Method for predicting fluid flow fields

Abstract

A method for predicting fluid flow fields in a subsurface reservoir or aquifer inputs training sets of geologically realistic models, fluid properties, and reservoir properties to a dynamic simulator to generate a set of labelled geologically realistic models. The labels relate to reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation. The labelled geologically realistic models are then used to train a backpropagation-enabled process. The trained backpropagation-enabled process is then used to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation for a non-training model.

Description

METHOD FOR PREDICTING FLUID FLOW FIELDSFIELD OF THE INVENTION

[0001] The present invention relates to backpropagation-enabled processes, and in particular, to a method for predicting fluid flow fields in a subsurface reservoir or aquifer using a backpropagation-enabled process.BACKGROUND OF THE INVENTION

[0002] Fault seal analysis is important in hydrocarbon exploration and production, as well as in the selection of suitable subsurface formations for carbon sequestration. For these activities, it is important to understand the extent to which a fluid will migrate across a fault or be trapped by a fault.

[0003] One method of performing a fault seal analysis is to use a juxtaposition diagram, such as an Allan diagram (Allan, “Model for hydrocarbon migration and entrapment within faulted structures” AAPG Bulletin 73:7:803-811; 1989), to predict which traps have a high probability of containing hydrocarbons and the amount of hydrocarbons a trap can contain. A challenge with using a juxtaposition diagram is that the process is time-consuming. Accordingly, efforts have been made to predict fault seal behaviour using a computer model.

[0004] For example, Viard et al. (US 10209380B2, 19 Feb 2019) describes a method for juxtaposition across a geological discontinuity, such as a fault, by identifying the discontinuity in a set of seismic data and parameterizing each side of the discontinuity with first and second parameters. Isolines of each parameter are determined and used to divide the surface of the discontinuity. Overlapping areas of the divided surfaces are summed to model flow through the discontinuity.

[0005] Generally, fault seal analyses may be static or dynamic.

[0006] Static fault seal is the ability of a fault to maintain a column over geologic timescales. Apparent static seal can be calculated by measuring the column height and buoyancy pressure difference across a reservoir / reservoir juxtaposition.

[0007] As an example of a static fault seal analysis, Liu et al. (US10634805B2, 28 Apr 2020) relates to a method for predicting fault seal based on a digital seismic image. A fault coordinatesystem is defined parallel and perpendicular to a fault surface. Seismic amplitudes are extracted from a sub-volume surrounding the fault surface and are mapped in the fault coordinate system. Trace fitting of the seismic amplitudes is performed along directions locally perpendicular to the fault surface to separate fault seismic signal from other seismic energy. The fault seismic signal from the trace fitting is compared to a natural analogue and / or a synthetic model of responses to generate a predicted fault seal.

[0008] Dynamic fault seal is the ability of a fault to maintain a cross-fault pressure difference over the timescale during which fluids (e.g., hydrocarbons, CO2, water, and others) are injected and / or produced from a field. Conventional approaches are based on hydrocarbon production rate history-matching studies that report faults as either sealing, partially sealing, or open. Some examples of dynamic fault seal analysis are more quantitative and report actual transmissibility multipliers used for history-matching or production-induced cross-fault pressure differences.

[0009] Myers et al. (US8793110B2, 29 Jul 2014) provide that the current practice of evaluating effects of fault zone materials on cross-fault flow by a Shale Gouge Ratio (SGR) or Clay Smear Potential (CSP) is limited by predictive capability and robustness. Myers et al., therefore, aim to address the conventional problems of requiring local calibration of SGR to fault permeability, and failure to consider apparent low permeability of faults in high netgross (N:G) sections. Accordingly, Myers et al. describe a dynamic seal analysis method for predicting fluid flow by predicting a volume of shale in a fault, calculating the thickness of the fault, and calculating the permeability of the fault based on the shale volume, thickness, an estimated shale smear continuity and cataclastic reduction factors.

