Estimating a cleanup curve for a downhole fluid sampling operation

Abstract

A method for predicting a cleanup curve for a downhole sampling operation includes providing a set of wellbore parameters and fluid parameters to a machine learning (ML) proxy model. A processor generates a cleanup curve for a downhole sampling operation corresponding to the provided set of wellbore parameters and fluid parameters using the ML proxy model. The ML proxy model is specifically configured to output a cleanup curve for a downhole sampling operation in which the cleanup curve represents a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application Ser. No. 63 / 764,055, entitled Estimating Cleanup of Downhole Fluid Sampling Contamination, filed Feb. 27, 2025, which is incorporated herein by reference in its entirety.BACKGROUND

[0002] Subterranean wells are commonly drilled to explore and recover natural hydrocarbon deposits located in the Earth's crust. During a drilling operation, formation fluids are commonly sampled and evaluated to obtain information about a reservoir's fluid composition and gas-oil ratio (GOR) as well as other properties of the fluid. This information may be used for field planning decisions and for the optimization of upstream and downstream production facilities.

[0003] Fluid samples are commonly obtained using Logging While Drilling (LWD) tools to evaluate the formation fluids, determine reservoir quality, and make real-time decisions for production or further exploration. One difficulty with LWD acquisition is that after several hours of drilling, the wellbore is typically surrounded by an “invasion zone” where drilling fluid (mud) filtrate permeates the formation. As a result, when the LWD tool initially extracts fluid from the formation, the sample may be contaminated with mud filtrate (rather than being pure formation fluid). The amount of contamination tends to be especially high at the beginning of the extraction (sampling) process. Sampling methods make use of Downhole Fluid Analysis (DFA) that monitors the contamination level during the extraction process until clean formation fluid is sampled. For example, as the contamination level decreases below a set threshold (such as 5% or 10%), the cleaner formation fluid is collected in the sample chambers for analysis, with the intent being to provide a more representative reservoir sample.

[0004] During a sampling while drilling operation, accurately predicting the contamination cleanup time is essential for efficient job planning and decision-making. The time it takes for the contamination level to drop below the acceptable threshold depends on various factors, including formation and fluid properties, sampling tool and wellbore geometries, and operational constraints like flow rate and pressure. Analytical models can be used to provide fast and accurate predictions of the contamination cleanup process for vertical wells. However, no such analytical models are available for highly deviated wells or horizontal wells. Instead, numerical simulations are performed on three dimensional (3D) high-resolution grids. While such simulations are generally reliable, they also tend to be computationally intensive and time-consuming, making them impractical for tool planners to quickly evaluate multiple scenarios to understand parameter sensitivities and assess the impact of uncertain reservoir and fluid properties on fluid sampling time.

[0005] There is a need in the industry for a high-fidelity proxy model for contamination cleanup that covers all relevant sampling conditions and provides a promising solution with fast and accurate contamination cleanup predictions for deviated and horizontal wells.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] For a more complete understanding of the disclosed subject matter, and advantages thereof, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, in which:

[0007] FIG. 1 depicts an example drilling rig including a disclosed system for evaluating contamination cleanup during a downhole fluid sampling operation.

[0008] FIG. 2 schematically depicts an example downhole fluid sampling and evaluation measurement tool.

[0009] FIG. 3 depicts a block diagram of one example system for predicting a cleanup time during a downhole fluid sampling operation.

[0010] FIG. 4 depicts a flow chart of one example method for predicting a cleanup time during a downhole fluid sampling operation.

[0011] FIGS. 5A and 5B (collectively FIG. 5) depict an example computational domain for model simulations used to train the proxy model in FIGS. 3 and 4.

[0012] FIGS. 6A, 6B, and 6C (collectively FIG. 6) depict an example computational mesh employed in the model simulations used to train the proxy model in FIGS. 3 and 4.

[0013] FIG. 7 depicts a flow chart of one example method for training a machine learning model for predicting a cleanup time during a downhole fluid sampling operation.DETAILED DESCRIPTION

[0014] Methods and systems for predicting a cleanup curve for a downhole sampling operation are disclosed. One example method includes providing a set of wellbore parameters and fluid parameters to a machine learning (ML) proxy model. The ML proxy model is specifically configured to output a cleanup curve for a downhole sampling operation in which the cleanup curve represents a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume. A processor generates a cleanup curve for a downhole sampling operation corresponding to the provided set of wellbore parameters and fluid parameters using the ML proxy model.

[0015] In advantageous embodiments, the ML proxy model is pre-trained using a synthetic training dataset generated by executing a first-principles numerical model across a multi-dimensional parameter space of the wellbore parameters and wellbore fluid parameters to compute corresponding simulated cleanup curves such that the ML proxy model approximates the first-principles model at a reduced computational latency suitable for real-time deployment in a wellbore fluid sampling operation. Use of the proxy model to generate the cleanup curve may advantageously reduce the computational latency by at least a factor of 1000 as compared to execution of the first-principles numerical model.

