A method and system for quantitative prediction of carbonate rock cement type and content through well logging
By establishing a cement prediction model and utilizing the correspondence between core experiments and well logging responses, the problem of continuous quantitative evaluation of cement type and content in carbonate reservoirs was solved, enabling accurate prediction of non-cored well sections and improving the consistency and reliability of well logging responses.
Patent Information
- Application Number
- CN202611496343.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies make it difficult to continuously and quantitatively evaluate the type and content of cement in carbonate reservoirs in non-cored well sections, and the logging response prediction results are not consistent with the actual response.
By establishing a cementation prediction model, utilizing the correspondence between core experimental data and well logging responses, cementation-sensitive residuals are constructed. Combined with depth and response scale matching, a random forest regression algorithm is used to predict cementation content. Forward modeling reconstruction and constraint correction are then performed to improve the reliability of the prediction results.
It enables continuous quantitative evaluation of cement type and content in non-cored well sections, improves the predictive consistency and reliability of logging response, reduces confusion between other occurrences of the same mineral, and enhances the continuity and accuracy of reservoir evaluation.
Smart Images

Figure CN123113033A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of well logging technology, specifically to a method and system for quantitative prediction of carbonate rock cement type and content through well logging. Background Technology
[0002] Carbonate reservoirs are an important type of oil and gas reservoir, widely distributed in marine sedimentary basins. Their reservoir properties are influenced by both sedimentary environment and subsequent diagenesis, exhibiting diverse reservoir space types and strong heterogeneity. Cementation is a crucial process in the diagenetic evolution of carbonate rocks, altering the reservoir's pore structure and compaction through pore filling. The formation environments, material sources, and diagenetic stages of calcite, dolomite, siliceous, ferruginous, and gypsum-salt cements vary, resulting in different impacts on reservoir properties. Therefore, identifying the main cement types and quantitatively evaluating their contents is of great significance for analyzing diagenetic processes, reconstructing reservoir evolution history, and assessing reservoir performance.
[0003] The identification and content evaluation of carbonate rock cements typically employ core experimental analysis methods, including thin section observation, cathodoluminescence analysis, scanning electron microscopy, elemental surface mapping, and X-ray diffraction mineral analysis. Among these, image observation and elemental analysis provide information on mineral morphology, occurrence, and elemental distribution, while X-ray diffraction analysis provides information on the overall mineral composition of the sample. However, these methods are primarily applied to cored samples, and due to limitations in the core sampling area and sample distribution, it is difficult to directly obtain information on the type and content of cements continuously distributed along the well depth in non-cored sections.
[0004] Well logging provides formation response data that is continuously distributed along the well depth; however, the logging response is influenced by factors such as particles, matrix, cement, porosity, and fluids. When the same mineral exists in different occurrences, its mineral composition and corresponding logging response may be similar. For example, calcite particles, calcite matrix, and calcite cement all have high Ca content, and it is difficult to distinguish their occurrences based solely on elemental content or single logging parameters such as natural gamma, sonic transit time, or density. Therefore, the overall mineral content cannot be directly equated with the corresponding cement content, and logging interpretation results for lithology identification, mineral content calculation, and porosity evaluation are also difficult to directly use for cement type and content evaluation.
[0005] When relying on single-well curve variations and interpreters' experience to judge cementation, the interpretation results are easily affected by the coupling of multiple factors in the well logging response and empirical differences. Machine learning methods can handle the nonlinear mapping relationship between multiple parameters, providing a technical approach for continuous reservoir evaluation using core experimental data and well logging responses. However, there are spatial scale differences between core experiments and well logging observations. When the location and response range of experimental samples and well logging data are inconsistent, it will affect the reliability of the correspondence between the two. If different occurrences of the same mineral are not sufficiently distinguished during experimental calibration, information about particles, matrix, or replacement products may be mixed into the cement content. Relying solely on statistical mapping to obtain prediction results also makes it difficult to directly determine whether there is consistency between cement type, various contents, and measured well logging responses.
[0006] Therefore, how to distinguish cement from other occurrences of the same mineral, establish a reliable correspondence between experimental calibration and logging response, achieve continuous quantitative evaluation of the main cement types and contents in non-cored well sections, and improve the consistency between predicted results and actual logging responses are technical problems that need to be solved in well logging interpretation of carbonate reservoirs. Summary of the Invention
[0007] The purpose of this invention is to solve the technical problems of easy confusion between well logging responses of cement and other occurrences of the same mineral, difficulty in continuously and quantitatively evaluating the type and content of cement in non-cored well sections, and insufficient consistency between predicted results and actual measured well logging responses. This invention provides a method and system for quantitative prediction of the type and content of carbonate rock cement through well logging.
[0008] To achieve the above objectives, this invention provides a method for quantitative prediction of carbonate rock cement type and content through well logging, comprising: Obtain the measured logging response and component parameters of the target carbonate rock section, determine the low-cementation reference response according to the lithofacies, and construct the cement-sensitive residual based on the difference between the measured logging response and the low-cementation reference response. The measured logging response and cement-sensitive residuals are input into a pre-established cement prediction model to predict the total cement content and the composition ratio of various cements. The cement prediction model is established based on cement occurrence decoupling labels and logging data matched with depth and response scales. The content of each type of cementitious material is determined based on the total cementitious material content and composition ratio; The logging response is reconstructed using the contents and component parameters of the various cementing materials in a forward model to obtain the reconstructed logging response. Based on the deviation between the reconstructed logging response and the measured logging response, the contents of various cementing materials are constrained and corrected to obtain the contents of various cementing materials after processing. The type corresponding to the maximum value among the various cementitious contents after the treatment is taken as the main cementitious type, and the cementitious type and content prediction results along the well depth are formed.
[0009] As a further improvement to the above technical solution, the construction of the cement occurrence decoupling tag includes: acquiring cast thin section images, cathodoluminescence images, and SEM-EDS elemental maps of the same core sample from the same field of view, and performing spatial registration; classifying the same mineral into four occurrence types—grains, matrix, replacement products, and cement—based on the mineral positions, pore filling relationships, crystal boundaries, grain contacts, and luminescent bands in the registered images; constructing a cement-specific mask based on the cement occurrence region, and only statistically analyzing the area and elemental content of each region within the cement-specific mask; determining the cement type of each region based on the elemental content, and determining the corresponding content by the ratio of the area of each cement region to the effective statistical area of the thin section, thereby generating the cement occurrence decoupling tag.
[0010] As a further improvement to the above technical solution, the establishment of the cementitious prediction model includes: obtaining the overall mineral composition information of the core sample through X-ray diffraction; applying an upper limit constraint or a mineral quantity balance constraint to the cementitious content corresponding to the cementitious occurrence decoupling tag based on the corresponding total mineral content in the overall mineral composition information; generating the tag confidence score of the cementitious occurrence decoupling tag based on the consistency of the results of the core sample's cast thin section, cathodoluminescence, SEM-EDS, and X-ray diffraction experiments; and weighting the corresponding calibration samples in the establishment of the cementitious prediction model based on the tag confidence score.
[0011] As a further improvement to the above technical solution, the depth and response scale matching includes: performing correlation analysis on the experimental reference sequence of the core sample and the measured logging curves of the corresponding well section to determine the core depth offset and matching confidence level; correcting the sample depth based on the core depth offset; setting a corresponding longitudinal response kernel according to the longitudinal resolution of each logging instrument; selecting a matching window near the corrected sample depth that does not cross obvious lithological interfaces; convolving the longitudinal response kernel with the logging curves within the matching window to obtain a representative logging value for the corresponding sample depth; using the representative logging value as the logging input when establishing the cement prediction model; and reducing the weight of samples with low matching confidence based on the matching confidence level.
[0012] As a further improvement to the above technical solution, the measured logging response includes logging parameters such as natural gamma, sonic transit time, density, compensated neutron, deep lateral resistivity, and photoelectric absorption cross section, as well as elemental logging parameters for Ca, Mg, Si, Fe, S, and Sr. Based on the elemental logging parameters at the same depth, Mg / Ca, Si / Ca, Fe / Ca, and S / Ca elemental ratio parameters are constructed. The measured logging response and elemental ratio parameters are combined with the input parameters of the cementitious prediction model to construct a logging input matrix, and standardized according to the minimum and maximum values of the corresponding parameters in the training samples. The standardized logging input matrix and the cementitious sensitive residual are jointly input into the cementitious prediction model.
[0013] As a further improvement to the above technical solution, the low-cementation reference response includes the density, photoelectric absorption cross section, and elemental background response under the combined influence of particles, matrix, pores, and fluids in the corresponding lithofacies. The corresponding low-cementation reference response is determined based on the lithofacies at the target depth location. The measured logging response at the target depth location is matched with the low-cementation reference response according to logging parameters. The corresponding background response is subtracted from the measured logging response to construct the cement-sensitive residual. The cement-sensitive residual is used as a supplementary variable and input into the cement prediction model along with the measured logging response at the target depth location.
[0014] As a further improvement to the above technical solution, the cementitious material prediction model adopts a random forest regression algorithm to predict the total cementitious material content and the composition ratio of five types of cementitious materials: calcite, dolomite, silica, ferrous, and gypsum. The non-negative composition ratio prediction values of the five types of cementitious materials are divided by their sum to obtain the normalized composition ratio with a sum of 1. The total cementitious material content is multiplied by the normalized composition ratio of each type of cementitious material to obtain the corresponding content of each type of cementitious material.
