Machine learning estimation of reservoir fluid properties

Abstract

A method for estimating reservoir fluid properties includes classifying the reservoir fluid as normal or abnormal based on measured gas composition and classified fluid types using a trained machine learning model, and when classifying the reservoir fluid as normal, using another trained machine learning model to estimate the properties of the reservoir fluid. When the reservoir fluid is classified as abnormal, a heavy hydrocarbon fraction of the reservoir fluid is predicted based on the measured composition and the classified fluid type using a further trained machine learning model, and when the reservoir fluid is classified as abnormal, the heavy hydrocarbon fraction of the reservoir fluid is predicted based on the measured composition and the classified fluid type using a further trained machine learning model.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of European Application No. 23306629.9, filed on September 28, 2023, the entire contents of which are incorporated herein by reference. Background Art

[0003] When drilling a well for the production of hydrocarbons, a drilling fluid is typically circulated through the well for a number of purposes. For example, the drilling fluid is typically intended to provide pressure to the underground formation, cool and lubricate the drill bit, flush cuttings from the drill bit and transport them to the surface, and provide hydraulic power to various downhole tools. The drilling fluid also typically transports formation fluids and dissolved formation gases to the surface. This gas can be released by the drill bit as it cuts through the formation.

[0004] The released gases are typically evaluated at the surface while drilling (e.g., via gas chromatography) to measure the amount of various alkane gases (e.g., methane (C1), ethane (C2), propane (C3), butane (C4), pentane (C5)) in the gas stream. Such measurements can provide valuable information to the mud logging tool and can provide information about the maturity and nature of the hydrocarbons in the reservoir, the demarcation of intervals in the reservoir being drilled and the quality of the oil, as well as information about production zones, lithologic changes, reservoir accumulation history, seal effectiveness, and the environmental impact of the drilling operation.

[0005] While such gas measurements can provide valuable insights into the contents of a reservoir, there is room for further improvement. For example, further evaluation of gas composition measurements is needed to infer (or estimate) additional reservoir fluid properties, such as fluid type, gas-to-oil ratio, and heavy hydrocarbon fractions (e.g., C6, C7, C8, etc.). 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 description taken in conjunction with the accompanying drawings, in which:

[0007] Figure 1 An example drilling rig including the disclosed system for taking mud logging measurements is depicted.

[0008] Figure 2 Another embodiment of a surface system for taking mud logging measurements is shown.

[0009] Figure 3 A flow chart of an example method for estimating reservoir fluid properties from gas composition measurements is depicted.

[0010] Figure 4A flow chart of another example method for estimating reservoir fluid properties from gas composition measurements is depicted. DETAILED DESCRIPTION

[0011] Embodiments of the present disclosure include methods and systems for estimating reservoir fluid properties from gas composition measurements. In an example embodiment, the disclosed method includes obtaining a measured composition of a gas sample obtained during a mud logging operation, the measured composition including a selected alkane gas; using a trained machine learning model to classify the reservoir fluid as normal or abnormal based on the measured composition and the classified fluid type; when the reservoir fluid is classified as normal based on the measured composition and the classified fluid type, using another trained machine learning to predict the heavy hydrocarbon fraction of the reservoir fluid; and when the reservoir fluid is classified as abnormal based on the measured composition and the classified fluid type, using another trained machine learning model to predict the heavy hydrocarbon fraction of the reservoir fluid. The method may optionally further include predicting the gas-oil ratio of the reservoir fluid from the measured composition, the classified fluid type, and the predicted heavy hydrocarbon fraction using another trained machine learning model.

[0012] The disclosed embodiments can advantageously enable accurate prediction of reservoir fluid properties, including reservoir fluid type, C6+ fraction in the gas-to-oil ratio of the fluid. In addition, the disclosed embodiments advantageously account for the presence of irregular or abnormal fluids, such as those subject to biodegradation, and can advantageously provide a robust and accurate prediction of the gas-to-oil ratio of a reservoir.

[0013] Figure 1 An example drilling rig 20 is depicted, which includes a system 80 for making mud logging measurements (e.g., making gas composition measurements and estimating reservoir fluid properties based on the measurements). The drilling rig 20 can be positioned above a subterranean formation or reservoir (not shown). The drilling rig 20 can include, for example, a derrick and a lifting device (also not shown) for raising and lowering a drill string 30, which, as shown, extends into a wellbore 40 and includes, for example, a drill bit 32 and one or more downhole measurement tools 38 (e.g., logging while drilling tools or measurement while drilling tools) in a bottom hole assembly (BHA) above the drill bit 32. Suitable drilling systems (e.g., including drilling, steering, logging and other downhole tools) are well known in the industry.