[0010] Many conventional approaches are limited to clastic formations. As discussed by Solum (“Static and dynamic fault seal potential in carbonates” Paper Wo FTS 05, presented at EAGE Fourth International Conference on Fault and Top Seals; Almeria, Spain, 20-24 September 2015), fault seal analysis for carbonates is more complicated than for faulted clastics due to the greater chemical reactivity of carbonates, and requires integration of geochemistry / diagenesis with structural geology.

[0011] Further, Solum et al. (“Toward the creation of models to predict static and dynamic fault-seal potential in carbonates” Petroleum Geoscience 23:70-91; 2017) provides a summary of carbonate reservoirs with cross-fault height difference that represent examples of fault seal.

[0012] Alpak et al. (“Simplified dynamic modelling of faulted turbidite reservoirs: A deeplearning approach to recovery -factor forecasting for exploration” SPE 197053 SPE Reservoir Evaluation & Engineering; November 2019) describe a deep-leaming-based predictor for estimating structure connectivity factor and the relative change in time to water breakthrough for a set of user-input options of fault patterns, fault transmissibility, reservoir permeability, and waterflood duration.

[0013] Alpak et al. (“Augmenting deep residual surrogates with Fourier neural operators for rapid two-phase flow and transport simulations” SPE 217441 SPE Journal; July 2023) relates to predictions of pressure and saturation fields using Fourier Neural operators with quasilinear time complexity to capture spectral information from feature maps.

[0014] There remains a need for improving the accuracy of fluid flow field predictions, and for improving the time required to interpret data for generating such a prediction.SUMMARY OF THE INVENTION

[0015] According to one aspect of the present invention, there is provided a method for predicting fluid flow fields in a subsurface reservoir or aquifer, comprising the steps of: providing a training set of geologically realistic models, wherein each of the geologically realistic models have one or more faults defining one or more fault blocks; providing a training set of fluid properties; providing a training set of reservoir properties; inputting the training set of geologically realistic models, the training set of fluid properties, and the training set of reservoir properties to a dynamic simulator to generate a set of labelled geologically realistic models with labels relating to reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation for each of the one or more fault blocks; training a backpropagation- enabled process using the labelled geologically realistic models to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation, thereby producing a trained backpropagation-enabled process; inputting a non-training model to the trained backpropagation-enabled process to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation.BRIEF DESCRIPTION OF THE INVENTION

[0016] The method of the present invention will be better understood by referring to the following detailed description of preferred embodiments and the drawings referenced therein, in which:

[0017] Fig. 1 is a schematic diagram of an embodiment of the present invention;

[0018] Fig. 2 is an example of one embodiment of a geologically realistic model used for training in the method of the present invention;

[0019] Fig. 3 is a schematic diagram showing a simplified example of inputs to and outputs from a dynamic simulator; and

[0020] Fig. 4 is an example of training properties inputted with a training set of geologically realistic models.DETAILED DESCRIPTION OF THE INVENTION

[0021] The present invention provides a method for predicting fluid flow fields. Training sets of geologically realistic models, fluid properties and reservoir properties are inputted to a dynamic simulator to generate labelled geologically realistic models with labels relating to reservoir pressure and one or both of reservoir fluid saturation and reservoir fluid mass accumulation for one or more fault blocks in the geologically realistic models. The labelled geologically realistic models are used to train a backpropagati on-enabled process to predict reservoir pressure and one or both of reservoir fluid saturation and reservoir fluid mass accumulation. The trained backpropagation-enabled process is then used on a non-training model to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation. The predictions may be generated for the effect on a reservoir or aquifer for different scenarios of injection and / or production well locations, injection and / or production rates, rock properties, and / or fluid properties.

[0022] By “dynamic simulator,” we mean a computer program to compute flow responses over time within voxels of a fault block in a three-dimensional reservoir or aquifer model.

[0023] By “geologically realistic model,” we mean a digital three-dimensional representation of a subsurface reservoir or aquifer in which the spatial dimensions of the reservoir and the number, position, spacing, and displacement of faults cutting the reservoir are within the range of what has been observed in nature.