[0015] As a further improvement to the above technical solution, the component parameters include the volume fractions of particles, matrix, and pore fluid, as well as the density, photoelectric absorption cross section, and elemental weight fraction of particles, matrix, pore fluid, and various cements; the constraint correction includes: determining the upper limit of the content of various cements based on the total mineral content, elemental mass balance, and fillable porosity; using the content of various cements and component parameters to forward model the logging response of density, photoelectric absorption cross section, and elements to obtain the logging reconstruction response; comparing the logging reconstruction response with the measured logging response at the same depth and with the same logging parameters to obtain the corresponding deviation; when the absolute value of the deviation exceeds the corresponding set threshold, correcting the content of various cements based on the deviation, under the constraint that the content of various cements is non-negative and does not exceed the corresponding upper limit of content; otherwise, retaining the original predicted content.
[0016] As a further improvement to the above technical solution, it also includes: determining the prediction confidence by using the depth matching confidence, model divergence, and the forward modeling residual between the reconstructed logging response and the measured logging response; and determining the cementation strength level based on the sum of the contents of various cementing materials after correction and the preset level classification conditions. When the prediction confidence is not lower than the first threshold, the main cementing material type, the single value of the content of each type of cementing material after processing, and the cementing strength level are output; when the prediction confidence is not lower than the second threshold but lower than the first threshold, the cementing material content range is output; when the prediction confidence is lower than the second threshold, the deterministic cementing material type and content result are not output, and the rejection mark and corresponding well section information are output; the second threshold is lower than the first threshold.
[0017] This invention also provides a well logging quantitative prediction system for carbonate rock cement type and content, comprising: The data acquisition module is used to acquire the measured logging response and composition parameters of the target carbonate rock well section; The background processing module is used to determine the low-cementation reference response according to the lithofacies, and to construct the cementation sensitive residual based on the difference between the measured logging response and the low-cementation reference response. The joint prediction module is used to input the measured well logging response and cement-sensitive residuals into a pre-established cement prediction model to predict the total cement content and the composition ratio of various cements; the cement prediction model is established based on cement occurrence decoupling labels and well logging data matched with depth and response scales; The content determination module is used to determine the content of various cementitious substances based on the total cementitious substance content and composition ratio. The forward modeling correction module is used to reconstruct the logging response using the contents and component parameters of the various cementing materials to obtain the reconstructed logging response. Based on the deviation between the reconstructed logging response and the measured logging response, the content of various cementing materials is constrained and corrected to obtain the content of various cementing materials after processing. The result generation module is used to take the type corresponding to the maximum value among the various cement contents after processing as the main cement type, and to generate cement type and content prediction results along the well depth.
[0018] The method and system for quantitative prediction of carbonate rock cement type and content through well logging provided by this invention have the following beneficial effects: 1. By establishing the correspondence between core test calibration results and logging responses, and using a pre-established cement prediction model to predict the target well section, the experimental understanding of core samples is extended to non-core well sections with corresponding logging data. This results in the prediction of cement types and contents distributed along the well depth, thereby improving the continuity of the evaluation of cementation in carbonate reservoirs.
[0019] 2. By spatially registering thin section images, cathodoluminescence images, and SEM-EDS elemental maps of the same field of view, the occurrence of grains, matrix, replacement products, and cements of the same mineral is distinguished. A cement-specific mask is used to limit the statistical region, reducing the interference of non-cement components on the content calibration. Combined with X-ray diffraction mineral composition constraints and label confidence weighting, the reliability of cement calibration data is improved.
[0020] 3. By correcting core depth offset, setting response kernels based on the longitudinal resolution of logging instruments, and extracting representative logging values within a matching window that does not cross obvious lithological interfaces, the influence of depth misalignment, response scale mismatch, and mixed responses of different lithologies on the calibration relationship is reduced, thereby improving the correspondence between core experimental data and logging data.
[0021] 4. By comprehensively utilizing conventional logging parameters, elemental logging parameters, and elemental ratio parameters, and constructing the difference between the measured logging response and the low-cementation reference response of the same lithofacies as a cement-sensitive residual, this residual is used in model prediction to supplement the response difference information of different cements. A random forest regression algorithm is employed to establish a nonlinear mapping relationship between multiple parameters and cement content, improving the ability to utilize complex logging responses.
[0022] 5. By predicting the total cement content and the composition ratio of various cements separately, the non-negative composition ratios are normalized and multiplied by the total cement content, so that the sum of the contents of various cements calculated in the joint prediction stage is consistent with the total predicted amount; the main cement types are determined based on the contents of various cements after constraint correction, so that the final type determination is consistent with the content ranking, and the correlation evaluation between cement type and content is realized.
[0023] 6. Determine the upper limit of cement content based on total mineral content, elemental mass balance, and fillable porosity. Reconstruct density, photoelectric absorption cross section, and elemental logging response using the predicted content and component parameters. Correct prediction results that exceed the deviation limit based on the deviation between the reconstructed response and the measured response, under the constraints of non-negative content and upper limit of content, thereby improving the physical rationality of the prediction results and their consistency with the measured logging response.
[0024] 7. The prediction confidence is determined by combining the confidence level of deep matching, the model divergence degree, and the forward modeling residuals. Based on this, the output method of single content value, content range, or rejection is adopted to make the output result reflect the corresponding reliability level and reduce the interference of low confidence deterministic results on reservoir evaluation. The cementation strength level is determined by combining the sum of the contents of various cementing materials to form a correlation evaluation of cementing material type, content, and cementation degree. Attached Figure Description
[0025] Figure 1 This is a flowchart illustrating the operation of the method for quantitative prediction of carbonate rock cement type and content using well logging in this embodiment of the invention. Figure 2 This is a structural diagram of the quantitative prediction system module for the type and content of carbonate rock cement in well logging, as described in this embodiment of the invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.
[0027] Example 1 like Figure 1 As shown, this embodiment provides a method for quantitative prediction of carbonate rock cement type and content using well logging, including: acquiring the measured well logging response and component parameters of the target carbonate rock interval; determining a low-cementation reference response according to lithofacies; constructing a cement-sensitive residual based on the difference between the two; inputting the measured well logging response and cement-sensitive residual into a pre-established cement prediction model to predict the total cement content and the composition ratio of various cements, and determining the content of various cements accordingly; reconstructing the well logging response using the content and component parameters of various cements through forward modeling; performing constraint correction processing based on the deviation between the reconstructed well logging response and the measured well logging response; and using the type corresponding to the maximum value among the processed cement contents as the main cement type to form a prediction result of cement type and content along the well depth. The cement prediction model is established based on cement occurrence decoupling labels and well logging data matched with depth and response scales.
[0028] Figure 1 The processing steps are sequentially arranged as follows: data acquisition, joint prediction, content determination, forward modeling correction, and result formation. In the model building stage, the experimental results from the core section are used to establish a calibration relationship with the corresponding logging response; in the application stage, the established model is used to process the logging data of the target well section. Cementation occurrence decoupling tags are used to distinguish the calibration contribution of the same mineral when it exists as cement or in other occurrences. Depth and response scale matching are used to establish the correspondence between core experimental points and representative logging values.
[0029] The low-cementation reference response represents the background response of the corresponding lithofacies, while the cement-sensitive residual provides information on the difference between the measured response and this background. The total cement content represents the overall cement content, and the proportions of various components represent the percentages of different cements in the total cement content. The forward modeling correction process re-maps the content prediction results to the logging response, verifying their consistency with measured data at the same depth. The processed content includes the content obtained after constraint correction when the deviation exceeds the limit, and the predicted content retained when the deviation does not exceed the limit.
[0030] This method is applicable to the identification, quantitative calculation, and vertical continuous prediction of different types of cement in complex carbonate reservoirs. It can be used for diagenesis evaluation, reservoir space evolution analysis, reservoir prediction, and fine reservoir characterization during oil and gas exploration and development. The following section uses a specific deep carbonate reservoir as an example to illustrate the experimental calibration, model establishment, and application process in non-cored well sections, combining each processing step.
[0031] Example 2 In this embodiment, when constructing cement occurrence decoupling tags, thin section images, cathodoluminescence images, and SEM-EDS elemental maps of the same core sample from the same field of view are acquired and spatially registered. Based on mineral location, pore filling relationship, crystal boundary, grain contact, and luminescent zonation, the same mineral is classified into four occurrence types: grains, matrix, replacement products, and cement. A cement-specific mask is constructed based on the cement occurrence regions. Only the area and elemental content of each region within the mask are statistically analyzed. The cement type is determined based on the elemental content, and the corresponding content is determined by the ratio of the area of each cement region to the effective statistical area of the thin section, thus generating cement occurrence decoupling tags.
[0032] First, representative carbonate core samples were selected according to the research objectives. Sample types included grainy limestone, micritic limestone, bioclastic limestone, dolomite, and other carbonate reservoir rock types. The following experimental analyses were conducted on the core samples: In this example, three core wells (A, B, and C) in the study area were selected as model wells, yielding a total of 292 carbonate rock core samples. The sample quantities for each experiment are shown in Table 1. Each sample was identified and statistically analyzed based on its available experimental data. Samples with multi-source images of the same field of view were used for corresponding attitude decoupling calibration.
[0033] Table 1. Statistics of Core Samples and Experimental Data
[0034] Thin section analysis of cast bodies is used to observe the occurrence, crystal morphology, filling pattern, and relationship with pore space of cements. Calcite cements typically appear as bright calcite filling intergranular pores; dolomite cements appear as subhedral to euhedral crystals, while dolomite replacement structures are classified as replacement products for distinction; siliceous cements appear as microcrystalline quartz aggregates or intergranular fillings; ferrous cements appear as reddish-brown iron oxides and hydroxides; and gypsum-based cements appear as sulfate mineral fillings.
[0035] Cathodoluminescence analysis utilizes the luminescence response of minerals under electron beam excitation to obtain luminescence color, intensity, and zonal structure characteristics. Calcite cements exhibit varying degrees of orange-yellow or dark red luminescence depending on the Mn and Fe element content; dolomite cements show weak luminescence or zonal luminescence characteristics. This information serves as auxiliary basis for experimental identification and label construction. Scanning electron microscopy analysis further observes crystal morphology, grain size, cement filling pattern, and pore blockage degree, providing microstructural characteristics.