[0014] The drilling rig 20 also includes a surface system 50 for controlling the flow of drilling fluid used on the drilling rig (e.g., for drilling the wellbore 40). In the depicted example drilling rig, the drilling fluid 35 can be pumped downhole (as depicted at 92), for example, via a conventional mud pump 57. The drilling fluid 35 can be pumped to the drill string 30, for example, through a riser 58 and a mud hose 59 in the route. The drilling fluid 35 typically exits the drill string 30 at or near the drill bit 32 and creates an upward flow 94 of mud through the wellbore annulus 42 (the annular space between the drill string and the wellbore wall). The drilling fluid 35 then flows through the return conduit 52 to the mud pit system 56, where the drilling fluid 35 can be recirculated. It should be understood that the terms drilling fluid and mud are used synonymously herein.

[0015] The circulating drilling fluid 35 is intended to perform many functions during the drilling operation, one of which is to transport the drill cuttings 45 to the surface (in the upward flow 94). The drill cuttings 45 are typically removed from the return mud via a mud shale shaker 55 (or other similar solids control equipment) in a return conduit (e.g., immediately upstream of the mud pit 56). Gases released or generated during drilling can also be transported to the surface in the circulating drilling fluid. These gases, which may be dissolved in the mud or dissolved in the form of bubbles, are typically removed from the drilling fluid, for example, via one or more degassers 54 located in or near a header box 53, which is immediately upstream of the mud shale shaker 55 in the example depiction. The drill cuttings 45 and the extracted gas are typically examined at the surface to assist the drilling operation and evaluate the formations and reservoirs through which the wellbore passes.

[0016] As is known to those of ordinary skill in the art, formation gas can be released into the wellbore 40 via the drilling process (e.g., by crushing the formation rock through the mechanical action of the drill bit), and can also migrate into the wellbore 40, for example, via fractures in the formation rock. The drilling process can also produce gas, for example, via drill bit metamorphism (DBM). Once in the wellbore, the gas can be transported to the surface via the drilling fluid (in the upward flowing fluid 94). The gas can be dissolved (dissolved) in the drilling fluid and / or exist in the form of bubbles, and can be sampled in the surface system, for example, via one or more drilling fluid degassers 54 and / or headspace gas probes. The disclosed embodiments are not necessarily limited to how the gas is sampled.

[0017] Further references Figure 1, the drilling rig 20 may also include a testing facility 60 (e.g., a mud logging system or a laboratory trailer including one or more instruments suitable for performing various measurements on sampled gases in the drilling fluid). In the illustrated embodiment, the testing facility 60 has instruments (e.g., gas chromatography equipment) and is configured to measure the formation gas composition. Of course, the testing facility 60 may include many other testing instruments known to those of ordinary skill in the art. The facility may also include a system 80 configured to estimate reservoir fluid properties from gas composition measurements. The system 80 may include an instrument, or may be configured to receive gas composition measurements from an instrument (e.g., a gas chromatography equipment). The system 80 may also include computer hardware and software configured to estimate reservoir fluid properties from gas composition measurements. The hardware may include one or more processors (e.g., microprocessors) that may be connected to one or more data storage devices (e.g., hard drives or solid-state memory) and a user interface. The system 80 may also be configured to receive a trained machine learning model. It will be further understood that the disclosed embodiments may include processor executable instructions stored in a data storage device. The executable instructions may be configured, for example, to perform methods 100 and 150 to predict reservoir fluid properties from received gas measurements, as described in more detail below. The disclosed embodiments are of course not limited to the use or configuration of any particular computer hardware and / or software.

[0018] Of course it should be understood that Figure 1 A land drilling rig 20 is depicted, but the disclosed embodiments are equally applicable to land drilling rigs or offshore drilling rigs. As known to those of ordinary skill in the art, an offshore drilling rig typically includes a platform deployed on top of a riser that extends from the seafloor to the surface. A drill string extends downward from the platform, through the riser, and through a blowout preventer (BOP) located on the seafloor into the wellbore. The disclosed embodiments are expressly not limited in these respects.

[0019] Figure 2 Another view of a portion of a surface system 50 is depicted. As described above, the return conduit 52 is configured to carry the drilling fluid 35 (sometimes including gas bubbles 37) from the wellbore 40 to the mud pit 56. The depicted example system 50 includes at least one degasser 54, which is deployed, for example, in or near a header box 53, which is immediately upstream of the shale shaker 55 and the mud pit 56. In this exemplary configuration, the degasser 54 is configured to remove gas (referred to in the industry as gas-out) from the drilling fluid exiting the wellbore 40. A small volume of drilling fluid can be circulated through the degasser 54, where the drilling fluid is agitated and optionally heated to remove alkanes and other gases from the fluid volume.

[0020] It should be understood that the degasser 54 may include substantially any suitable type of degasser, including, for example, vacuum degassers, centrifugal degassers, and impeller degassers. The degasser may also be configured to heat the drilling fluid 35 to promote enhanced degassing of the fluid. The disclosed embodiments are not limited to the type of degasser employed.

[0021] like Figure 2 As further depicted in the example embodiment, the degasser 54 can be piped directly to a mud logging unit or a rig lab 60 (e.g., as depicted at 65), for example, to automatically deliver the sampled gas to a gas chromatograph for composition testing. In addition, although not shown, the system 50 can include one or more pumps (e.g., suction or booster pumps) configured to pump the sampled gas from the degasser 54 to the lab 60. Of course, the disclosed embodiments are not limited to any sampling, pumping, or gas delivery configuration.