[0024] Referring to Fig. 1 illustrating a schematic of the method of the present invention 10, a training set of geologically realistic models 12, a training set of fluid properties 14, and a training set of reservoir properties 16 are provided. The training sets of geologically realistic models 12, fluid properties 14, and reservoir properties 16 are inputted to a dynamic simulator 20. The output of the dynamic simulator 20 is a set of labelled geologically realistic models 22. The labelled geologically realistic models 22 are used to train a backpropagation-enabled process 30 to predict flow-simulation solution variables selected from reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation. In a preferred embodiment, the backpropagation-enabled process 30 is trained to predict additional flow-simulation solution variables including, without limitation, reservoir density, and reservoir composition. Preferably, the backpropagation-enabled process 30 is trained to predict flow-simulation solution variables as a function of time.

[0025] The training set of geologically realistic models 12 may be based on field-acquired data and / or augmentations of field-acquired data. Field-acquired data comprises seismic data, well data, and data learned from surface outcrops. Preferably, the geologically realistic models are generated synthetically.

[0026] Synthetic models may be generated, for example, without limitation, according to Griffith et al. (US2021 / 0223422A1 and / or US2021 / 0223423 Al, 2021 Jul 22) and / or according to Zarian et al. (WO2021 / 259912A1, 2021 Dec 30).

[0027] As an example, a set of models with a realistic range of reservoir permeabilities and fault patterns can be generated using a software, for example, without limitation, PETROSIGNS® Earth (Shell). Fault populations statistics may be produced from compilations of domain knowledge and data.

[0028] Examples of backpropagation-enabled processes include, without limitation, artificial intelligence, machine learning, and deep learning. It will be understood by those skilled in the art that advances in backpropagation-enabled processes continue rapidly. The method of the present invention is expected to be applicable to those advances even if under a different name. Accordingly, the method of the present invention is applicable to the further advances in backpropagation-enabled process, even if not expressly named herein.

[0029] A preferred embodiment of a backpropagation-enabled process is a deep learning process, including, but not limited to a convolutional neural network.

[0030] The backpropagati on-enabled process may be supervised, semi -supervised, or a combination thereof. In one embodiment, a supervised process is made semi-supervised by the addition of an unsupervised technique. In another embodiment, a subset of the seismic data is labelled in a semi-supervised process. Examples of a semi-supervised backpropagation-enabled process include, without limitation, a semi-supervised VAE process and a semi-supervised GAN process.

[0031] In one embodiment, the supervised backpropagation-enabled process is a classification process. The classification process may be conducted voxel-wise, slice-wise and / or volume-wise.

[0032] By using data from the synthetic models to train a backpropagation-enabled process, the effectiveness and accuracy of the training is significantly improved. The use of synthetic data for training a backpropagation-enabled process for fluid flow fields has two principal benefits. First, the model and associated labels can be generated in accordance with the present invention in a significantly shorter period of time, with related cost savings. Conversely, the generation of labels from field-acquired data can take years and involves sorting through excess details of information. Also, the interpretation and labelling of field-acquired data has a degree of human error and / or bias involved. In accordance with the present invention, the synthetic training model is substantially free of human error.

[0033] In accordance with the method of the present invention, synthetic geologically realistic models are generated to produce imaginary realizations of subsurface features. The models are generated by introducing variations in the subsurface features. The variations can be geologically realistic features simulating the outcome of a geologic process, simulations of geologic processes, simulations of noise sources, and combinations thereof.

[0034] By “model variation,” we mean introducing a change in a 3D series of layers having substantially horizontal and parallel boundary layers. Model variations preferably provide different types of geologically realistic features simulating the outcome of a geologic process, simulation of a geologic process, simulation of a noise source, and combinations thereof.

[0035] By generating a plurality of synthetic geologically realistic models, many subsurface features can be labelled to train the backpropagation-enabled process with different models of the same subsurface features and / or images of different subsurface features. An advantage of using synthetic geologically realistic models in the method of the present invention is the abilityto efficiently generate a significant number of images for training. By producing models of a subsurface feature in different scenarios, the training accuracy of the backpropagation-enabled process is improved. For example, labels identifying a river that is wide and shallow in one realization, narrow and deep in another realization, and wide and deep in yet another realization will more effectively train a backpropagation-enabled process to learn what a river looks like in field-acquired data.