[0036] Elemental surface scanning was used to obtain the spatial distribution of elements such as Ca, Mg, Si, Fe, Mn, S, and Sr. Within the identified cementation occurrence area, calcite cement exhibited a high Ca content and low Mg, Si, and Fe contents; dolomite, silica, ferruginous, and gypsum-based cements were identified by combining the characteristics of increased Mg / Ca, Si / Ca, Fe / Ca, and S / Ca ratios, respectively. In this example, five types of cement were identified by combining thin section analysis, luminescence, microstructure, and elemental distribution. Their experimental identification characteristics and corresponding well logging auxiliary information are shown in Table 2.
[0037] Table 2 Experimental identification characteristics of different cementing materials
[0038] During image statistics, color correction, noise removal, and mineral region segmentation are first performed on the thin section images. Then, regions are automatically identified by combining crystal color, texture, and elemental characteristics. After spatial registration of different images, mineral regions are classified by occurrence. A special mask for cementing materials limits the area statistics to avoid including grains, matrix, or replacement areas of the same mineral in the cementing material. The area content is calculated as follows: %, where, This indicates the relative content of the i-th type of cementitious material. This indicates the area of the identification region for this type of cementitious material. This represents the effective statistical area of the thin section. Among them, The index represents the type of cementitious material; all area items use the same area unit, making the area ratio a dimensionless quantity.
[0039] Taking sample A-01 from well A as an example, the effective statistical area of the thin section is: The occurrence of calcite, dolomite, quartz, iron minerals, and sulfate minerals was determined in the registration field of view, and the results are shown in Table 3.
[0040] Table 3. Decoupling and calibration results of mineral occurrence in sample A-01
[0041] The total area of the mask for cementitious materials is The remaining areas in the thin slice are Dolomite replacement products Silicide replacement products All of these are placed outside the cementitious material mask, so that the cementitious material content statistics do not include these replacement-genetic regions.
[0042] Substitute the mask statistics of the sample into ,in , and Various area content, mask area and Effective statistical area analysis yielded: calcite cement. Dolomite cement Siliceous cement Iron cementitious material ; Gum and salt cement The subscripts of the above content symbols , , , and The following categories are identified as calcite, dolomite, silica, ferrous and gypsum-based cements; the content and classification probability of the same type in the following text follow this type correspondence.
[0043] To organize experimental calibration information, a database of quantitative cement parameters was established, including thin section images, cathodoluminescence, elemental composition, mineral composition, and well logging response characteristics. Thin section image parameters include cement area ratio, grain size, crystal morphology, pore filling ratio, and image grayscale; area ratio reflects the degree of cementation. Cathodoluminescence parameters include color coding value, intensity, homogeneity, number of zonations, and zonation width, used to help distinguish cements from different diagenetic stages. The above experimental information is associated with cement type, content, and corresponding well logging responses and stored accordingly.
[0044] Some images and calibration parameters of sample A-01 are shown in Table 4. The content recorded is 26.37% after mineral composition constraint correction, as described later, to distinguish it from the uncorrected thin film mask area ratio.
[0045] Table 4. Images and calibration parameters of sample A-01
[0046] Experimental element parameters can be expressed as And record the element combination parameters. , , , .in , , and These elements are used to characterize changes in elemental ratios related to cementation processes involving dolomite, silica, ferrous metals, and gypsum salts. The database also records combinations of well logging parameters. This ensures that the experimental element information and the downhole logging information maintain their respective sources and correspondences.
[0047] In the elemental statistics of sample A-01, the following method was used: The corresponding ratio is denoted as , , and The content of the experimental elements used is shown in Table 5.
[0048] Table 5 Experimental elemental parameters of sample A-01
[0049] Based on the elemental content calculations described above, we obtain... , , and These values are core test parameters, while the model input ratios for the target well section are constructed separately using elemental logging parameters at the same depth.
[0050] Example 3 In the calibration phase of model building, the overall mineral composition information of the core sample is obtained through X-ray diffraction. Based on the total amount of the corresponding minerals, the content upper limit constraint or mineral quantity balance constraint is applied to the content corresponding to the cement occurrence decoupling label. According to the consistency of the results of casting thin section, cathodoluminescence, SEM-EDS and X-ray diffraction experiments, the label confidence score is generated, and the corresponding calibration samples are weighted accordingly.
[0051] XRD analysis provides information on the overall mineral composition of the sample, including calcite, dolomite, quartz, and sulfate minerals. Cementitious materials constitute only a portion of the corresponding minerals; therefore, the overall mineral composition is used to constrain the masking statistical results and cannot be directly used as the cementitious material content. Thin section observation is a two-dimensional statistical method, while XRD provides overall mineral analysis. Under the combined constraints of both, a mineral composition correction factor is used to correct the area statistical results. .in, This indicates the corrected relative content of cementitious material. This represents the correction factor for the i-th type of cement; this factor is determined jointly by the statistical results of the thin section area and the XRD mineral analysis results.
[0052] To unify the measurement standards of thin section area statistics and XRD mineral analysis results, both are standardized to a mass fraction standard for comparison and constraint. Thin section mask statistics yield the area fraction of various cementing materials, which can be approximated as a volume fraction according to Delesse's principle. For the same mineral, the proportion of its cementing occurrence in the total mineral area is the same as its proportion in the total mineral mass; therefore, the occurrence proportion is not affected by area-mass conversion. The correction factor is determined jointly by this occurrence proportion and the XRD mineral mass fraction: for paired samples with available XRD data, the area proportion of cementing occurrence in the mineral corresponding to the i-th type of cementing material is statistically analyzed. The whole-rock mass fraction of this mineral was determined by XRD. Product of this percentage As the content of this type of cementitious material at the mass fraction level, it is compared with the content at the sheet area level. In comparison, for multiple paired samples of the same lithofacies from the same well, a least-squares fitted straight line passing through the origin is used, and the slope of the line is the correction coefficient for that lithofacies in that well. For samples without XRD data, the correction factor is taken as the determined value of paired samples from the same well and the same lithofacies.
[0053] Taking the acquisition of the dolomite cement correction coefficient in the paired grain limestone samples from Well A as an example: the area ratio of cement occurrence in the dolomite of each paired sample was statistically analyzed according to its occurrence. and the whole-rock mass fraction of dolomite by XRD The product of this percentage and the product of the area and caliber percentages is taken as the mass caliber content. In this example, the caliber content of the sheet area of the five paired samples is 3.9%, 4.8%, 6.2%, 7.5%, and 8.1%, respectively, and the corresponding mass caliber contents are 4.3%, 5.4%, 6.9%, 8.4%, and 9.1%, respectively. The slope of the fitted line is 1.12, that is... Listed in Table 6. Substitute sample A-01. Subsequently, the dolomite cement content was corrected from 4.8% of the area caliber to 5.38%, which is consistent with the calibration value in Table 7 and is lower than the upper limit of mass balance determined by the whole rock mass fraction of dolomite in XRD for this sample. This indicates that the corrected area statistics result and the XRD mass content are mutually verified under the same measurement caliber.
[0054] The corrected content of various cementitious materials is checked against the following upper limits, and is truncated if it exceeds the upper limit: ① Upper limit of total mineral content: the whole-rock mass fraction of the corresponding mineral measured by XRD, i.e., the corrected content must not exceed the total amount of all occurrences of that type of mineral; ② Upper limit of elemental mass balance: for ferrous and gypsum-salt cements, the Fe and S content obtained from elemental logging or core elemental analysis is divided by the weight fraction of that element in the corresponding cement; ③ Upper limit of fillable porosity: the original porosity estimate determined based on sedimentary facies and adjacent well data, and the sum of the contents of various cements must not exceed this value. Therefore, the contents of various cements obtained in the calibration stage are simultaneously constrained by the upper limits of total mineral content, elemental mass balance, and fillable porosity, ensuring that the label values used for subsequent modeling maintain mass balance with the XRD mineral composition.
[0055] Sample A-01 according to Perform calibration correction, among which The corrected content, Take the corresponding coefficients listed in Table 6.
[0056] Table 6 Mineral Composition Correction Factors for Sample A-01
[0057] This yields calcite cement. Dolomite cement The calculations for the other types are performed in the same manner, and the final calibration values are shown in Table 7.
[0058] Table 7. Results of cementitious material content determination for sample A-01
[0059] Based on the consistency of identification across multiple experiments for this sample, the label confidence scores for calcite, dolomite, silica, ferrous, and gypsum-based cements were 0.95, 0.90, 0.88, 0.82, and 0.85, respectively. These label confidence scores were used to weight the corresponding calibration samples in the model building process, incorporating the reliability of experimental identification into the training. The label confidence scores correspond to the experimental calibration quality, while the prediction confidence scores (described later) are used to control the output of the prediction results for the target well section.
[0060] The label confidence level was derived from the consistency of multiple experimental identifications. The conversion rule was as follows: For each cement type, the number of consistent results from four experiments (cast thin section, cathodoluminescence, SEM-EDS, and XRD) supporting the identification of that type was counted. Consistency across four experiments was assigned a score of 0.90–0.95, consistency across three experiments was assigned a score of 0.85–0.90, consistency across two experiments was assigned a score of 0.75–0.85, and single-item evidence was assigned a score of 0.60–0.75. Within the same grade, the score was adjusted according to the significance of elemental and occurrence characteristics. In sample A-01, calcite cement showed consistency across four experiments, assigned a score of 0.95; dolomite cement showed consistency across three experiments and was supported by XRD mineral total content, assigned a score of 0.90; siliceous cement showed consistency across three experiments, assigned a score of 0.88; gypsum-salt cement was assigned a score of 0.85; and ferrous cement showed relatively weak identification characteristics, assigned a score of 0.82.