[0022] As mentioned above, the mud logging system or laboratory may include a gas analyzer or measuring equipment, such as a gas chromatography (GC) measuring equipment. Common GC analyzers include a gas sample injection port, which is configured to feed a gas sample into a column assembly that includes at least a main GC column. The gas sample is usually mixed with a carrier gas such as nitrogen, argon, helium or air and transported by the column assembly. An optional pre-cut column can remove heavier hydrocarbon compounds, such as those compounds with a carbon number higher than a threshold value (such as C6, C8 or C10 and above). The main column usually includes a stationary phase and is intended to separate various gaseous compounds in the gas sample so that they arrive at a detector at different elution times, such as so that C1 arrives before C2, C2 arrives before C3, etc. The detector can include substantially any suitable GC detector, such as an FID detector, a TC detector or a mass spectrometer.

[0023] The gas measurement device can be configured to quantify the amount of light alkanes and optionally other gases (e.g., olefins and alcohols) in a gas sample. In suitable embodiments, at least the following light alkane gases can be quantified: methane (C1), ethane (C2), propane (C3), n-butane (nC4), isobutane (iC4), n-pentane (nC5), and isopentane (iC5). However, it should be understood that the disclosed embodiments are not limited in this regard, as in other embodiments, the gas measurement device can quantify other compounds as aromatic compounds (e.g., benzene and toluene), cyclic compounds (e.g., cyclopentane, methylcyclohexane), and inorganic substances (e.g., hydrogen, helium, carbon monoxide, and carbon dioxide).

[0024] In an exemplary embodiment, the gas measurement system (including the degasser in the measurement equipment) may include an advanced gas service, such as the FlairFlex service (available from SLB), which is configured to monitor the composition of the reservoir (particularly C1 to C5). The Flairflex service provides laboratory quality gas composition measurements of reservoir fluids during drilling and can be used for hydrocarbon and fluid contact identification and inter-well and intra-well fluid phase mapping. Gas composition measurements from Flairflex can guide optimal sampling and downhole fluid analysis, as well as early detection of reservoir complexity (tight, thin layers, etc.). Although the following embodiments are described with respect to receiving Flairflex gas composition measurements, it should be understood that these are merely examples and the disclosed embodiments are not limited to use with any particular gas chain configuration. Generally, the disclosed embodiments only require measuring C1 to C5 compounds. It should also be understood that these measurements can be obtained using essentially any suitable gas measurement equipment, including, for example, GC-FID, GC-TCD, GC-MS, MS, adsorption spectroscopy (e.g., IR, MID) and / or Raman spectroscopy.

[0025] Now turn to Figure 3 and Figure 4 , depicts a flow chart of example methods 100 and 150 for estimating reservoir fluid properties from mud logging gas composition measurements. Methods 100 and 150 may be performed in real time during a mud logging operation, for example, while drilling a subterranean wellbore. Figure 3 , gas measurements are received at 102. The gas measurements may include C1, C2, C3, C4, and C5 measurements, such as C1, C2, C3, iC4, nC4, iC5, and nC5 measurements. For example, the gas measurements may be performed using the FlairFlex service described above. In an exemplary embodiment, the gas measurements include circulating a drilling fluid in a wellbore while drilling, degassing a portion of the drilling fluid to obtain a gas sample, and performing a GC measurement to obtain a composition (e.g., C1 to C5 composition) of the gas sample.

[0026] A first machine learning fluid classification model may be used at 104 to evaluate the gas measurement received at 102 to estimate the reservoir fluid type. For example, the reservoir fluid type may be classified as gas, gas condensate, or oil (e.g., black oil). A second machine learning model may be used at 106 to evaluate the gas measurement received at 102 and the estimated fluid type classified at 104 to classify the fluid as normal or abnormal. The gas measurement received at 102 and the estimated fluid type classified at 104 may be further evaluated at 110 or 112 using a third or fourth machine learning model to predict the heavy hydrocarbon fraction (referred to herein as the C6+ fraction) of the reservoir fluid. When the fluid is classified as normal at 106, the third machine learning model may be used at 110, and when the fluid is classified as abnormal at 106, the fourth machine learning model may be used at 112. The gas measurement received at 102, the estimated fluid type classified at 104, and the C6+ fraction predicted at 110 or 112 may be further evaluated at 114 using a fifth machine learning model to predict a gas-oil ratio (GOR) of the reservoir fluid. The estimated fluid type classified at 104, the C6+ fraction predicted at 110 or 112, and the GOR predicted at 114 may then be output, for example, via mud gas logging, at 116. This process may then be repeated substantially any number of times with continuous gas composition measurements.