[0016] In certain embodiments, the disclosed method may further include determining, based on the generated cleanup curve, a minimum fluid threshold including at least one of a minimum pumping time or a minimum fluid volume required to reach a target drilling fluid filtrate contamination level. Fluid samples may then be acquired during a downhole logging operation (e.g., a LWD fluid sampling operation) at a pumping time or pumping volume that exceeds the determined minimum pumping time or minimum pumping volume.

[0017] As described in more detail below, the disclosed embodiments may advantageously greatly reduce the time required to generate wellbore sampling cleanup curves such that the cleanup curves may be generated in real time during a sampling operation. The disclosed embodiments may further enable a large number of cleanup curves to be generated corresponding to a range of wellbore parameters and fluid parameters. The disclosed embodiments may therefore enable well planners to predict cleanup time in advance optimize multi-station sampling programs and minimize operational delays.

[0018] FIG. 1 depicts a schematic drilling rig 20 including a drill string 30 and a bottom hole assembly (BHA) including a drill bit 32 deployed in the string disposed within a wellbore 40. The drilling rig 20 may be deployed in either onshore or offshore applications (an onshore application is depicted). Moreover, the wellbore may include a deviated section, such as a building section, a highly inclined section, or a horizontal or near-horizontal section as depicted, however, the disclosed embodiments are not limited to any particular wellbore configuration. In the depicted example, the wellbore 40 may be formed in subsurface formations by rotary drilling in a manner that is well-known to those or ordinary skill in the art (e.g., via well-known directional drilling techniques).

[0019] As is known to those of ordinary skill, the drill string 30 may be rotated to drill the well (e.g., via a rotary table or via a hydraulically powered motor deployed in or above the BHA). A pump may deliver drilling fluid through the interior of the drill string 30 to the drill bit 32 where it exits the string via ports therein. The fluid may then circulate upwardly through the annular region 42 between the outside of the drill string 30 and the wall of the wellbore 40. In this known manner, the drilling fluid lubricates the drill bit 32 and carries formation cuttings up to the surface.

[0020] In the illustrated example embodiment, the BHA may include any number of downhole tools, for example, including a steering tool 34 and a measurement while drilling (MWD) tool 38. As depicted the BHA further includes an LWD fluid sampling and evaluation measurement tool 50. As described in more detail below, fluid sampling tool 50 may be configured to obtain a formation fluid sample and to analyze the sample to estimate a composition of the formation fluid. The BHA may further optionally include other LWD tools, one or more stabilizers, as well as other tools such as a reamer. The disclosed embodiments are not limited to any particular BHA configuration.

[0021] As further depicted, the rig 20 may include a system 100 configured to estimate a contamination cleanup curve or a cleanup time for a downhole fluid sampling operation. The system 100 may be located, for example, in a surface laboratory 80 (such as a laboratory trailer) at the rig site or offsite (e.g., and may be available via internet or other network connectivity). The disclosed embodiments are not limited in this regard. The system 100 may include computer hardware and software configured to store and employ a trained proxy model such as a trained machine learning (ML) model. To perform these functions, the hardware may include one or more processors (e.g., microprocessors) which may be connected to one or more data storage devices (e.g., hard drives or solid-state memory). As is known to those of ordinary skill, the processors may be further connected to a network, e.g., to receive the images from a networked camera system (not shown) or another compute system. It will, of course, be understood that the disclosed embodiments are not limited by the use of or the configuration of any particular computer hardware and / or software.

[0022] Turning now to FIG. 2, one example embodiment of a LWD fluid sampling tool 50 is depicted in a horizontal wellbore section (e.g., as shown in FIG. 1). It will be appreciated that the disclosed embodiments are not limited to any particular fluid sampling tool configuration. Measurement tool 50 may include a downhole tool body 52 such as an LWD tool body configured for deployment in (and coupling with) a BHA in a drill string. For example, the tool body may include threaded ends (not shown) for coupling with the drill string and may be configured to withstand the harsh drilling environment including severe shocks and vibrations. The measurement tool 50 may further include a probe 55 configured to sealingly engage a wellbore wall and to pump or draw wellbore fluid into the tool via an input port 56. The input port 56 is in fluid communication with an internal flowline 60 and at least one measurement sensor 65 (such as an optical sensor that is configured to make optical absorption measurements of the sampled wellbore fluid). A controller 70 may be configured to operate the sampling tool 50 as well as evaluate and interpret sensor measurements made by the measurement sensor 65. While the example embodiment depicted on FIG. 2 does not depict a sampling pump, it will be appreciated that a sampling pump may be deployed along flowline 60 or above (upstream of the) measurement sensor 65 such that it draws the fluid through the sensor 65. The disclosed embodiments are, of course, not limited in this regard.

[0023] As described above, fluid samples are commonly acquired downhole, for example, using a LWD tool, for evaluating formation fluids, determining reservoir quality, and making real-time decisions for production or further exploration. In drilling operations, the pressure of the drilling mud in the borehole is typically larger than the formation pressure, giving rise to mud-filtrate invasion. After several hours of drilling, the wellbore is typically surrounded by an invasion zone where drilling mud filtrate permeates the formation. As a result, when the fluid sampling tool initially extracts fluid from the formation, the sample is often contaminated with this mud filtrate rather than being pure formation fluid. The degree of contamination is especially high at the beginning of the extraction process. Once the contamination level falls below a predefined threshold (e.g., 5%), cleaner formation fluid is diverted into the sample chambers for subsequent analysis, yielding a more representative reservoir sample.