[0061] For samples lacking complete SEM-EDS or XRD data, labels were obtained as follows: ① Occurrence decoupling and type identification were primarily based on cast thin sections and cathodoluminescence (all 292 samples in this example possessed both data), with SEM-EDS elemental maps and XRD serving as supporting evidence; ② Mass balance constraints were applied sample-by-sample to samples with XRD data, while for samples lacking XRD data, upper limits were applied based on the XRD mineral composition range of already calibrated samples from the same well and lithofacies. For iron and gypsum-salt cements, elemental logging or elemental analysis were also used for elemental mass balance constraints; ③ Label confidence levels were determined according to the above grading based on the actual number of consistent experimental items available for the sample, and weighted by confidence level for model establishment. Thus, all 292 samples had clear calibration criteria: samples with complete data were used to form sample-by-sample mass balance constraints, while samples with incomplete data participated in model establishment under constraints from the same well and lithofacies and confidence level weighting. Differences in calibration reliability were quantitatively expressed through label confidence levels.
[0062] This example compiles experimental calibration data from 292 core samples to form sample records for subsequent modeling and validation. All statistics are performed within the same measurement caliber across all 292 samples, without distinguishing between sample subsets or type dominance. The detailed records of some samples listed in Table 16 are consistent with the statistical results. The statistical average contents of various cementing materials are 18.6%, 23.4%, 7.8%, 5.2%, and 9.6%, respectively; the statistical maximum values are 52.4%, 61.8%, 26.5%, 18.4%, and 35.2%, respectively; and the statistical minimum values are 3.2%, 5.38%, 0.85%, 0.22%, and 0.31%, respectively. The minimum value for calcite cement corresponds to low-content samples not listed in Table 16, while the minimum values for dolomite, silica, ferrous, and gypsum-based cements correspond to samples A-01, A-02, A-03, and A-03, respectively. The specific contents of each sample are listed in the subsequent training data table.
[0063] Example 4 In the depth and response scale matching stage, correlation analysis was performed between the experimental reference sequence of the core sample and the measured logging curves of the corresponding well section to determine the core depth offset and matching confidence level, and the sample depth was corrected based on the offset. A vertical response kernel was set according to the vertical resolution of each logging instrument. A matching window that does not cross obvious lithological interfaces was selected near the corrected sample depth. The convolution result of the response kernel and the logging curve within the window was used as the representative logging value of the sample for model building, and the training weight of samples with low matching confidence was reduced.
[0064] The core depth corresponding to the experimental sample is denoted as The well logging sampling depth is recorded as Establishment was achieved through core repositioning, well depth correction, and marker layer calibration. The depth correspondence is represented. The range used for window positioning can be written as... ,in In this window representation, half-width is used. The local area represented by the thin section differs from the longitudinal integrated response of the logging tool; therefore, response scale matching is required after depth localization. Thin section analysis is typically at the millimeter-centimeter level, while the logging response reflects meter-level longitudinal integrated information. In the window representation... The depth of the well logging sampling points within the window. The sample depth after repositioning is used; the offset in the aforementioned window half-width and depth correction relationship is denoted as... However, each value should be taken according to the definition of its respective relationship.
[0065] The original core depths and experimental records for some samples from Well A are shown in Table 8. Core shrinkage, missing cores, and well depth errors during the coring process may cause depth misalignment. Calibration can be performed first by combining the characteristic stratigraphic positions of natural gamma and sonic curves, and then the offset can be determined by correlation analysis between the experimental reference sequence and the measured curves.
[0066] Table 8 Depth Record of Core Experiment Samples
[0067] Depth correction relationship is ,in To correct the core depth, The original record depth, In this correction relationship, the depth correction amount is represented. The correction amount for sample A-01 is... The original depth is After correction, the result is The corresponding logging depth is 4210.70m.
[0068] In the coring section of well A, the XRD mineral content sequence was used as the experimental reference sequence and cross-correlated with the logging curves to obtain the depth offset and matching confidence shown in Table 9.
[0069] Table 9. Depth offset and matching confidence levels between core and well logging data
[0070] After completing the depth calibration, response kernels corresponding to the instrument's longitudinal resolution were set for GR, AC, DEN, CNL, RT, PE, and elemental logging, respectively. The parameters are shown in Table 10.
[0071] Table 10. Longitudinal resolution and response kernel parameters of various logging instruments
[0072] In this example, a triangular kernel is used as the response kernel, and its weight decreases linearly with increasing distance from the corrected sample depth. Each logging curve is convolved with its corresponding response kernel. For example, the DEN response kernel for A-01 has a half-width of 0.20m, and the resulting representative logging value is... PE and elemental logging data are processed using their respective kernels. When there are significant lithological abrupt changes in the window, only curve data within the lithologically consistent segment are used for convolution.
[0073] Match confidence also participates in sample weighting. For example, A-05 has a match confidence of 0.51, so its training weight is reduced to 0.40; A-01 and A-02 have higher match confidence, so their training weights in the matching stage are both 1.00. This weight is used to reduce the impact of deep matching error on model building, while label confidence reflects the calibration quality from the degree of consistency in experimental identification.
[0074] To compare differences in response scaling, a fixed-window arithmetic mean can be used. The calculation is performed, where n is the number of sampling points within the window, and the summation term represents the summation of each sampled value within the window. This average value is used for comparison and does not replace the modeling input obtained from the instrument response kernel. The average value in this formula... This represents the window arithmetic mean of the currently calculated logging parameters. For this parameter, the first one in the window Well logging values at each sampling point The index of the sampling point within the window is used; summation is performed on all sampling points within the window.
[0075] Fixed window comparison Define the scope, This represents the half-width of the window. (Take...) At that time, the range of window A-01 is ;according to The arithmetic mean of the sampling points within the window is calculated, and the five sets of DEN data are listed in Table 11.
[0076] Table 11 Density sampling data in fixed-window control
[0077] From the five sample values listed The fixed window result and the representative value of the response kernel convolution are used for comparison. The formal modeling uses the representative logging value processed according to the response scale of each instrument.
[0078] Example 5 In constructing the logging input, the measured logging responses include natural gamma, sonic transit time, density, compensated neutrons, deep lateral resistivity, and photoelectric absorption cross-section parameters, as well as elemental logging parameters for Ca, Mg, Si, Fe, S, and Sr. Mg / Ca, Si / Ca, Fe / Ca, and S / Ca ratios are constructed from elemental logging parameters at the same depth. These are combined with the model's input parameters to form the logging input matrix. The matrix is standardized based on the minimum and maximum values of each parameter in the training samples, and then input into the model along with the cementation-sensitive residuals.
[0079] The aforementioned conventional logging parameters are denoted as GR, AC, DEN, CNL, RT, and PE, respectively. GR reflects variations in clay content and radioactive elements; AC reflects the rock's elastic properties and pore structure; DEN reflects the overall mineral composition and cementation degree; CNL reflects pore space and hydrogen-containing substances; RT reflects fluid properties and pore structure; and PE provides information on variations in mineral composition. In elemental logging, Ca and Mg mainly reflect carbonate mineral composition, while Si, Fe, and S provide information related to silica, ferric, and sulfate, respectively, and are used in conjunction with parameters such as Sr for joint analysis.
[0080] The element ratios used for well logging prediction are expressed as follows: , , , After adding measured well logging parameters, it constitutes The ratios are calculated using logging data from the same depth, avoiding the direct substitution of the aforementioned core experiment element ratios into the logging input for the well section.
[0081] Table 12 shows the representative logging values of sample A-01 at 4210.70m and the ratios calculated from them. Among them, Mg / Ca, Si / Ca, Fe / Ca and S / Ca are 0.089, 0.033, 0.016 and 0.012 respectively.
[0082] In this example, the sonic transit time is uniformly measured in μs / m. The sonic transit time recorded in μs / ft in the original well logging data is converted to approximately 3.2808 μs / m by 1 μs / ft before being used for standardization and model input. For example, the sonic transit time recorded at 61.2 μs / ft is converted to 200.8 μs / m to avoid the mixing of units affecting model training and prediction.
[0083] Table 12. Logging input and elemental ratio parameters for sample A-01
[0084] After matching, data records are established according to the samples, and the basic correlation structure is as follows: The basic logging portion of X is... Y stores the corresponding experimental labels; when used for content modeling, the label records include cement occurrence decoupled content data and corresponding total amount and composition information; when used for type comparison, the main cement type is read. Table 13 lists some basic logging parameters and type records.
[0085] Table 13. Well logging and main cementing material records for some samples.
[0086] Standardization Adoption In the formula These are the standardized logging parameters, where X is the original value of the parameter. and These are the minimum and maximum values of the corresponding parameters in the training samples, respectively. The same parameter standardization criteria are used for both model building and application.
[0087] The above standardization process is demonstrated using GR on the sample, according to... In the training samples , The measured value of A-01 is ,get The standardized inputs for this sample are shown in Table 14.
[0088] Table 14 Standardized logging inputs for sample A-01
[0089] Example 6 When constructing cement-sensitive residuals, the low-cementation reference response includes the density, photoelectric absorption cross section, and elemental background response under the combined influence of particles, matrix, pores, and fluids in the corresponding lithofacies. The reference response is determined based on the lithofacies at the target depth. The measured response and reference response are matched item by item according to logging parameters. The corresponding background response is subtracted from the measured response, and the resulting residual is used as a supplementary variable. This residual, along with the measured logging response at that depth, is input into the cement prediction model.