[0027] Although the present disclosure refers to the heavy hydrocarbon fraction as the C6+ fraction, it should be understood that the disclosed embodiments are not limited thereto. Gas composition measurements typically include C1, C2, C3, C4, and C5 measurements (light hydrocarbons or light alkane gases). Therefore, the heavy hydrocarbon fraction may be referred to as the C6+ fraction. However, in alternative embodiments, the heavy hydrocarbon fraction may refer to other hydrocarbons, such as C8+ or C10+, depending on the nature of the gas composition measurement and the nature of the reservoir fluid.

[0028] Continue to refer Figure 3 , the first machine learning model for fluid type classification at 102 may include a statistical learning model that classifies hydrocarbon fluids as black oil, gas condensate, or dry gas. Advanced statistical learning tools may be used to build a classification model to accurately identify fluid type with a given set of input parameters, such as molar gas composition (C1-C5 mol%). Statistical learning refers to a broad range of tools for exploring and understanding data through statistical models. Classification models may be used to estimate / predict outputs based on one or more inputs and are trained with historical reservoir fluid data.

[0029] A database containing fluid properties of reservoir fluids can be used to build, train, and validate statistical models. In an advantageous embodiment, the database can contain thousands of samples distributed into three types of fluids (oil, gas condensate, and gas). Exploratory data analysis techniques can be used to identify a set of relevant input parameters for the model based on their respective effects on the classification accuracy of the model.

[0030] The database may include the composition of light hydrocarbons (C1-C5 mole %) and the corresponding fluid types (oil, gas condensate, and gas). To ensure data consistency, statistical tools (e.g., Mahalanobis distance) may be advantageously used to identify and remove outliers from the database. The compositional mass balance of the sample may also be checked. After such quality control, the composition may be normalized to 100% to make the data comparable to the composition measured at 102. For fluid type identification, the C1-C5 mole % may be used as an input (predictor) to a classification model.

[0031] The fluid classification model 104 may advantageously utilize various predictors, including humidity Wh, equilibrium Bh, and characteristics Ch of the gas sample, as defined below:

[0032]

[0033] Where C1, C2, C3, C4, and C5 represent the mole % or proportion of each light alkane gas in the measurement made at 102. In such embodiments, training may establish a correlation between one or more of the predictors and the corresponding fluid type.

[0034] In an example embodiment, in order to generate and train a fluid type classification model, a random forest (RF) algorithm can be selected as a classification model. RF is a supervised learning model and is a subclass of a rule-based decision tree algorithm. Given a training data set with input parameters and a target category, a decision tree algorithm can generate a set of rules that can then be used to predict the category using parameters from a new data set. In RF, the model randomly selects predictors from the available input parameters to build a decision tree, and combines many decision trees into a single model. The model calculates votes for each predicted target category and considers the category with the highest vote as the final prediction. However, other classification models, either alone or in combination, can be selected.

[0035] Samples in a pre-existing database can be split for training and validation purposes. For example, 80% of the samples (randomly selected) can be used as a training set, and the remaining (20%) samples can be used for validation. However, other proportions can also be used within the scope of the present disclosure. The fluid type classification model can be trained using, for example, cross-validation techniques. Bayesian optimization (BO), grid search, random search, or random gradient search can be used to optimize RF parameters (the number of trees, etc.) to improve performance. Typically, a large range can be defined for each parameter to be optimized, and then the algorithm can iteratively determine the best parameter combination that minimizes the cross-validation error of the training data.

[0036] In an example embodiment, the workflow for training and implementing a model may use a hierarchical classification model with two sub-models to improve the accuracy of fluid type classification (see Figure 4 ). The first sub-model may identify the fluid as an "oil" type or a "gas condensate / gas" type. The second sub-model may then be used to classify the sample identified as "gas condensate / gas" as either a "gas condensate" or a "gas" type.

[0037] It should be understood that the disclosed embodiments are not limited to the above three fluid types. For example, the classification model (or models) can be extended to include other categories or subcategories of the above three categories, such as light oil, medium oil, heavy oil, volatile oil, etc. In addition, the disclosed embodiments are not limited to using only the gas measurements received at 102. On the contrary, the disclosed embodiments can also allow the use of other measurements, for example, including various logging while drilling measurements. In some embodiments, when a pre-existing database includes such other measurements, these other measurements can improve the classification ability.

[0038] Continue to refer Figure 3 , at 106, the reservoir fluid may be further classified as normal or abnormal. In an example embodiment, at 106, the reservoir fluid may be classified as abnormal when the reservoir fluid shows signs of having undergone a biodegradation process, and the reservoir fluid may be classified as normal when the reservoir fluid does not show such signs of having undergone biodegradation (or shows only minimal signs thereof). In other example embodiments, the reservoir fluid may be classified as abnormal at 106 for other reasons, such as geological processes including fractionation, mixing, secondary processes, temperature and pressure conditions, and changes in source materials. It may be advantageous to distinguish between normal reservoir fluids and abnormal reservoir fluids because such altered fluids may have specific components that may degrade C6+ and GOR predictions. In an example embodiment, the second machine learning model may include a random forest model.