[0036] By “augmented data,” we mean field-acquired and / or synthetically generated data that is modified, for example, by conventional DL data-augmentation techniques, as described in Taylor et al. (“Improved deep learning with generic data augmentation” IEEE Symposium - Symposium Series on Computational Intelligence SSCI 2018 1542-1547; 2018) which describes conventional augmenting by geometrical transformation (flipping, cropping, scaling and rotating) and photometric transformations (amending colour channels to change lighting and colour by colour jittering and Fancy Principle Component Analysis). Augmented data may also be generated, for example, as described in Liu et al. (US2020 / 0183035A1, 2020 Jun 11), which relates to data augmentation for seismic interpretation, recognizing that standard data augmentation strategies may produce limited plausible alternative samples and / or may lead to geologically or geophysically infeasible to implausible alternative samples. The machine learning method involves extracting patches from input data and transforming that data based on the input data and geologic and / or geophysical domain knowledge to generate augmented data. Transforming data is selected from an identity transformation, a spatial filter, a temporal filter, an amplitude scaling, a rotational transformation, a dilatational transformation, a deviatoric transformation, a resampling using interpolation or extrapolation, a spatial and temporal frequency modulation, a spectral shaping filter, an elastic transformation, an inelastic transformation, and a geophysical model transformation. In another embodiment, two pieces of data are blended together to generate a new piece of data. Other geophysical augmenting methods may also be used to generate augmented data. The labels may be preserved or modified in the augmentation. In this way, the training sets may be augmented to improve the model by introducing variations of data without requiring resources of acquiring and labelling field- acquired data or generating new synthetic data.

[0037] Each of the geologically realistic models 12 have one or more faults defining one or more fault blocks. Each fault block is comprised of a plurality of voxels. A fault block may bebounded by faults and / or a contact boundary. Fig. 2 shows a training geologically realistic model 12, having a contact boundary 34, a fault 36, and a fault bock 38. The contact boundary 34 may be a representation of contact between hydrocarbons and water in a subsurface formation. Fault block 38' is defined by faults 36, while fault block 38" is defined by a fault 36 and a contact boundary 34.

[0038] Each fault 36 in the training set of geologically realistic models 12 has fault properties. Fault properties include, without limitation, displacement, throw, heave, location, orientation, cross-fault juxtaposition, fault core width, fault permeability, and combinations thereof. By “fault juxtaposition location,” we mean a location in the subsurface where there is a bedding plane discontinuity across which there is cross-fault reservoir / reservoir juxtaposition. Fault properties may be defined by assigned GPS coordinates and / or depth contours. Fault properties may also be defined by relationships within a model. For example, a fault location may be defined by spacing from another fault. Preferably, the fault properties are based on computation of fault statistics constrained to be realistic. One or more of the fault properties may change along the length of the fault. For example, a fault 36 preferably has a fault core width that fluctuates along the length of the fault 36, such that certain voxels may have a fault core width of 0 while other voxels may have a fault core width greater than 0.

[0039] Examples of fluid properties for the training set of fluid properties 14 include, without limitation, type of fluid, composition, density, viscosity, gas / oil ratio, buoyancy pressure, and combinations thereof. Certain fluid properties may change within a model 12 due to pressure, depth, temperature, interaction with other fluids, contact with certain lithologies, column height, and combinations thereof. Accordingly, in a preferred embodiment, a voxel in the model 12 may be selected as a control point and values of fluid properties in other voxels are calculated based on the properties of the control point voxel.

[0040] The faults 36 in the training set of geologically realistic models 12 and / or the training set of fluid properties may be further enhanced with a geologically realistic parameter selected from the group consisting of scaling relationships, burial depth, stress state, reactivation history, temperature, diagenesis, and combinations thereof.

[0041] Examples of reservoir properties for the training set of reservoir properties 16 include, without limitation, permeability, fluid saturation, lithology, capillary properties, interfacial tension, and combinations thereof. Values may be absolute or a relative scale. Forexample, the reservoir permeability may simply be a designation of high, medium, or low permeability.