[0090] The lithofacies determination at the target depth adopts a combination of core calibration and well logging discrimination: the lithofacies of the cored well section is determined by core description, and a lithofacies-well logging response template is established. The template parameters include well logging parameters sensitive to lithofacies, such as natural gamma, density, photoelectric absorption cross section, and elemental ratios. The lithofacies of the target well section are determined by cluster discrimination using this template, and the determination result is recorded as the corresponding lithofacies label. The establishment and selection rules for the low-cementation reference response of the same lithofacies are as follows: among the samples of the same lithofacies, the same well, or adjacent cored wells, the batch of samples with the lowest sum of the calibration contents of various cements is selected (in this example, samples with a total content not exceeding 5% are selected; if the quantity is insufficient, samples are supplemented in ascending order of total content). The representative well logging values are extracted according to the response kernel method in Example 4, and the statistical median is taken as the low-cementation reference response of the well and the lithofacies. The reference response library is formed by summarizing by well and lithofacies.
[0091] Under the same lithofacies conditions, particle type, matrix content, pore structure, and fluid properties are similar, and their common influence is already included in the low-cementation reference response. The difference between the measured response and this reference response mainly reflects the response change caused by cement filling, thus distinguishing the background changes of particles, matrix, pores, and fluids from the contribution of cement. For well D, which was not cored, the lithofacies was determined by the above logging discrimination method. The low-cementation reference response was read from the reference response library according to the lithofacies obtained by discrimination. The component parameters were taken from the statistical values of the reference sample of that lithofacies in the reference response library. The volume fractions of particles, matrix, and pore fluid were substituted according to the mean values of the reference samples. The pore fluid properties were uniformly taken according to the region. It is not necessary to obtain core experimental data again in well D.
[0092] For example, sample A-01 is a grainy limestone. Its measured responses, low-cementation background responses, and differences in density, PE, and some elemental parameters are shown in Table 15. The density background value is... , and the measured value Corresponding to The residuals; the remaining parameters are treated according to the same item-by-item correspondence.
[0093] Table 15 Background response and cement-sensitive residuals of sample A-01
[0094] The aforementioned difference provides information on response variations beyond the same lithofacies background, highlighting sensitive changes related to cement. The residuals and measured logging parameters participate in the prediction, allowing the model to utilize both the overall response and relative background changes. This difference does not directly replace cement content; the content is still determined by subsequent model predictions and constraint processing.
[0095] Taking sample A-01 (granular limestone) as an example, the formation process of the background values in Table 15 is explained: Seventeen low-cementation samples with a total cement content of no more than 5% were selected from the granular limestone samples of well A. The representative logging values of each sample were extracted according to the response kernel method in Example 4. The median was taken as the low-cementation reference response of the lithofacies, and the following values were obtained: DEN 2.69 g / cm³, PE 5.05 b / e, Ca 40.0%, Mg 3.0%, Si 1.3%, Fe 0.6%, and S 0.4%, as listed in Table 15. Subtracting each item from the measured response of A-01, the density residual was obtained as +0.02 g / cm³, the PE residual as +0.13 b / e, and the residuals of Ca, Mg, Si, Fe, and S as +2.6, +0.8, +0.1, +0.1, and +0.1 percentage points, respectively. After the residual and the measured response were input into the model together, the absolute value of the deviation between the joint predicted content of A-01 and the experimental calibration value did not exceed 0.30 percentage points (see Table 18), indicating that the background separation highlighted the response changes related to cement.
[0096] Example 7 In the joint prediction, the cementitious material prediction model employs a random forest regression algorithm to predict the total cementitious material content and the composition ratios of five types of cementitious materials: calcite, dolomite, silica, ferrous, and gypsum. The predicted values of each type of non-negative composition ratio are divided by their sum to obtain normalized composition ratios with a sum of 1. The total cementitious material content is then multiplied by each of these normalized composition ratios to obtain the corresponding cementitious material content.
[0097] The target values used for content modeling are derived from the cementitious occurrence decoupling calibration results. The well-logging input uses parameters that have undergone depth, response scale matching, and standardization, supplemented with corresponding cementitious-sensitive residuals. Various contents satisfy the following conditions when recorded using a fractional scale: ,in This represents the relative content of type i cementitious material; in the data tables of this example, the content is presented as a percentage. In the joint prediction stage of total amount and composition, the correspondence between the content of each type and the total amount is established through proportional normalization. The main cementitious material types are determined according to the one with the highest content after forward modeling correction.
[0098] The model training objective consists of two parts: the first is the total cementitious content objective, which is the sum of the labeled contents of each type of cementitious component; the second is the proportion objective of the five components, which is the ratio of the labeled contents of each component to the total content. The two parts are used to train a random forest regression model separately, with the training sample weights arranged according to... Calculation, where The label confidence level is calculated as follows: (the total content target is the average of the confidence levels of the five label categories, and the composition ratio target is the confidence level of the corresponding label category). The matching weights obtained from depth and response scale matching are shown in Table 9. These weights are applied to both the Bootstrap sampling and the weighted mean square error calculation for node splitting, ensuring that samples with high calibration quality and reliable depth matching play a greater role in tree structure formation. When the total cementitious content of a sample is zero, its composition ratio is undefined, and the sample only participates in the training of the total content target. When the sum of the predicted values of the five composition ratios at a certain depth after non-negation processing is zero, normalization is not performed, and the cementitious content of each category at that depth is set to zero, and a no-cementation label is output.
[0099] The data for modeling and validation consisted of 292 core samples, with some cement content labels shown in Table 16. This content data was used to train the total tissue content and composition, and also for comparative evaluation of the direct content regression described later.
[0100] Table 16. Labels showing cementitious content of some samples.
[0101] The random forest regression model consists of multiple regression decision trees, and the corresponding prediction is obtained by averaging the outputs of each tree. The regression parameter settings in this example are shown in Table 17, where m is the number of input parameters.
[0102] Table 17 Random Forest Regression Parameter Settings
[0103] The total number of samples in this example is Training and validation samples are allocated approximately 70% and 30% respectively, with the integer sample size denoted as . and ,in, The number of training samples. To verify the sample size, the allocation uses 204 and 88 groups (rounded down), and the verification samples are not used in the corresponding model training.
[0104] The proportions and contents of various components obtained from the combined total and composition prediction of A-01 are shown in Table 18. The proportions and contents of the components in the table are displayed with the number of decimal places listed, and the total contents shown for each item are 25.81%.
[0105] Table 18. Joint Prediction Results of Total Amount and Composition of Sample A-01
[0106] In this example, the absolute values of the deviations of the jointly predicted contents from the experimental calibration values do not exceed 0.30 percentage points. The composition ratio normalization constraint applies to the joint prediction stage, while the subsequent content correction based on the forward modeling bias follows the non-negativity and upper limit constraints for each content category.
[0107] To analyze the contribution of well logging parameters to the model, the importance of random forest features can be evaluated and calculated. In the formula, Let T represent the importance of the j-th input parameter, and T be the number of decision trees. To determine the split contribution of the j-th parameter in the t-th decision tree, we take the sum of the weighted mean squared error decreases caused by all split nodes of that parameter within the tree: For each node in the t-th tree with parameter j as the split variable, we sum the differences between the weighted mean squared error of the parent node and the weighted average of the weighted mean squared errors of the left and right child nodes, with node samples weighted according to their sample weights; the importance of each parameter is determined by... The calculation involves dividing by the sum of the importance of all parameters involved in the importance evaluation and multiplying by 100% to obtain the contribution rate as a percentage. After sorting by contribution rate, variables below a set threshold can be deleted during parameter optimization. In this example, the contribution rate is recorded as a percentage value, with the preferred retention condition being... That is, a high contribution parameter with a contribution rate exceeding 5%.
[0108] Table 19 Ranking of Well Logging Parameter Contribution Rate
[0109] According to Table 19, PE, Mg / Ca, DEN, Ca, Mg, AC, Si / Ca, and Fe / Ca are the main prediction parameters. Parameter selection is used for model optimization. In actual application, the input parameter combinations determined during model establishment are used to maintain the input correspondence.
[0110] To observe the effect of joint prediction on the consistency between type and content, this example also sets up random forest classification and direct regression of each type's content as controls. The input, mathematical expression, and calculation records of the control model are described below, with the classification voting results serving as auxiliary type information; the final main cementing type of the target well section is still determined by the content ranking after constraint correction.
[0111] The classification comparison uses a Random Forest Classification (RFC) model, which involves voting by multiple classification decision trees, and is expressed as follows: ,in, This represents the mode operation, which selects the category that appears most frequently from the categories given by each classification tree. In the formula, Y is the classification comparison output, and X is the well logging input. Let be the i-th decision tree, and T be the number of decision trees. The basic logging combination is: The type set is represented as ,in , , , and These correspond to calcite, dolomite, silica, ferrous and gypsum-based cements, respectively.
[0112] In the specific input configuration for classification comparisons, from Starting from the 12 logging parameters included, additional , , and The four ratios constitute The table shows 16 input variables; the five experimental types are labeled 1 to 5 respectively. Classification parameters and sample allocation are shown in Table 20.
[0113] Table 20 Random Forest Classification Parameters
[0114] The classification control group uses the same total sample size. According to the rounded-up Groups and The groups were divided into training and test sets, with a ratio of approximately 70% and 30%, respectively. The partially standardized input of A-01 is shown in Table 14, and the voting records in Table 21 are obtained after feeding it into 500 classification trees.
[0115] Table 21 Voting Results for Sample A-01 Classification Comparison
[0116] Classification tree type support ratio Calculation, where Let i be the classification probability of the i-th type of cementitious material. To support the number of decision trees for this type, Let be the total number of trees. The probability outputs are denoted as follows: It is used to store classification comparison information and is not used as an independent final type criterion for total quantity and composition model.
[0117] Substitute A-01 ,in Get the corresponding number of votes Take 500, and get , , , and .because The classification comparison indicates that the type is calcite cement.