[0024] In fluid sampling operations, such as sampling while drilling operations, accurate prediction of contamination cleanup time is critical for efficient job planning and operational decision-making. Cleanup time is governed by a complex interplay of formation and fluid properties, sampling-tool and wellbore geometries, and operational constraints such as flow rate and pressure drawdown. Analytical models can be used to provide fast and accurate predictions of the contamination cleanup process for vertical wells. However, no such analytical models are available for highly deviated wells or horizontal wells. Instead, as described above in the Background Section, numerical simulations are commonly performed on three dimensional (3D) high-resolution grids. While such simulations are generally reliable, they also tend to be computationally intensive and time-consuming (e.g., requiring up to several hours to perform a single simulation). Such numerical simulations therefore tend to be impractical for tool planners to quickly evaluate multiple scenarios to understand parameter sensitivities and assess the impact of uncertain reservoir and fluid properties on fluid sampling time. They are also entirely unsuitable for real-time simulation of cleanup curves during a sampling operation.

[0025] FIG. 3 depicts a block diagram of a system 100 for generating predicted formation fluid cleanup curves. System 100 includes a parameter input block 110 configured to receive fluid sampling input parameters for modeling. The parameter input block 110 may include, for example, a user interface that enables (or prompts) a user to input parameters values. Example input parameters may include, for example, various wellbore and sampling tool parameters such as the wellbore diameter, the wellbore inclination angle, the formation thickness, a distance between the sampling tool and the top boundary of the formation, and a probe orientation angle (e.g., a toolface angle). The input parameters may also include various fluid and formation parameters including, for example, a depth of filtrate invasion into the formation, a permeability anisotropy, a ratio of the mud filtrate viscosity to the formation fluid viscosity (or the individual viscosities of the drilling fluid, the mud filtrate, and the formation fluid), and a mud filtration rate (e.g., a rate at which the filtrate enters the formation). The input parameter values may be input into a trained proxy model 120 (e.g., a trained ML model). Example proxy models and model training are described in more detail below. The trained model 120 may be configured to receive the input parameters from the input block 110 and provide a model output 130 that may include, for example, at least a predicted contamination cleanup curve. Such a curve may include, for example, a plot or listing of mud filtrate contamination values or percents with respect to pumping / sampling time or fluid volume as well as an estimated cleanup time for a desired cleanup level or threshold. In example embodiments, the predicted contamination cleanup curve may predict a sampling time required to achieve the desired cleanup level. In other example embodiments, the predicted contamination cleanup curve may predict a pumping volume required to achieve the desired cleanup level. The system 100 may further include a display interface 140 configured to display the predicted cleanup curve (the model output 130). The display interface 140 may include, for example, a computer display and may optionally further display the selected model parameter values and may enable the parameter values to be selected and / or adjusted (e.g., using a pointer such as a mouse).

[0026] FIG. 4 depicts a flow chart of one example method 150 for estimating a contamination cleanup curve or a cleanup time / volume for a downhole fluid sampling operation. The method includes training a proxy model at 155 to obtain a trained ML model (such as trained model 120 in FIG. 4). As described in more detail below by way of example, the model may be trained using numerically simulated contamination cleanup curves obtained using a single-phase, two-component numerical (mathematical) model at various selected input parameter conditions. In other words, the training may make use of numerous sets of input parameters and corresponding simulated cleanup curves. In the example described below, a design of experiment (DOE) was conducted to select the input parameters (e.g., optimal input parameters) to maximize parameter space coverage using a minimum number of simulations (to reduce the number of sets of input parameters required to obtain a desirable coverage of the parameter space). The trained machine learning model may include substantially any suitable ML model, for example, including Random Forest, ExtraTrees, K-Nearest Neighbors, Long Short Term Memory (LSTM) networks, and Transformer models. The ExtraTrees model was found to outperform other non-sequential models. In advantageous embodiments, the trained machine learning model is a temporal deep learning model specifically configured for sequence-based cleanup curve modeling. The LSTM, TCN, and transformer models advantageously excelled in capturing sequence-based outputs, significantly boosting predictive accuracy. The LSTM achieved a mean absolute percentage error (MAPE) of about 0.7 percent in the example implementation described in more detail below while a transformer model achieved a MAPE of about 1.1 percent and a TCN achieved a MAPE of about 1.6 percent. Other traditional models had corresponding errors of over 4.5 percent.

[0027] A set of fluid sampling input parameters (e.g., wellbore and fluid property parameters) is received at 160. The set of parameters may include, for example, those parameters described above with respect to FIG. 3. In particularly advantageous embodiments, the parameters may include at least the filtrate invasion depth, the permeability anisotropy, the mud filtration rate, and the viscosity ratio. The mud filtration rate was found to be an important parameter particularly for LWD sampling operations. The wellbore inclination has also been observed to be an important parameter such that in advantageous embodiments the set of input parameters may further include wellbore inclination. The received sampling parameters may be evaluated by the trained ML model at 162 to estimate and output a corresponding contamination cleanup curve at 164. The method may further optionally include repeating 160, 162, and 164 at 166 for one or more additional sets of sampling input parameters to estimate corresponding contamination cleanup curves.