[0039] The second machine learning model (for classifying the reservoir fluid as normal or abnormal at 106) can be trained, for example, using a database as described above, the database including the composition of light hydrocarbons (C1-C5 mole % or C1-C6+ mole %), fluid type, and corresponding normal or abnormal classifications for a large number of gas samples. The second machine learning model can be advantageously trained to identify the relationship between C1-C5 and C6 or C6+, thereby classifying the reservoir fluid as normal or abnormal. The second machine learning model can further utilize many predictors or markers of fluid abnormalities. For example, markers of biodegradation can include one or more ratios of isobutane to normal butane (iC4 / nC4) and isopentane to normal pentane (iC5 / nC5) in the gas composition measurements received at 102, or the ratio of ethane to propane (C2 / C3). In such an embodiment, model training can establish correlations between one or more of these markers and / or relationships and normal or abnormal fluids.

[0040] like Figure 3 As further depicted in , the third and fourth machine learning models can be used to predict the C6+ score of the reservoir fluid. In particular, the third model is used when the reservoir fluid is classified as normal, and the fourth model is used when the reservoir fluid is classified as abnormal. The third and fourth machine learning models (for predicting the C6+ score) may include statistical or regression models, such as a Gaussian process regression (GPR) model. As described above, the database can be used to train and optimize the statistical (regression) models. These statistical models can be developed to predict the C6+ score for each classification (normal or abnormal). The trained third and fourth machine learning models can be configured to predict the C6+ score based on the gas composition measurements made at 102 and the fluid classification estimates made at 104.

[0041] In addition, it will be noted that each of the second classification model and the third and fourth predictive (or regression) models receive as input the gas measurement made at 102 and the fluid classification estimate made at 104. It has been found that the use of the fluid classification estimate tends to improve the predictive ability of the second machine learning model, and therefore can provide more accurate classification of normal and abnormal fluids and more accurate C6+ predictions. Although not wishing to be bound by theory, it is believed that the fluid classification estimate advantageously provides additional information about the underlying structure of the measurement made at 102, which enables the second machine learning model, the third machine learning model, and the fourth machine learning model to capture more complex relationships between continuous features in the measurement and fluid type. In addition, the models tend to be able to learn to distinguish between different groups of observations, which helps to avoid overfitting the data when training the models (because the models can better generalize to new data by taking into account the variability that exists between types).

[0042] In addition, in an example embodiment, the third machine learning model and the fourth machine learning model can estimate the C6+ score based on the "position" of the component in the multidimensional space of composition parameters (e.g., C1, C2, C3, C4, C5, C6, C6, Wh, Bh, Ch, etc.). For example, the set of composition parameters can be combined and the model can be configured to correlate these parameters with the C6+ score. These parameters can be correlated with the C6+ score so that in practice, the model assigns the C6+ score based on a set of values ​​of those parameters (or stated in another way based on the position of the fluid in the aforementioned multidimensional parameter space). It should be understood that the third model and the fourth model can utilize the same set of composition parameters. In such an embodiment, a specific position of the fluid in the multidimensional parameter space can be mapped to a first C6+ score in the third model (for normal fluids) and a different second C6+ score in the fourth model (for abnormal fluids). In other words, the third model can include a first correlation between the multidimensional feature space and the C6+ score, and the fourth model can include a second different correlation between the multidimensional feature space and the C6+ score.

[0043] Still further reference Figure 3 , the gas measurement received at 102, the estimated fluid type classified at 104, and the C6+ score predicted at 110 or 112 can be further evaluated using a fifth machine learning model at 114 to predict the GOR of the reservoir fluid. The fifth machine learning model (for predicting GOR) can include a statistical or regression model, such as a GPR model. As described above, a database can be used to train and optimize a statistical (regression) model to predict the GOR based on the gas composition measurement made at 102, the fluid classification estimate made at 104, and the C6+ estimate made at 110 or 112.

[0044] Note that the fifth machine learning model receives as input the gas measurement made at 102, the fluid classification estimate made at 104, and the C6+ prediction made at 110 or 112. It has been found that using the fluid classification and the C6+ prediction tends to improve the predictive ability of the fifth machine learning model and can therefore provide a more accurate prediction of the gas-to-oil ratio. While not wishing to be bound by theory, and as described above, it is believed that the fluid classification estimate advantageously provides additional information about the underlying structure of the gas measurements made and can help avoid overfitting the data during training.

[0045] Now turn to Figure 4, method 150 includes receiving gas composition measurements at 152. As described above, the gas measurements may include, for example, C1, C2, C3, C4, and C5 measurements made using the FlairFlex service described above. The gas measurements received at 152 may be evaluated using a set of random forest machine learning fluid classification models at 154 and 160 to estimate the reservoir fluid type. In the depicted embodiment, a first random forest model 154 may be configured to evaluate the received gas measurements to group (or classify) the fluid into oil 156 or a subgroup 158 containing gas condensate or gas. The received gas measurements may then be further evaluated using a second random forest model 160 to classify the reservoir fluid (and subgroup 158) into gas condensate 162 or gas 164. In this way, at 166, the reservoir fluid may be classified as oil, gas, or gas condensate.