[0118] The logging input for direct content regression control is divided into The conventional logging parameters shown are... The element parameters shown are supplemented. , , and Then, the input matrix is CNL is the compensated neutron parameter, and RT is the deep lateral resistivity parameter. Five content regression relationships were established in the control group: , , , and ; , , , and The values represent the direct regression content of the five types of cementing materials, in that order. to These represent the regression mappings of the contents of calcite, dolomite, silica, ferrous and gypsum-based cements formed during the control training, respectively.
[0119] The regression analysis used a random forest regression (RFR) model, with the predicted values for each tree calculated as follows: On average, of which Let T be the prediction content of the i-th class, and T be the number of decision trees. This represents the output of the t-th regression tree. Taking A-01 calcite cement as an example, some single-tree results are shown in Table 22.
[0120] Table 22 Single-tree prediction records of direct regression for calcite cementitious materials
[0121] In this single-sample calculation, the average of the 500 regression trees is calculated as follows: Obtained, among which This represents the predicted value for the t-th tree. The subscript here is used for tree numbering and is distinguished from the symbol for the total amount of cementitious material according to the formula it appears in.
[0122] The total number of records output by this set of trees is known to be ,but The direct regression value for A-01 calcite cement was 26.40%. Other types were processed using the same calculation method, and the average output for each type is shown in Table 23.
[0123] Table 23 Direct regression results of various cementitious materials in sample A-01
[0124] Direct regression summation of various contents When the sum of all the items is used as the total for proportional conversion, the result is as follows.
[0125] according to calculate, This is the content after proportional conversion. This represents the total content of the five direct regression categories in the sample. For example... and The total values used in the proportional conversion here are derived from the sum of various direct regression values, and the values of each component remain unchanged after the conversion.
[0126] Table 24 Deviation between direct regression and experimental calibration for sample A-01
[0127] Compared with Table 18, the bias of calcite cement in the direct regression control was 8.04 percentage points, while the corresponding bias in the joint prediction was... 0.30 percentage points. This example illustrates that joint prediction of background residuals with total amount and composition can improve the content correspondence of this sample; type consistency is achieved through the unified ranking of the contents after final processing.
[0128] Example 8 In the forward modeling correction phase, component parameters include the volume fractions of particles, matrix, and pore fluids, as well as the densities, photoelectric absorption cross sections, and elemental weight fractions of particles, matrix, pore fluids, and various cements. The upper limits of the content of various cements are determined based on total mineral content, elemental mass balance, and fillable porosity. The density, photoelectric absorption cross section, and elemental logging responses are reconstructed using the predicted contents and component parameters. The reconstructed values are compared with measured values at the same depth and with the same parameters. When the absolute value of the deviation exceeds the corresponding threshold, correction is performed based on the constraint that the content is non-negative and does not exceed the upper limits of each category; otherwise, the original predicted contents are retained.
[0129] Forward modeling treats various cementing materials, particles, matrix, and pore fluids as components. Density is calculated by adding the products of each component's volume fraction and density. Photoelectric absorption cross-section and elemental weight fraction are weighted according to the corresponding component densities. The reconstruction equations are as follows:
[0130]
[0131]
[0132] In the formula, The volume fraction of the i-th type of component, which includes particles, matrix, pore fluid and various cements; and These represent the density and photoelectric absorption cross section of the corresponding components, respectively. This represents the weight fraction of the k-th element in the corresponding component. The reconstructed density, photoelectric absorption cross section, and elemental response are compared with the measured DEN, PE, and corresponding elemental logging parameters.
[0133] In the above reconstruction formula, Pe is the photoelectric absorption index of the component (unit b / e, i.e., bar / electron), which characterizes the photoelectric absorption capacity corresponding to a unit number of electrons. It is an inherent parameter of the component and can be obtained from the standard mineral value after calibration, or it can be calibrated from the core measurement value. The applicable conditions for reconstructing the photoelectric absorption index and element weight fraction in the form of density-weighted volume average are: the sum of the volume fractions of each component is 1 (volume closure), and each component is uniformly mixed within the vertical detection range of the well logging. At this time, the density is linearly weighted according to the volume fraction, and the photoelectric absorption index and element weight fraction are normalized according to the component density weight. The joint prediction reconstruction of A-01 is PE of 5.16b / e and DEN of 2.71g / cm³, which is compatible with the measured 5.18b / e, 2.71g / cm³ and the allowable deviation thresholds of ±0.10b / e and ±0.03g / cm³ (see Table 25), indicating that this reconstruction form is compatible with the well logging records listed in Table 25.
[0134] The volume fractions of various cements required for forward modeling are obtained by converting the calibrated content according to the diameter: the area fraction of thin sections is approximated as the volume fraction according to the Delesse principle, and the calibrated content of the mass fraction diameter is converted into the volume fraction according to the corresponding mineral density. After conversion, the sum of the volume fractions of particles, matrix, pore fluid, and various cements is 1, satisfying the component total closure. For non-core sections, the volume fractions of particles, matrix, and pore fluid are determined by porosity logging and statistical values of lithofacies reference samples. The upper limits of the content of various cements are obtained as follows: ① The upper limit of total mineral content is the whole-rock mass fraction of the corresponding mineral measured by XRD (for non-core sections, the statistical value of the same lithofacies reference sample is used); ② The upper limit of elemental mass balance is the Fe and S content obtained by elemental logging or core elemental analysis divided by the weight fraction of the element in the corresponding cement; ③ The upper limit of fillable porosity is the original porosity estimated based on sedimentary facies and adjacent well data, and the total amount of cement must not exceed this value.
[0135] The update and correction of the content of various cementitious materials after the deviation exceeds the limit is performed as follows: Calculate the adjustment amount according to the deviation direction of each parameter and the sensitivity of each component to the corresponding parameter. Under the constraints that the content of various cementitious materials is non-negative, does not exceed the corresponding upper limit, and the sum of the volume fractions of each component remains at 1, update the content of various cementitious materials and re-perform forward modeling reconstruction according to the updated content. If the absolute value of the deviation of each parameter after the update falls into the corresponding threshold, the correction ends, and the updated content is used as the processed content of various cementitious materials. If there are still parameters exceeding the limit, repeat the above update process until all parameters meet the threshold or the maximum number of corrections is reached (3 times in this example). If the maximum number of corrections is reached but not all are met, take the content with the smallest weighted deviation as the processing result and mark the correction as not fully converged.
[0136] Table 25 shows the forward modeling comparisons for A-01, which also lists the reconstructed responses of the direct regression control and the joint prediction content. The allowable bias is defined by the corresponding thresholds for each parameter; the values for density, PE, Ca, and Mg are as shown in the table.
[0137] Table 25 Forward Reconstruction and Bias Evaluation of Sample A-01
[0138] The biases of DEN, PE, Ca, and Mg in the direct regression control were respectively The values of +0.24b / e, +5.2 percentage points, and +1.1 percentage points all exceed the corresponding thresholds. Therefore, this group of contents cannot be used as the final output without constraint correction. The corresponding biases for the joint prediction are 0, ... 0.02b / e 0.3 percentage points and The absolute value of 0.1 percentage points did not exceed the limit, so this example enters the branch that retains the predicted content.
[0139] After this processing, A-01 retains the jointly predicted contents listed in Table 18. The highest content is calcite cement, and its main cement type is accordingly determined to be calcite. If the forward modeling deviation of the target depth exceeds the limit, the deviation is used as the correction basis. After processing within the non-negative and corresponding upper limit range of various contents, the type is determined based on the maximum content after processing.
[0140] The above-mentioned sample A-01 enters the retention branch. To illustrate the correction process when the deviation exceeds the limit, sample A-04 (4223.45m) is used as an example. The initial joint prediction of the sample's content was 38.5% calcite, 9.6% dolomite, 2.3% silica, 7.2% iron, and 0.6% gypsum salts, totaling 58.2%. Forward modeling using the aforementioned reconstruction formula yielded DEN 2.82 g / cm³, PE 5.58 b / e, Ca 46.8%, Mg 5.0%, and Fe 5.6%. Compared to the measured values of 2.72 g / cm³, 5.26 b / e, 41.5%, 4.1%, and 3.2%, the deviations were +0.10 g / cm³, +0.32 b / e, +5.3, +0.9, and +2.4 percentage points, respectively. The absolute values all exceeded the corresponding thresholds of ±0.03 g / cm³, ±0.10 b / e, ±1.5, ±0.5, and ±0.5 percentage points, triggering constraint correction. The upper limits for the content of various cementing materials in A-04 are determined to be 60.5% calcite, 12.0% dolomite, 3.5% silica, 4.6% iron, and 2.1% gypsum salts. The upper limit for iron is obtained based on elemental mass balance, i.e., the measured Fe content of 3.2% divided by the weight fraction of Fe in the iron cementing material (0.70). The initially predicted iron content of 7.2% exceeded this upper limit. After adjustment according to the aforementioned update method, the contents of each type are 36.9% calcite, 8.3% dolomite, 1.5% silica, 4.3% iron, and 0.5% gypsum salts, totaling 51.5%. The correction results for each parameter are shown in Table 26.
[0141] Table 26 Constraint Correction Process When Sample A-04 Exceeds the Limit
[0142] After one update, all parameter deviations in this example fell within the threshold, and the correction was complete. The absolute values of the deviations between the corrected cementitious contents and the experimental calibration values were 0.15, 0.06, 0.08, 0.04, and 0.05 percentage points, respectively. The main cementing material type remained calcite, consistent with experimental findings. If the initial predicted content had been directly output, the total cementing content would have been approximately 7.0 percentage points higher than the experimental calibration value. This example demonstrates that constraint correction can update out-of-limit predicted contents to values consistent with the measured logging response and meeting the upper limit of the content.