[0028] It has been found that the disclosed proxy model dramatically reduces the evaluation time for contamination cleanup simulations and thereby provides reliable results for highly deviated wells where analytical solutions are not available. By using such a high-fidelity proxy model (trained proxy model 120), cleanup curve predictions tend to be nearly instantaneous (e.g., with in a fraction of a second using a desktop or laptop PC), enabling rapid assessment of multiple sampling operation scenarios and significantly improving and expediting job planning. By way of comparison, a single numerical simulation (generating a single cleanup curve) may require several hours of desktop or laptop processing time. The trained ML proxy model may therefore reduce computational latency for generating a cleanup curve by at least a factor of 1000 (e.g., at least a factor of 10,000).

[0029] As noted above with respect to FIGS. 3 and 4, the proxy model may be trained using numerically simulated contamination cleanup curves (a synthetic training data set). This training data set may be generated by executing a first-principles differential equation model across a multi-dimensional parameter space of simulated wellbore and wellbore fluid conditions. The trained ML proxy model may therefore be specifically configured to approximate the first-principles differential equation model with significantly reduced computational resources and may therefore provide cleanup (or contamination) curves at a correspondingly significantly reduced computational latency such that the ML proxy model may be advantageously utilized in real-time wellbore fluid sampling (e.g., LWD fluid sampling) operations to estimate pumping times and / or pumping fluid volumes required to obtain fluid sample having a desired purity (or a contamination level below a desired threshold). Moreover, the trained ML may advantageously enable cleanup curves to be rapidly and interactively generated for a large number of hypothetical input parameters, thereby enabling well planners to evaluate the impact well design parameters on subsequent fluid sampling operations. Moreover, it will be appreciated that certain wellbore parameters and / or fluid parameters are not known with certainty. Rapid generation of the simulated cleanup curves may advantageously enable well planners and / or field personnel to conduct a cleanup curve uncertainty analysis in substantially real time. For example, parameters of interest (such as the depth of filtrate invasion, the permeability anisotropy, and mud filtration rate) may be varied over predetermined ranges (e.g., related to the uncertainty of the parameter estimation) to determine a corresponding uncertainty in the predicted cleanup curve.

[0030] It will be appreciated that under miscible fluid conditions, a single-phase, two-component differential equation based fluid model may be employed that assumes complete miscibility between the mud filtrate and the reservoir fluid. Example model differential equations may include the single-phase continuity equation, momentum conservation with Darcy's Law, and a contamination transport equation, for example, as expressed mathematically below:∂ϕ⁢ρ∂t+∇·ϕρ⁢u=0ϕρ⁢ (∂u∂t+u·∇u)=-∇p+μ⁢∇2u-μK⁢u∂ϕ⁢ρ⁢ω∂t+∇·ωϕρ⁢u=0where φ represents formation porosity, ρ represents the fluid density, u represents a velocity vector, p represents the pressure, μ represents the dynamic viscosity, K represents the permeability of the porous medium, and w represents the contaminant mass fraction. Gravity is not considered in the foregoing equations (although the disclosed embodiments are not limited in this regard). The mixture density may be given as follows:ρ=(ωρmf+1-ωρo)-1where ρmf and ρo represent the densities of the pure filtrate and the pure formation fluid. A linear mixing rule may be used for the effective viscosity, μ, for example, as follows:μ=μo(1-ω)+μmf*ωThe contaminant volume fraction may be determined as follows:v=ωρmfωρmf+1-ωρoA geometry model pertaining to a deviated wellbore intersecting a homogeneous and anisotropic formation with a thickness H may be employed. For example, the geometry model may assume that the LWD sampling tool is positioned at a distance h (0≤h≤H) from the top of the formation.FIGS. 5A and 5B (collectively FIG. 5) depict an example computational domain for model simulations used to train the proxy model in FIGS. 3 and 4. In this example, the computational domain includes a tool height h=0.5H, the wellbore inclination is 30 degrees, and the probe orientation is 0 degrees. In this example, owing to the inherent symmetry, it is only necessary to consider half of the total cylindrical domain. However, in instances such as a 90 degree probe orientation, where no symmetry plane exists, a full cylindrical domain may be utilized. In this example the domain was selected to extend 20 meters radially around the probe to ensure (or increase the likelihood) that boundary effects from the far-field are negligible. The lateral boundary may be treated as a pressure outlet. The boundary condition at the probe may be set as a negative velocity inlet, with the inlet velocity derived from a specified flow rate. The wellbore boundary surface may be set as a velocity inlet based on the mud filtrate rate. Other boundary conditions may be defined as symmetry.

