4 in this subsection, as later clarified on in the subsection as “The effect of the particular choices of the stopping parameters is further illustrated in Figures 2 through 18. Figures 2 through 7 depict, in a 109-10g scale, the norm of the old nonlinear residual UF(xk)U versus the quotient of the new and old nonlinear residuals IIF(xk+1)U/IIF(xk)H for each Newton iterate. Superimposed on those figures are the stopping criteria corresponding to Table 1”
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings from Rowan on a system for reservoir simulation which included nonlinear solver iterations (Wang, ¶¶ 2 and 71) with the teachings from Diaz on stopping criteria for nonlinear reservoir simulators (Diaz, abstract) The motivation to combine would have been that “An optimal criterion should reflect the requirements for the nonlinear (outer-) iteration…. The design of this criteria includes knowledge of the approximate location of the boundary between the regions of linear and quadratic convergence for the nonlinear (outer-) iteration… The results illustrate that criteria using relaxed constraints in the linear convergence region are most efficient overall…”
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Wang, Qinghua, Graham Fleming, and Zhiqiang Gu. "A Multi-Pass Procedure for Redistributing Linearized Mass Balance Errors to Improve Simulator Performance." SPE Annual Technical Conference and Exhibition?. SPE, 2015. Abstract
Fung. US 2012/0179436. Abstract, ¶¶ 71-73
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DAVID A. HOPKINS whose telephone number is (571)272-0537. The examiner can normally be reached Monday to Friday, 10AM to 7 PM EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Ryan Pitaro can be reached at (571) 272-4071. 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.
/David A Hopkins/Primary Examiner, Art Unit 2188
Prosecution Insights