[0143] Example 9 In result evaluation and output, the prediction confidence is determined using depth matching confidence, model divergence, and the forward modeling residual between the well logging reconstruction response and the measured response. The cementation strength is determined based on the sum of various contents after constraint correction and preset grade conditions. When the prediction confidence is not lower than the first threshold, the main cementing material type, individual values of various contents, and cementation strength grade are output. When it is not lower than the second threshold but lower than the first threshold, the cementing material content range is output. When it is lower than the second threshold, no deterministic type or content is output; instead, a rejection flag and corresponding well section information are output, where the second threshold is lower than the first threshold.
[0144] Prediction confidence is synthesized from three indices: deep matching confidence, model divergence, and multi-parameter forward modeling residuals, using the following formula: Where R is the prediction confidence level. For deep matching confidence, For model divergence, For multi-parameter forward modeling residual index, , , For the composite weights, we use 0.4, 0.3, and 0.3 respectively in this example, and truncate R in the range of 0 to 1.
[0145] Deep matching confidence According to the correlation analysis in Example 4, the correlation coefficient between the core sample reference sequence and the measured logging curves in the cored well section was determined. For Well D, which was not cored, since there was no core sample reference sequence, the lithofacies-logging response template established by the adjacent cored well (Well C) was used to cross-correlate with the natural gamma and sonic transit time curves of the same layer in Well D, and the correlation coefficient was taken as the correlation coefficient. The consistency of the curves from the repeated measurement segments of well D was used for verification. In this example, the curves at each depth point of well D were verified. Between 0.80 and 0.95. Model divergence. Quantification based on the degree of dispersion of the predicted values from each decision tree: ,in The standard deviation of the predicted total cementitious content for each decision tree is given. The standard deviation of the total cement content in the training samples was used as the normalization benchmark. A smaller value indicates more consistent output across all trees. Multi-parameter forward modeling residuals. Quantified by the average of the normalized deviations of each comparison parameter: ,in and These are the reconstructed response and the measured response for the k-th parameter, respectively. The allowable deviation threshold for this parameter is shown in Table 25, and K is the number of parameters involved in the comparison. A value less than 1 indicates that the average of the parameters being compared is within the limit.
[0146] Bond strength is classified based on the total content of various treatments, which can be expressed as: In the case of the five types of cementitious materials, the specific range of summation is as follows:
[0147] In the formula, The total cement content after constraint correction treatment at the same depth. For this depth The content of cementitious material after treatment is calculated by adding the five sub-items together under a consistent content measurement standard.
[0148] Table 27 Cementation Strength Grade and Reservoir Impact
[0149] The total content of various types retained by A-01 is 25.81%, and a medium cementation grade is obtained according to Table 27. This, together with the main types and the content of various types, forms the evaluation result.
[0150] Prediction confidence level reflects the reliability of the output results. For the D well example, 0.8 is used as the first threshold and 0.6 as the second threshold. The proportion and processing methods for each depth level are shown in Table 28. Internally generated content calculation records must be controlled for external output according to this confidence level.
[0151] The first and second thresholds are determined based on the confidence-error correspondence of the verification samples: the predicted confidence of the verification samples is calculated using a composite formula and sorted from high to low. The proportion of samples within each confidence interval whose content error for single-value output does not exceed the corresponding accuracy requirement is statistically analyzed. The lower confidence limit with a proportion of 90% or higher is taken as the first threshold (0.8 in this example), and the lower confidence limit with a proportion of 70% or higher is taken as the second threshold (0.6 in this example). No deterministic result is output when the confidence level is below the second threshold. When outputting intervals, the half-width of the interval for the content of various cementing materials is calculated according to... Confirmed, among which , The root mean square error for the verification of the cementing materials listed in Table 36 is used. The half-width of the interval for the total cementing material content is calculated as 3.65% of the average root mean square error of the verification per well as listed in Table 38. When the confidence level is 0.8, λ is taken as 1.0, and when the confidence level is 0.6, λ is taken as 1.5. The lower the confidence level, the wider the interval. The output interval for each type of content is the point value plus or minus the half-width of the interval. When the lower limit is below 0, it is truncated to 0 according to the non-negative content.
[0152] Table 28 Statistics on the Reliability Grading of Well D Prediction
[0153] Among them, the well section from 4232.0 to 4235.5m crosses a clear lithological interface, the model has a high degree of divergence and the forward modeling residual exceeds the limit, and the prediction confidence is less than 0.6. Therefore, the rejection mark and depth information of this well section are output, but the type and content of deterministic cement are not given, so as to reduce the interference of low confidence results on reservoir evaluation.
[0154] Taking the 4247.00m depth point of well D as an example, this illustrates the output of the confidence interval: the depth matching confidence level at this point. Model divergence Multi-parameter forward modeling residuals According to the composition formula, the confidence level R is obtained as R = 0.4 × 0.88 + 0.3 × (1 0.28)+0.3×(1 The value of 0.45) = 0.733, which falls within the range of 0.6 ≤ R < 0.8, and is output in interval format. The various content point values jointly predicted for this point and the output interval calculated using λ = 1.17 are shown in Table 29.
[0155] Table 29 Examples of confidence interval outputs in well D at 4247.00m
[0156] Compared with the core data of the same layer of the adjacent well C, the total cement content of the corresponding sample was 43.1%, which falls within the output range of 40.0% to 48.6%, indicating that the range output under medium confidence level can cover the actual content variation. Compared with the high confidence level, no single content value is output at this depth point, which is different from the rejection processing of the low confidence level, and realizes the technical effect of controlling the output according to confidence level.
[0157] In the application of non-cored well sections, well D was selected as the application well, and no core samples were taken from its study section. In this stage, the already trained model was called up using continuous logging data; thin section casting, cathodoluminescence, SEM-EDS, and XRD experiments were used for the aforementioned calibration and model establishment, and it was not required to obtain core experimental results again at each predicted depth in well D.
[0158] Well D was studied at depths ranging from 4200 to 4350 m, with a logging sampling interval of 0.125 m. This example used 1200 sets of logging records. The measured parameters and ratio combinations are as follows: ,according to The minimum and maximum values from the training phase are standardized, and the cement-sensitive residuals based on lithofacies structure are used for prediction. For example, the logging input for a depth of 4215.25m is shown in Table 30.
[0159] Table 30 Well logging input at depth 4215.25m in Well D
[0160] After processing at depth along the target well section, the output of various contents is recorded sequentially as follows: This is used to generate content curves that vary with well depth. For each depth, the content is determined, forward modeled and corrected, and a type is generated. Then, based on the confidence level, single-value, interval, or rejection outputs are performed.
[0161] To demonstrate the auxiliary type analysis on the D well data, the parameters of 4215.25m were fed into the classification control model to obtain the voting records of the 500 trees listed in Table 31.
[0162] Table 31 Auxiliary classification voting results at depth 4215.25m in Well D
[0163] Classification probabilities by Calculations yielded the following results: , , , and .because The classification and comparison indicate dolomite cement. This probability record is retained as auxiliary information; the content determination and external output still follow the above-mentioned joint prediction, forward modeling correction, and confidence processing procedures.
[0164] Table 32 shows the direct regression control content records for some depths in Well D. The contents of the five categories at 4215.25m were 12.4%, 34.8%, 1.6%, 0.7%, and 0.3%, respectively, with a total content of [missing information]. ,satisfy The table shows strong cementation conditions, with dolomite cement being the most abundant component in this record. The type probability columns are for auxiliary classification statistics and do not replace the primary type or prediction reliability determined by the processed content.
[0165] Table 32 Direct regression comparison records of non-coring sections of Well D
[0166] In the records listed at 4285.00m in well D, The corresponding result is strong bonding.
[0167] As recorded in the table, the points listed from 4205.00 to 4215.25m are dominated by dolomite cement, while calcite cementation is enhanced from 4220.50 to 4225.75m. The content of siliceous cement increases near 4240.25m. The total cement content at all listed points exceeds 35%, but the overall evaluation of the well section still requires a credibility rating, especially for the aforementioned rejected sections, where no definitive conclusion should be reached.
[0168] Regarding data completeness, the enhanced mode with elemental logging data can be compared with the simplified mode with only conventional logging data. The enhanced mode uses GR, AC, DEN, CNL, RT, PE and elemental logging information to obtain the content of five types of cementitious materials; the simplified mode mainly outputs the total cementitious material content Ct, main cementitious material types and cementation strength based on the changes in parameters such as PE, DEN, and AC, without detailing the complete results of the five types of content.
[0169] In the simplified prediction control, the input is reduced to... Taking well D at 4215.25m as an example, the simplified model record obtained by using the six conventional logging values in Table 30 is shown in Table 33.
[0170] Table 33 Simplified prediction comparison results for well D at 4215.25m
[0171] Records of elemental logging inputs In comparison, the total difference in the simplified forecast is This example demonstrates that conventional logging data can be used for supplementary evaluation of cementation levels; this simplified comparison does not replace the elemental background response and forward modeling consistency processing when relevant data are available.
[0172] The model's effectiveness was validated using core samples not used in the corresponding training, thin section data from adjacent wells, and production test data. The type accuracy was [percentage missing]. ,in To predict the number of correct samples, To verify the total number of samples; the content error was calculated using... ,in To predict the content, denoted as the experimental content, and n as the number of validation samples. The coefficient of determination is also used. The degree of agreement between the predicted and experimental values is evaluated, and the results are assessed using the bond strength compliance rate as the evaluation grade.
[0173] For 88 groups of samples that did not participate in the classification control training, the classification results were compared with the thin-slice experiment identification results. The number of correct types, the number of errors, and the accuracy rate are shown in Table 34.
[0174] Table 34. Verification Results of Cement Type Comparison
[0175] Summarize the correct numbers for each category to obtain This value reflects the performance of the listed classification controls on this set of test samples, and the type output for joint content prediction is determined by the one with the highest processed content.