[0036] FIGS. 6A, 6B, and 6C (collectively FIG. 6) depict an example computational mesh employed in the model simulations used to train the proxy model in FIGS. 3 and 4. A mesh sensitivity study was conducted to achieve a balance between computational efficiency and numerical accuracy. A very fine mesh was first established as the reference case, using a probe control size of 0.00075 m, a cell growth rate of 1.1, and a curvature normal angle of 4.5°. For the horizontal well case, this configuration resulted in approximately 8.7 million cells. To reduce computational cost while maintaining acceptable predictive accuracy for contamination cleanup, progressively coarser mesh configurations were evaluated. The probe control size was systematically increased to decrease the total cell count, and the corresponding impact on contamination cleanup predictions was assessed. Through this parametric mesh refinement study, the total number of cells was reduced to approximately 0.9 million (see FIG. 6), while maintaining good agreement with the very fine base mesh results.

[0037] In the example implantation described herein, the computational mesh applies a sampling probe control size of 0.003 meters. The sampling probe was modeled as a circle with a radius of 0.010414 meters. Two circular mesh refinement zones were defined around the probe, with radii of 0.05 meters and 0.08 meters, respectively. The mesh size increased progressively from the probe with a growth rate of 1.2 (to reduce computational overhead, the growth rate is at least 1.15). A curvature normal angle of 18° was applied to ensure adequate resolution of the cylindrical wellbore geometry (again to reduce computational overhead, the curvature normal angle is at least 12 degrees). For invasion depths equal or greater than 4 in., a localized refinement zone (mesh size=0.012 m) was defined around the entire wellbore. For invasion depths less than 4 in., refinement was applied around the entire wellbore to improve resolution of the near-wellbore flow field.

[0038] A wide range of simulation cases were set up and run using the commercial software ANSYS Fluent (www.ansys.com). The contamination transport equation was implemented by defining an additional user-defined scalar. The depth of filtrate invasion was included in the initial condition with compiled user-defined functions. The density mixing rule and the viscosity mixing rule were also defined with user-defined functions in ANSYS Fluent. The development of the proxy model required a large number of simulations covering various parameter combinations. To streamline this process, an automated workflow was created using ANSYS. Automated geometry control was managed through macros in ANSYS Workbench, while mesh generation and simulation setup were handled with PyFluent. This fully automated workflow enabled efficient case generation, simulation execution, and results storage using Python, significantly speeding up the model development process.

[0039] As noted above, ANSYS Fluent was used to conduct finite volume numerical simulations with a single-phase, two-component fluid first principles numerical model. This approach simulated miscible fluid behavior under the assumption of complete miscibility between mud filtrate and reservoir fluid. Key simulation parameters were varied to build a comprehensive database of fluid behavior. Such parameters included wellbore diameter, wellbore inclination, formation thickness, tool placement, probe orientation, filtration invasion depth, permeability anisotropy, viscosity ratio, and mud filtration rate. Machine learning models were then trained to predict contamination cleanup curves for non-simulated cases. The ML models included Random Forest, ExtraTrees, K-Nearest Neighbors, Long Short Term Memory (LSTM) networks, a temporal convolutional network (TCN), and Transformer models.

[0040] A Design of Experiment (DOE) was used to select the parameter sets for training the proxy model. Table 1 lists nine parameters that were identified as potentially individually influencing the cleanup behavior. It was determined that simulating all of the related cases listed in Table 1 would involve approximately 20 million scenarios (simulations), requiring excessive computational resources. The DOE may be configured to reduce the number of selected scenarios (sets of parameters) by at least a factor of 100 (e.g., by a factor of 200, 500, or even 1000). In the example implementation described herein, a first stage focused on 2,728 cases to assess the sensitivity of each of the parameters and the corresponding results at reduced computational demand.

[0041] In the first stage, 1728 simulation cases were selected by expert as a fractional factorial design. Then an additional 1000 cases were generated using a Latin Hypercube experimental design where the true model was evaluated. In this first stage, 300 of the cases were reserved for validation and the remaining 2428 were used to train the proxy model. The validation study assessed the performance of the proxy model and identified critical parameter zones requiring further simulation. The first-phase proxy model successfully demonstrated the feasibility of applying machine learning to accurately predict contamination cleanup curves. The results showed a strong link between cleanup time and parameters such as filtrate invasion depth, permeability anisotropy, viscosity ratio, and mud invasion rate. ExtraTrees outperformed other non-sequential models, while LSTM and Transformer models excelled in capturing sequence-based (temporal) outputs, significantly boosting predictive accuracy.