[0046] Continue to refer Figure 4 , a third random forest machine learning model 170 can be used to classify the reservoir fluid as normal 172 or abnormal 174 based on the gas measurements received at 152 and the fluid type classification estimated at 166. The third random forest machine learning model can be configured to utilize various markers and / or predictors that may indicate evidence of biodegradation or fluid mixing, as described above with respect to Figure 3 The first Gaussian process regression model 182 and the second Gaussian process regression model 184 can be configured to predict the C6+ fraction 186 and 188 of the reservoir fluid for normal and abnormal reservoir fluids based on the gas measurement received at 152 and the fluid type classification estimated at 166. When the fluid is classified as normal 172, the first GPR model can be used at 182, and when the fluid is classified as abnormal 174, the second GPR model can be used at 184. The third GPR model 192 can be configured to predict the gas-oil ratio 194 of the reservoir fluid based on the gas measurement received at 152, the estimated fluid type classified at 166, and the C6+ fraction predicted at 190. The process can then be repeated substantially any number of times based on successive gas composition measurements.

[0047] Continue to refer Figure 3 and Figure 4 , prediction of C6+ fraction and GOR can advantageously use as input gas measurements including C1-C5 components and one or more features extracted using K-means. Extracted features mean that the cluster assignments of the data points generated by the unsupervised K-means clustering algorithm are used as new features or variables. Using cluster assignments as features can sometimes be useful, especially when the original features have complex relationships that are not linearly separable. K-means clustering can help identify underlying patterns in the data and create new features that capture these patterns.

[0048] Adding K-means generated features to a pre-trained model advantageously transfers knowledge of the underlying structure of the data from one database to another. This can help the pre-trained model generalize better to new databases and capture patterns and correlations that may not be present in the original data. In addition, K-means can be easily retrained with new data to establish new relationships and extract new features. Retraining the K-means model with new data can allow for a better fit to the underlying structure of the current data, which can lead to more accurate and relevant feature extraction, which is beneficial for the regression model to predict new data. Adding K-means generated features can act as a form of regularization, helping to avoid overfitting the pre-trained model to the original data. In addition, these additional features can improve model performance by providing a richer and more complete representation of the data.

[0049] In an example embodiment of method 150, the random forest fluid classification submodel utilizes C1-C5 input data and moisture Wh and balance Bh predictors. The random forest normal / abnormal classification model can utilize C1-C5 input data and one or more of the above-mentioned biodegradation markers. The Gaussian process regression model for predicting C6+ score and GOR can utilize C1-C5 input data and moisture Wh and balance Bh predictors and at least one K-means feature.

[0050] As described above, the disclosed embodiments utilize multiple machine learning models, for example, including five different machine learning models. As described above, these models can be trained using a database including thousands of samples. In an advantageous embodiment, the model can also be retrained using a subsequent database including samples associated with a particular basin. By retraining the machine learning model to align with the unique characteristics of each basin, the accuracy and applicability of the model output to a particular basin can be significantly enhanced. The accuracy of the model and its applicability to that particular context.

[0051] It should be understood that the present disclosure includes many embodiments. These embodiments include but are not limited to the following embodiments.

[0052] In a first embodiment, a method for predicting properties of a reservoir fluid during a mud logging operation, the method comprising: obtaining a measured composition of a gas sample obtained during the mud logging operation, the measured composition comprising a selected alkane gas; classifying the fluid type of the reservoir fluid based on the measured composition using a first trained machine learning model; classifying the reservoir fluid as normal or abnormal based on the measured composition and the classified fluid type using a second trained machine learning model; when the reservoir fluid is classified as normal, predicting the heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a third trained machine learning model; when the reservoir fluid is classified as abnormal, predicting the heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a fourth trained machine learning model; and predicting the gas-oil ratio of the reservoir fluid based on the measured composition, the classified fluid type, and the predicted heavy hydrocarbon fraction using a fifth trained machine learning model.

[0053] A second embodiment may include the first embodiment further comprising outputting the classified fluid type, the predicted heavy hydrocarbon fraction, and the predicted gas-oil ratio to a mud log.

[0054] A third embodiment may include any one of the first to second embodiments, wherein the obtaining further comprises: circulating a drilling fluid in a wellbore during drilling; degassing a portion of the drilling fluid to obtain a gas sample; and performing gas chromatography measurement on the gas sample to obtain measured components of the gas sample.

[0055] The fourth embodiment may include any one of the first to third embodiments, wherein: the first trained machine learning model and the second trained machine learning model include a first random forest classification model and a second random forest classification model; and the third trained machine learning model, the fourth trained machine learning model and the fifth trained machine learning model include a first Gaussian process regression model, a second Gaussian process regression model and a third Gaussian process regression model.