[0042] The dynamic simulator 20 provides labels for flow-simulation solution variables selected from reservoir pressure and one or both of reservoir fluid saturation and reservoir fluid mass accumulation. The dynamic simulator 20 may also provide labels for other flow-simulation solution variables, including, for example, without limitation, reservoir density and / or reservoir composition prior to fluid injection and / or production and labels after a period of injection and / or production. For example, without limitation, the labels may provide flow-simulation solution variables after 10 years of injection and / or production. The labels may be absolute or a relative scale ranging from high to low. A suitable dynamic simulator is for example, without limitation, PETROSIGNS® Flow (Shell). It will be understood by those skilled in the art that the value of one or more flow- simulation solution variable may be different across the fault block.Accordingly, the labels are preferably provided for each or most of the plurality of voxels in a fault block. In certain embodiments, the labels for one or more flow-simulation solution variable is an average value for the voxels in one or more fault blocks of the geologically realistic model 12.

[0043] A simplified example of input and output to the dynamic simulator 20 is illustrated in Fig. 3. The three geologically realistic models 12 depicted in Fig. 3 are structurally the same. The models 12, however, have different fault transmissibility multipliers (FTM). FTM’s are a function of reservoir permeability on each side of a fault, the thickness of the fault, and permeability of the fault. The FTM’s were inputted as high, medium, and low FTM.

[0044] The labelled geologically realistic models 22', 22", 22"' generated by the dynamic simulator 20 show an example of labels generated for reservoir pressure after 10 years of production. Labelled model 22' is the result for high-FTM, labelled model 22" is the result for mid-FTM, and labelled model 22'" is the result for low-FTM. The labelled geologically realistic models 22', 22", 22'" show the predicted effect on reservoir pressures in each fault block 38, by labelling as low, medium-low, medium, medium-high, and high reservoir pressure.

[0045] Preferably, the labelled geologically realistic models 22 provide flow-simulation solution variables as a function of time. With fluid injection and / or production, flow-simulation solution variables are expected to change as a function of time. In this way, the method of the present invention allows a user to efficiently predict flow-simulation solution variables fordifferent scenarios and to resolve uncertainties in, for example, without limitation, rock properties and / or fluid properties.

[0046] Fig. 4 illustrates an example of permutations of training properties that could be inputted to the dynamic simulator 20. As shown by Fig. 4, one structural model could have 27 permutations of high, medium, and low representations of FTM, fault permeability, and reservoir permeability. This shows that a single synthetically generated structural model could result in 27 models by variations in the associated training fluid properties 14 and training reservoir properties 16. This improves the efficiency of training the backpropagati on-enabled process by effectively multiplying the number of geologically realistic models by 27, in this case. Those skilled in the art would understand that even greater efficiencies could be attained by more variations in the properties for each structural model.

[0047] The labelled geologically realistic models 22 are used to train a backpropagati on- enabled process 30 to predict flow-simulation solution variables selected from reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation. Other flowsimulation solution variables such as, for example, without limitation, reservoir density, and reservoir composition may be predicted. Preferably, the flow-simulation solution variables are predicted as a function of time.

[0048] Once trained, the backpropagation-enabled process is then used on a non-training model to predict flow-simulation solution variables selected from reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation. The trained backpropagation-enabled process may also be used to predict other flow-simulation solution variables such as, without limitation, reservoir density and reservoir composition. Preferably, the flow-simulation solution variables are predicted as a function of time.

[0049] The trained backpropagation-enabled process 30 can be used to develop a number of scenarios for de-risking an injection and / or production operations. For example, the user may specify an initial reservoir pressure, a reservoir temperature, a reservoir thickness, an injection rate, a production rate, perforation thickness, water saturation, location of an injection and / or production well, number of injection and / or production wells, rock properties, fluid properties, and combinations thereof. For example, there may be some uncertainty in one or both of rock properties and fluid properties. The user may efficiently test effect of different options for the rock property and / or fluid property uncertainties.