[0042] Based on the first-stage outcomes, second-stage cases (parameter sets for simulation) may be designed. Further development in the second phase, with expanded training data, is expected to enhance the model's accuracy and enable it to address corner cases (in the input parameter set) and generalize effectively across a wider range of sampling conditions and input parameters. Moreover, future efforts may be intended to refine the model for robust performance under diverse reservoir and operational scenarios. If necessary, a third stage may follow depending on the results of the second stage. The objective is to minimize the number of simulations while achieving the desired model performance (accuracy and precision of the predicted contamination curves).TABLE 1Parameters used for Proxy Model TrainingParametersymbolUnitRangeGeometrywellboreDwIn8.5, 10.25diameterwellDeg0 (vertical), 30, 45, 60, 75, 90, 95inclinationangleformationHM1, 1.5, 2, 2.5, 3, 5, 10, 15, 20, 30, 50thicknesstool distancez = h / H0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5from the topboundaryprobeα0, 45, 90, 135, 180orientationOtherdepth ofDOIInch0, 0.5, 1, 2, 4, 8, 12, 20, 40filtrateinvasionpermeabilityKv / Kh0.01, 0.1, 0.2, 0.5, 1, 1.5anisotropyformationμmf / μ00.002, 0.1, 0.25, 0.5, 1, 2.5, 5, 10,viscosity / mud100, 200filtrateviscositymud filtrationcm3 / s / cm21E−3, 1E−4, 1E−5, 1E−6, 0rate

[0043] With reference again to FIGS. 3 and 4, a variety of methods, incorporating advanced Machine Learning techniques, may be employed to develop the trained proxy model to ensure promote an efficient and accurate approximation of the high-fidelity model's behavior. Such methods may include Kriging (also known as Gaussian Process Regression), which is particularly suited for spatial modeling; Radial Basis Function (RBF) Interpolation, ideal for interpolating multi-dimensional data; and Artificial Neural Networks (ANNs), which are widely used for capturing complex nonlinear relationships within data. Additionally, Polynomial Regression may be applied for simpler, lower-dimensional approximations, while ensemble learning techniques such as Random Forests and Gradient Boosting may be leveraged to enhance predictive accuracy by combining multiple decision trees. Moreover, reduced-order modeling may be employed to offer a simplified representation of the high-fidelity model that retains essential characteristics while significantly reducing computational complexity. Each of these methods may be rigorously evaluated against a set of criteria, primarily focusing on their precision in replicating the behavior of the original model (from which the simulations were obtained), computational efficiency, and scalability. One objective is to identify the most appropriate approach that balances accuracy and computational cost, ensuring the proxy model can reliably substitute the high-fidelity model in large-scale simulations and real-time applications.

[0044] The disclosed embodiments advantageously introduces a robust ML-based proxy model designed for contamination cleanup simulation at any wellbore inclination (inclination independent). Advanced ML models specialized in sequence processing such as LSTMs and Transformers have been found to significantly enhance the prediction of sequence-based cleanup curves. Moreover, the disclosed embodiments have been found to provide a precise and reliable solution for complex fluid simulations under challenging well conditions.

[0045] FIG. 7 depicts a flow chart of one example method 200 for training a machine learning model for predicting a cleanup time during a downhole fluid sampling operation. Sets of input wellbore parameters and wellbore fluid parameters for simulation may be selected at 205. For example, as described above a DOE may be employed to select the input parameters (e.g., optimal input parameters) to maximize parameter space coverage using a minimum number of simulations. A training data set may be generated at 210, for example, using a single-phase, two-component fluid first principles numerical model. The model may be a differential equations model, for example, as described above. The training data set may include a large number (e.g., several thousand) of the selected sets of input wellbore parameters and wellbore fluid parameters and corresponding computed (predicted) cleanup curves (plots of mud filtrate contamination versus time or flow volume during a sampling operation.

[0046] The computed training data set may then be used to train the ML model 215. The training data set may be split into a training subset and a validation subset, as described above by way of example, and then used to train the ML model. The training may include identifying relationships and / or correlations between the wellbore parameters and wellbore fluid parameters and the corresponding cleanup curves. The training and validation may further include tuning model hyper parameters and optimizing to achieve the lowest mean absolute percentage error (MAPE). The training may make use of customized deep learning architectures suitable for regression and may further compare and contrast the predictive performance of many different artificial intelligence (AI) based regression methods, but particularly temporal models as described above. The trained model may then be deployed in the field or in an offsite computer system at 220.

[0047] It will be understood that the present disclosure includes numerous embodiments. These embodiments include, but are not limited to, the following embodiments.

[0048] In a first embodiment a method for generating a cleanup curve for use in a wellbore fluid sampling operation comprises providing a set of wellbore parameters and fluid parameters to a machine learning (ML) proxy model, wherein the ML proxy model is specifically configured to output a cleanup curve for a downhole sampling operation, the cleanup curve representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume; and generating a cleanup curve for a downhole sampling operation corresponding to the provided set of wellbore parameters and fluid parameters using the ML proxy model via a processor.

[0049] A second embodiment may include the first embodiment, wherein the ML proxy model is pre-trained using a synthetic training dataset generated by executing a first-principles numerical model across a multi-dimensional parameter space of the wellbore parameters and wellbore fluid parameters to compute corresponding simulated cleanup curves such that the ML proxy model approximates the first-principles model at a reduced computational latency suitable for real-time deployment in a wellbore fluid sampling operation.

[0050] A third embodiment may include the second embodiment, wherein the generating the cleanup curve using the ML proxy model reduces the computational latency by at least a factor of 1000.

[0051] A fourth embodiment may include any one of the first through third embodiments, further comprising determining, based on the generated cleanup curve, a minimum fluid threshold including at least one of a minimum pumping time or a minimum fluid volume required to reach a target drilling fluid filtrate contamination level.