[0056] A fifth embodiment may include any one of the first to fourth embodiments, wherein the classified fluid type consists of oil, gas condensate, or gas.

[0057] The sixth embodiment may include any one of the first to fifth embodiments, wherein the first trained machine learning model further includes: a first sub-model configured to classify the fluid type into a first type of reservoir fluid and a group including at least a second type of reservoir fluid and a third type of reservoir fluid; and a second sub-model configured to classify the group into a second type of reservoir fluid or a third type of reservoir fluid.

[0058] A seventh embodiment may include any of the first to sixth embodiments, wherein classifying the reservoir fluid as normal or abnormal further comprises evaluating at least one biodegradation marker with a second trained machine learning model.

[0059] The eighth embodiment may include any one of the first to seventh embodiments, wherein: when the reservoir fluid is classified as normal, the third machine learning model maps the multidimensional gas composition feature space to a first heavy hydrocarbon fraction of the reservoir; and when the reservoir fluid is classified as abnormal, the fourth machine learning model maps the multidimensional gas composition feature space to a second different heavy hydrocarbon fraction of the reservoir.

[0060] A ninth embodiment may include any one of the first to eighth embodiments, wherein the third machine learning model, the fourth machine learning model, and the fifth machine learning model are further configured to evaluate at least one K-means extracted feature.

[0061] The tenth embodiment may include any one of the first to ninth embodiments, wherein: the first machine learning model and the second machine learning model are configured to further evaluate at least one of the humidity Wh, balance Bh and characteristic Ch of the gas sample; and the second machine learning model, the third machine learning model and the fourth machine learning model are configured to evaluate at least one of the humidity Wh, balance Bh and characteristic Ch of the gas sample and at least one K-means extraction feature.

[0062] In an eleventh embodiment, a system for predicting properties of a reservoir fluid during a mud logging operation comprises: a processing system comprising a processor and a memory storing program code instructions, wherein the program code instructions are executable by the processor to: obtain measured compositions of a gas sample obtained during the mud logging operation, wherein the measured compositions comprise selected alkane gases; classify the fluid type of the reservoir fluid according to the measured compositions using a first trained machine learning model; classify the reservoir fluid as normal or abnormal according to the measured compositions and the classified fluid type using a second trained machine learning model; when the reservoir fluid is classified as normal, predict the heavy hydrocarbon fraction of the reservoir fluid according to the measured compositions and the classified fluid type using a third trained machine learning model; when the reservoir fluid is classified as abnormal, predict the heavy hydrocarbon fraction of the reservoir fluid according to the measured compositions and the classified fluid type using a fourth trained machine learning model; and predict the gas-oil ratio of the reservoir fluid according to the measured compositions, the classified fluid type, and the predicted heavy hydrocarbon fraction using a fifth trained machine learning model.

[0063] A twelfth embodiment may include the eleventh embodiment, further comprising: a degasser configured to extract a gas sample from the circulating drilling fluid during the mud logging operation; and a gas chromatography device configured to measure a composition of the measured gas sample.

[0064] The thirteenth embodiment may include any one of the eleventh to twelfth embodiments, wherein: the first trained machine learning model and the second trained machine learning model include a first random forest classification model and a second random forest classification model; and the third trained machine learning model, the fourth trained machine learning model, and the fifth trained machine learning model include a first Gaussian process regression model, a second Gaussian process regression model, and a third Gaussian process regression model.

[0065] A fourteenth embodiment may include any of the eleventh to thirteenth embodiments, wherein the second trained machine learning model is configured to evaluate at least one biodegradation marker when classifying the reservoir fluid as normal or abnormal.

[0066] The fifteenth embodiment may include any one of the eleventh to fourteenth embodiments, wherein: the first machine learning model and the second machine learning model are configured to further evaluate at least one of the humidity Wh, balance Bh and characteristic Ch of the gas sample; and the second machine learning model, the third machine learning model and the fourth machine learning model are configured to evaluate at least one of the humidity Wh, balance Bh and characteristic Ch of the gas sample and at least one K-means extraction feature.

[0067] In a sixteenth embodiment, a method for predicting properties of a reservoir fluid during a mud logging operation, the method comprising: obtaining measured components of a gas sample obtained during the mud logging operation, the measured components comprising a selected alkane gas; classifying the reservoir fluid as normal or abnormal based on the measured components and the classified fluid type using a trained machine learning classification model; predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured components and the classified fluid type when the reservoir fluid is classified as normal using a first trained machine learning regression model; and predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured components and the classified fluid type when the reservoir fluid is classified as abnormal.

[0068] A seventeenth embodiment may include the sixteenth embodiment, wherein classifying the reservoir fluid as normal or abnormal further comprises evaluating at least one biodegradation marker using a second trained machine learning model.

[0069] An eighteenth embodiment may include any one of the sixteenth to seventeenth embodiments, wherein: when the reservoir fluid is classified as normal, the first trained machine learning regression model maps the multidimensional gas composition feature space to a first heavy hydrocarbon fraction of the reservoir; and when the reservoir fluid is classified as abnormal, the second trained machine learning regression model maps the multidimensional gas composition feature space to a second different heavy hydrocarbon fraction of the reservoir.