[0050] By training the backpropagation-enabled process according to the method of the present invention, a multitude of factors can be processed accurately and timely to produce a prediction of reservoir pressure and one or both of reservoir fluid saturation and reservoir fluid mass accumulation, preferably as a function of time.

[0051] As a non-limiting example, a user using the trained backpropagation-enabled process 30 may specify a contact boundary in a non-training model. The user may then specify an existing or proposed well in the non-training model. Fault blocks and other geometric compartment boundaries, for example a sub-seismic tip extension compartment may be identified. Permeability for non-training model may be derived from field-acquired data or provided for different scenarios.

[0052] A prediction of reservoir pressure and one or both of reservoir fluid saturation and reservoir fluid mass accumulation for the non-training model is generated by the backpropagation-enabled process 30. The prediction may be provided as alternate scenarios for one or more specified parameters. The parameters may be selected from an initial reservoir pressure, a reservoir temperature, a reservoir thickness, an injection rate, a production rate, perforation thickness, water saturation, location of an injection and / or production well, number of injection and / or production wells, rock properties, fluid properties, and combinations thereof.

[0053] This allows a user to predict fluid flow fields in a subsurface reservoir or aquifer for different scenarios of injection and / or production well locations, injection and / or production rates, rock properties, and / or fluid properties as a function of time.

[0054] While preferred embodiments of the present invention have been described, it should be understood that various changes, adaptations, and modifications can be made therein within the scope of the invention(s) as claimed below.

Claims

What is claimed is:

1. A method for predicting fluid flow fields in a subsurface reservoir or aquifer, comprising the steps of: providing a training set of geologically realistic models, wherein each of the geologically realistic models have one or more faults defining one or more fault blocks; providing a training set of fluid properties; providing a training set of reservoir properties; inputting the training set of geologically realistic models, the training set of fluid properties, and the training set of reservoir properties to a dynamic simulator to generate a set of labelled geologically realistic models with labels relating to reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation for each of the one or more fault blocks; training a backpropagation-enabled process using the labelled geologically realistic models to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation, thereby producing a trained backpropagation- enabled process; inputting a non-training model to the trained backpropagation-enabled process to predict reservoir pressure and one or more of reservoir fluid saturation and reservoir fluid mass accumulation.

2. The method of claim 1, wherein the one or more faults in the set of geologically realistic models have fault properties selected from the group consisting of displacement, throw, heave, location, orientation, cross-fault juxtaposition, fault core width, fault permeability, and combinations thereof.

3. The method of claim 1, wherein the prediction of one or more reservoir saturation, reservoir pressure, reservoir mass composition, reservoir density, and reservoir composition as a function of time includes effect of one or more injection wells, one or more production wells, predicted changes over time, and combinations thereof.

4. The method of claim 1, wherein the training data set comprises field-acquired data, augmented data, synthetic data, and combinations thereof.

5. The method of claim 1, wherein the labels generated in the dynamic simulator are provided for each voxel of the fault block.

6. The method of claim 1, wherein the labels generated in the dynamic simulator are averaged for voxels in the fault block.

7. The method of claim 1, wherein the geologically realistic models are generated from synthetic data.

8. The method of claim 1, wherein the training set of fluid properties is selected from the group consisting of type of fluid, composition, density, viscosity, gas / oil ratio, buoyancy pressure, and combinations thereof.

9. The method of claim 1, wherein the training set of reservoir properties is selected from the group consisting of permeability, fluid saturation, lithology, capillary properties, interfacial tension, and combinations thereof.

10. The method of claim 1, wherein the training set of geologically realistic models and / or the training set of fluid properties further comprise a geologically realistic parameter selected from the group consisting of scaling relationships, burial depth, stress state, reactivation history, temperature, diagenesis, and combinations thereof.

11. The method of claim 1, wherein the prediction or one or more of reservoirs saturation and reservoir pressure for the non-training model is provided for a parameter selected from the group consisting of an initial reservoir pressure, a reservoir temperature, a reservoir thickness, an injection rate, a production rate, perforation thickness, water saturation, location of an injection and / or production well, number of injection and / or production wells, rock properties, fluid properties, and combinations thereof.

Images