[0052] A fifth embodiment may include the fourth embodiment, further comprising repeating the providing, the generating, and the determining to output a plurality of cleanup curves and minimum fluid thresholds for corresponding distinct sets of wellbore and fluid parameters.

[0053] A sixth embodiment may include the fifth embodiment, wherein the repeating comprises varying selected ones of the wellbore parameters and fluid parameters over predetermined ranges related to an uncertainty of the selected parameter to determine a corresponding uncertainty in the predicted cleanup curve.

[0054] A seventh embodiment may include any one of the fourth through sixth embodiments, further comprising causing a downhole sampling tool to sample wellbore fluid for a time duration or fluid volume that satisfies the minimum fluid threshold to obtain a wellbore fluid sample having a contamination level less than the target drilling fluid filtrate contamination level.

[0055] An eighth embodiment may include any one of the first through seventh embodiments, wherein the wellbore parameters comprise at least a non-vertical wellbore inclination.

[0056] A ninth embodiment may include any one of the first through eighth embodiments, wherein the fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

[0057] A tenth embodiment may include any one of the first through ninth embodiments, wherein the trained proxy model comprises a long short term memory network, a temporal neural network, or a transformer model.

[0058] In an eleventh embodiment a system for estimating a cleanup curve for a downhole sampling operation comprises an input module configured to receive a set of wellbore parameters and fluid parameters for a downhole sampling operation; a trained machine learning (ML) based proxy model specifically configured to output a cleanup curve for a downhole sampling operation corresponding to the input set of wellbore parameters and fluid parameters, receive the set of wellbore and fluid parameters and output a corresponding contamination cleanup curve, the cleanup curve representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume; and a processor configured to input the received set of wellbore parameters and fluid parameters into the trained ML model to generate the corresponding contamination cleanup curve.

[0059] A twelfth embodiment may include the eleventh embodiment, wherein the ML proxy model is pre-trained using a synthetic training dataset generated by executing a first-principles numerical model across a multi-dimensional parameter space of the wellbore parameters and wellbore fluid parameters to compute corresponding simulated cleanup curves such that the ML proxy model approximates the first-principles model at a reduced computational latency suitable for real-time deployment in a wellbore fluid sampling operation.

[0060] A thirteenth embodiment may include the twelfth embodiment, wherein the system reduces a computational latency of generating the cleanup curve cleanup curve by at least a factor of 1000 as compared using the first-principles differential equation model.

[0061] A fourteenth embodiment may include any one of the eleventh through thirteenth embodiments, wherein the trained proxy model comprises a long short term memory network, a temporal neural network, or a transformer model; the wellbore parameters comprise at least a non-vertical wellbore inclination; and the fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

[0062] A fifteenth embodiment may include any one of the eleventh through fourteenth embodiments, further comprising a display interface configured to display the generated cleanup curve.

[0063] In a sixteenth embodiment a method for training a machine learning (ML) model to generate a cleanup curve for a downhole sampling operation comprises selecting a plurality of sets of wellbore parameters and fluid parameters; generating a training data set including a plurality of simulated cleanup curves corresponding to the selected plurality of sets of wellbore parameters and fluid parameters, each of the plurality of simulated cleanup curves representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume, wherein the simulated cleanup curves are simulated by executing a first-principles numerical model; and training a ML model with the generated training data set to obtain a trained ML model.

[0064] A seventeenth embodiment may include the sixteenth embodiment, wherein the selecting comprises (i) selecting the wellbore parameters and fluid parameters for simulation, (ii) defining a parameter space including parameter values for each of the selected wellbore parameters and fluid parameters, and (iii) conducting a design of experiment to select the plurality of sets of wellbore parameters and fluid parameters from the defined parameter space.

[0065] An eighteenth embodiment may include any one of the sixteenth or seventeenth embodiments, wherein the numerical model comprises a finite volume differential equation model in which a downhole sampling probe is modeled as a circle with circular mesh refinement zones having a curvature normal angle of at least 12 degrees defined around the sampling probe and a mesh size increasing progressively from the probe at a growth rate of at least 1.15.

[0066] A nineteenth embodiment may include any one of the sixteenth through eighteenth embodiments, wherein the ML model comprises a long short term memory network, a temporal neural network, or a transformer model; the wellbore parameters comprise at least a non-vertical wellbore inclination; and the fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

[0067] A twentieth embodiment may include any one of the sixteenth through nineteenth embodiments, wherein the trained ML model is configured to generate the cleanup curve with a computational latency that is at least a factor of 1000 less than that of the first-principles numerical model.

[0068] Although estimating a cleanup curve for a downhole fluid sampling operation has been described in detail, it should be understood that various changes, substitutions and alternations can be made herein without departing from the spirit and scope of the disclosure as defined by the appended claims.