[0070] A nineteenth embodiment may include any one of the sixteenth to eighteenth embodiments, further comprising: predicting the gas-oil ratio of the reservoir fluid based on the measured components, the classified fluid type, and the predicted heavy hydrocarbon fraction using a third trained machine learning regression model.

[0071] The twentieth embodiment may include the nineteenth embodiment, wherein: the first trained machine learning regression model, the second trained machine learning regression model, and the third trained machine learning regression model are configured to evaluate at least one of the humidity Wh, balance Bh, and characteristic Ch of the gas sample and at least one K-means extracted feature.

[0072] Although the method for machine learning estimation of reservoir fluid properties has been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the disclosure as defined by the appended claims.

Claims

1. A method for predicting properties of a reservoir fluid during a mud logging operation, the method comprising: obtaining a measured composition of a gas sample obtained during a mud logging operation, the measured composition including a selected alkane gas; classifying a fluid type of the reservoir fluid based on the measured composition using the first trained machine learning model; classifying the reservoir fluid as normal or abnormal based on the measured composition and the classified fluid type using the second trained machine learning model; when the reservoir fluid is classified as normal, predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a third trained machine learning model; when the reservoir fluid is classified as anomalous, predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a fourth trained machine learning model; and The fifth trained machine learning model is used to predict the gas-oil ratio of the reservoir fluid based on the measured composition, the classified fluid type, and the predicted heavy hydrocarbon fraction.

2. The method according to claim 1, wherein the obtaining further comprises: Circulating drilling fluid in the wellbore while drilling; degassing a portion of the drilling fluid to obtain a gas sample; and A gas chromatography measurement is performed on the gas sample to obtain measured components of the gas sample.

3. The method of claim 1, wherein the first trained machine learning model further comprises: a first sub-model configured to classify the fluid types into a first type of reservoir fluid and a group including at least a second type of reservoir fluid and a third type of reservoir fluid; and A second sub-model is configured to classify the group as a second type of reservoir fluid or a third type of reservoir fluid.

4. The method of claim 1, wherein classifying the reservoir fluid as normal or abnormal further comprises evaluating at least one biodegradation marker using a second trained machine learning model.

5. The method according to claim 1, wherein: a third machine learning model mapping the multidimensional gas composition feature space to a first heavy hydrocarbon fraction of the reservoir when the reservoir fluid is classified as normal; and When the reservoir fluid is classified as anomalous, a fourth machine learning model maps the multi-dimensional gas composition feature space to a second different heavy hydrocarbon fraction of the reservoir.

6. The method of claim 1, wherein the third machine learning model, the fourth machine learning model, and the fifth machine learning model are further configured to evaluate at least one K-means extracted feature.

7. A system for predicting properties of a reservoir fluid during a mud logging operation, the system comprising: A processing system comprising a processor and a memory storing program code instructions executable by the processor to: obtaining a measured composition of a gas sample obtained during a mud logging operation, the measured composition including a selected alkane gas; classifying a fluid type of the reservoir fluid based on the measured composition using the first trained machine learning model; classifying the reservoir fluid as normal or abnormal based on the measured composition and the classified fluid type using the second trained machine learning model; when the reservoir fluid is classified as normal, predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a third trained machine learning model; when the reservoir fluid is classified as anomalous, predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a fourth trained machine learning model; and The fifth trained machine learning model is used to predict the gas-oil ratio of the reservoir fluid based on the measured composition, the classified fluid type, and the predicted heavy hydrocarbon fraction.

8. The system of claim 7, wherein: The first trained machine learning model and the second trained machine learning model include a first random forest classification model and a second random forest classification model; and The third trained machine learning model, the fourth trained machine learning model, and the fifth trained machine learning model include a first Gaussian process regression model, a second Gaussian process regression model, and a third Gaussian process regression model.

9. The system of claim 7, wherein: The first machine learning model and the second machine learning model are configured to further evaluate at least one of the humidity Wh, the balance Bh, and the characteristic Ch of the gas sample; and The second machine learning model, the third machine learning model, and the fourth machine learning model are configured to evaluate at least one of the humidity Wh, the balance Bh, and the characteristic Ch of the gas sample and at least one K-means extracted feature.

10. A method for predicting a property of a reservoir fluid during a mud logging operation, the method comprising: obtaining a measured composition of a gas sample obtained during a mud logging operation, the measured composition including a selected alkane gas; using the trained machine learning classification model to classify reservoir fluids as normal or abnormal based on the measured composition and classified fluid type; when the reservoir fluid is classified as normal, predicting a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type using a first trained machine learning regression model; and When the reservoir fluid is classified as anomalous, a second trained machine learning regression model is utilized to predict a heavy hydrocarbon fraction of the reservoir fluid based on the measured composition and the classified fluid type.

Images