Claims

1. A computer-implemented method for generating a cleanup curve for use in a wellbore fluid sampling operation; the method comprising:providing a set of wellbore parameters and fluid parameters to a machine learning (ML) proxy model, wherein the ML proxy model is specifically configured to output a cleanup curve for a downhole sampling operation, the cleanup curve representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume; andgenerating, via a processor, the cleanup curve for the downhole sampling operation corresponding to the provided set of wellbore parameters and fluid parameters using the ML proxy model.

2. The method of claim 1, wherein the ML proxy model is pre-trained using a synthetic training dataset generated by executing a first-principles numerical model across a multi-dimensional parameter space of the wellbore parameters and wellbore fluid parameters to compute corresponding simulated cleanup curves such that the ML proxy model approximates the first-principles model at a reduced computational latency suitable for real-time deployment in a wellbore fluid sampling operation.

3. The method of claim 2, wherein the generating the cleanup curve using the ML proxy model reduces the computational latency by at least a factor of 1000.

4. The method of claim 1, further comprising determining, based on the generated cleanup curve, a minimum fluid threshold including at least one of a minimum pumping time or a minimum fluid volume required to reach a target drilling fluid filtrate contamination level.

5. The method of claim 4, further comprising repeating the providing, the generating, and the determining to output a plurality of cleanup curves and minimum fluid thresholds for corresponding distinct sets of wellbore and fluid parameters.

6. The method of claim 5, wherein the repeating comprises varying selected ones of the wellbore parameters and fluid parameters over predetermined ranges related to an uncertainty of the selected parameter to determine a corresponding uncertainty in the predicted cleanup curve.

7. The method of claim 4, further comprising causing a downhole sampling tool to sample wellbore fluid for a time duration or fluid volume that satisfies the minimum fluid threshold to obtain a wellbore fluid sample having a contamination level less than the target drilling fluid filtrate contamination level.

8. The method of claim 1, wherein the wellbore parameters comprise at least a non-vertical wellbore inclination.

9. The method of claim 1, wherein the fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

10. The method of claim 1, wherein the trained proxy model comprises a long short term memory network, a temporal neural network, or a transformer model.

11. A system for estimating a cleanup curve for a downhole sampling operation, the system comprising:an input module configured to receive a set of wellbore parameters and fluid parameters for a downhole sampling operation;a trained machine learning (ML) based proxy model specifically configured to output a cleanup curve for a downhole sampling operation corresponding to the input set of wellbore parameters and fluid parameters, receive the set of wellbore and fluid parameters and output a corresponding contamination cleanup curve, the cleanup curve representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume; anda processor configured to input the received set of wellbore parameters and fluid parameters into the trained ML model to generate the corresponding contamination cleanup curve.

12. The system of claim 11, wherein the ML proxy model is pre-trained using a synthetic training dataset generated by executing a first-principles numerical model across a multi-dimensional parameter space of the wellbore parameters and wellbore fluid parameters to compute corresponding simulated cleanup curves such that the ML proxy model approximates the first-principles model at a reduced computational latency suitable for real-time deployment in a wellbore fluid sampling operation.

13. The system of claim 12, wherein the system reduces a computational latency of generating the cleanup curve cleanup curve by at least a factor of 1000 as compared using the first-principles differential equation model.

14. The system of claim 11, wherein:the trained proxy model comprises a long short term memory network, a temporal neural network, or a transformer model;the wellbore parameters comprise at least a non-vertical wellbore inclination; andthe fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

15. The system of claim 11, further comprising a display interface configured to display the generated cleanup curve.

16. A method for training a machine learning (ML) model to generate a cleanup curve for a downhole sampling operation, the method comprising:selecting a plurality of sets of wellbore parameters and fluid parameters;generating a training data set including a plurality of simulated cleanup curves corresponding to the selected plurality of sets of wellbore parameters and fluid parameters, each of the plurality of simulated cleanup curves representing a concentration of drilling fluid filtrate contamination in sampled wellbore fluid versus time or versus pumped wellbore fluid volume, wherein the simulated cleanup curves are simulated by executing a first-principles numerical model; andtraining a ML model with the generated training data set to obtain a trained ML model.

17. The method of claim 16, wherein the selecting comprises (i) selecting the wellbore parameters and fluid parameters for simulation, (ii) defining a parameter space including parameter values for each of the selected wellbore parameters and fluid parameters, and (iii) conducting a design experiment to select the plurality of sets of wellbore parameters and fluid parameters from the defined parameter space.

18. The method of claim 16, wherein the numerical model comprises a finite volume differential equation model in which a downhole sampling probe is modeled as a circle with circular mesh refinement zones having a curvature normal angle of at least 12 degrees defined around the sampling probe and a mesh size increasing progressively from the probe at a growth rate of at least 1.15.

19. The method of claim 16, wherein:the ML model comprises a long short term memory network, a temporal neural network, or a transformer model;the wellbore parameters comprise at least a non-vertical wellbore inclination; andthe fluid parameters comprise at least a depth of filtrate invasion, a permeability anisotropy, a mud filtrate viscosity to formation fluid viscosity ratio, and a mud filtration rate.

20. The method of claim 16, wherein generating each of the plurality of simulated cleanup curves reduces a computational latency by at least a factor of 1000 less than that of the first-principles numerical model.

Images