[0176] Content prediction was performed using 88 validation samples for error analysis. When the sample numbers are explicitly listed, the root mean square error is calculated as follows:
[0177] in Let be the model prediction value for the i-th sample. The values are the corresponding experimental measurements, where n is the number of validation samples, and the subscripts are... and The model predictions and experimental observations are labeled separately. This is the sample number for verification. When calculating the error of a single type of cement, the summation is performed across all verification samples of that type. Taking calcite cement as an example, the experimental values, predicted values, and squared errors of some samples are shown in Table 35. This table only lists some records; the overall index is calculated from all 88 verification samples.
[0178] Table 35 Verification Records of Some Calcite Cementitious Materials
[0179] The overall root mean square error of the calcite cement is calculated as follows:
[0180] The coefficient of determination is used to evaluate the degree of correspondence between predictions and experimental results, and is adopted as follows:
[0181] In the formula, As the coefficient of determination, These are the model's predicted values. These are the corresponding experimentally measured values. The average experimental values of the verification samples for this type of cementitious material constitute the baseline term in the denominator. The sum of the numerator and denominator covers all verification samples of this type of cementitious material, with the denominator reflecting the degree of fluctuation of the experimental values relative to their average value. When all experimental values of this type of verification samples are identical and the denominator is zero, the coefficient of determination is no longer applicable, and the root mean square error (RMSE) is used as the consistency evaluation index for this type. A summary of the coefficients of determination and RMSE for various types of cementitious materials is shown in Table 36, with the RMSE for percentage content recorded according to the corresponding content scale.
[0182] Table 36 Prediction and Evaluation Results of Cementitious Material Content
[0183] The cementation strength verification compares the grade corresponding to the predicted total content with the degree of cementation reflected by the core porosity and permeability tests. The results are shown in Table 37.
[0184] Table 37 Bond strength verification results
[0185] Based on the number of compliant samples in the three categories, the overall compliance rate was [percentage missing]. The verification of type, content, and cementation strength reflects different evaluation objects, while the prediction reliability classification plays a controlling role in the output of the target well section.
[0186] To verify the cross-well application capability, a separate well-based tissue validation was conducted. Each time, one well (A, B, or C) was used as the validation well, and the other two wells were used as training wells, ensuring that samples from the same well were not included in the training and validation sets simultaneously. The results of the well-based validation are shown in Table 38.
[0187] Table 38 Verification results of cross-well allowance based on well location
[0188] The accuracy rate of the average type of well-drilled deposits was 90.6%, and the calculated value for random deposit type verification was... The results of the two partitioning methods were used to observe the impact of sample partitioning on the evaluation; the well-based reservation was used to reduce the overestimation of accuracy caused by the spatial correlation of samples from the same well.
[0189] When the validation error exceeds the set range, model optimization can be performed by increasing training samples, adjusting input parameter combinations, optimizing random forest parameters, and recalculating feature importance. For classification comparisons, the classification probability threshold can also be adjusted. Even after optimization, independent data verification is still required, and the content error and coefficient of determination should maintain the following evaluation criteria:
[0190]
[0191] In the above formulas for root mean square error and coefficient of determination, the denominator n in the RMSE calculation is the number of validation samples. Calculation in progress These are the model's predicted values. These are the corresponding experimentally measured values. The summation of the numerator and denominator is the average of the experimental values. The summation range of the numerator and denominator is all the verification samples of this type of cementitious material. The meaning of the symbols is consistent with the aforementioned content prediction and evaluation section. When the experimental values do not change and the denominator is zero, R² is not applicable, and RMSE is used for evaluation alone.
[0192] By linking experimental calibration, continuous logging prediction, and independent verification, this method extends core point experimental information to well sections with corresponding logging data. Combined with content consistency processing and reliability output, it improves the continuity of cementation evaluation in complex carbonate reservoirs.
[0193] Example 10 like Figure 2As shown, this embodiment also provides a well logging quantitative prediction system for carbonate rock cement type and content, including a data acquisition module, a background processing module, a joint prediction module, a content determination module, a forward modeling correction module, and a result generation module. Figure 2 The connection between modules represents the connection relationship of data processing. Each module completes the processing in sequence based on the measured data of the target well section, the content prediction, and the forward modeling verification.
[0194] The data acquisition module obtains the measured logging response and composition parameters of the target carbonate rock well section, provides the measured data to the subsequent background processing and prediction stages, and provides the composition information required for forward modeling. The background processing module determines the low-cementation reference response based on the corresponding lithofacies, calculates the difference between the measured response and the reference response item by item, and outputs the cement-sensitive residual.
[0195] The joint prediction module calls upon a pre-established cementitious prediction model, combining the measured logging response with cementitious sensitive residuals to predict the total cementitious content and the proportions of various components. The model's calibration is based on cementitious occurrence decoupling labels and logging data matched for depth and response scales. The content determination module receives the total amount and component proportions, generates the content of various cementitious components, and transfers it to the forward modeling correction module.
[0196] The forward modeling correction module reconstructs various contents into logging responses based on component parameters. After comparing these responses with the measured responses, constraint correction is performed according to the deviations to obtain the processed contents. The results generation module reads these contents, uses the type corresponding to the maximum value as the main cementing type, and predicts the cementing type and content according to well depth. This module thus achieves... Figure 1 The data connection method shown is illustrated.
[0197] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.
Claims
1. A method for quantitative prediction of carbonate rock cement type and content through well logging, characterized in that, include: Obtain the measured logging response and component parameters of the target carbonate rock section, determine the low-cementation reference response according to the lithofacies, and construct the cement-sensitive residual based on the difference between the measured logging response and the low-cementation reference response. The measured logging response and cement-sensitive residuals are input into a pre-established cement prediction model to predict the total cement content and the composition ratio of various cements. The cement prediction model is established based on cement occurrence decoupling labels and logging data matched with depth and response scales. The content of each type of cementitious material is determined based on the total cementitious material content and composition ratio; The logging response is reconstructed using the contents and component parameters of the various cementing materials in a forward model to obtain the reconstructed logging response. Based on the deviation between the reconstructed logging response and the measured logging response, the contents of various cementing materials are constrained and corrected to obtain the contents of various cementing materials after processing. The type corresponding to the maximum value among the various cementitious contents after the treatment is taken as the main cementitious type, and the cementitious type and content prediction results along the well depth are formed.
2. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The construction of the cement occurrence decoupling tag includes: acquiring cast thin section images, cathodoluminescence images, and SEM-EDS elemental maps of the same core sample from the same field of view, and performing spatial registration; classifying the same mineral into four occurrence types—grains, matrix, replacement products, and cement—based on the mineral positions, pore filling relationships, crystal boundaries, grain contacts, and luminescent bands in the registered images; constructing a cement-specific mask based on the cement occurrence regions, and only statistically analyzing the area and elemental content of each region within the cement-specific mask; determining the cement type of each region based on the elemental content, and determining the corresponding content by the ratio of the area of each cement region to the effective statistical area of the thin section, thereby generating the cement occurrence decoupling tag.
3. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 2, characterized in that, The establishment of the cementitious material prediction model includes: obtaining the overall mineral composition information of the core sample through X-ray diffraction; applying an upper limit constraint or a mineral quantity balance constraint to the cementitious material content corresponding to the cementitious material occurrence decoupling tag based on the corresponding total mineral content in the overall mineral composition information; generating the tag confidence score of the cementitious material occurrence decoupling tag based on the consistency of the results of the core sample's cast thin section, cathodoluminescence, SEM-EDS, and X-ray diffraction experiments; and weighting the corresponding calibration samples in the establishment of the cementitious material prediction model based on the tag confidence score.
4. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The depth and response scale matching includes: performing correlation analysis on the experimental reference sequence of the core sample and the measured logging curves of the corresponding well section to determine the core depth offset and matching confidence level; correcting the sample depth based on the core depth offset; setting a corresponding longitudinal response kernel according to the longitudinal resolution of each logging instrument; selecting a matching window near the corrected sample depth that does not cross obvious lithological interfaces; convolving the longitudinal response kernel with the logging curves within the matching window to obtain a representative logging value for the corresponding sample depth; using the representative logging value as the logging input when establishing the cement prediction model; and reducing the weight of samples with low matching confidence based on the matching confidence level.
5. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The measured logging responses include logging parameters for natural gamma, sonic transit time, density, compensated neutrons, deep lateral resistivity, and photoelectric absorption cross section, as well as elemental logging parameters for Ca, Mg, Si, Fe, S, and Sr. Based on the elemental logging parameters at the same depth, Mg / Ca, Si / Ca, Fe / Ca, and S / Ca elemental ratio parameters are constructed. The measured logging responses and elemental ratio parameters are combined according to the input parameters of the cementation prediction model to construct a logging input matrix, and standardized based on the minimum and maximum values of the corresponding parameters in the training samples. The standardized well logging input matrix and the cement-sensitive residual are input together into the cement prediction model.
6. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The low-cementation reference response includes the density, photoelectric absorption cross section, and elemental background response under the combined influence of particles, matrix, pores, and fluids in the corresponding lithofacies. The corresponding low-cementation reference response is determined according to the lithofacies at the target depth location. The measured logging response at the target depth location is matched with the low-cementation reference response item by item according to the logging parameters. The cementation sensitive residual is constructed by subtracting the corresponding background response from the measured logging response. The cement-sensitive residual is used as a supplementary variable and input into the cement prediction model along with the measured logging response at the target depth.
7. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The cementitious material prediction model employs a random forest regression algorithm to predict the total cementitious material content and the composition ratios of five types of cementitious materials: calcite, dolomite, silica, ferrous, and gypsum. The predicted non-negative composition ratios of the five types of cementitious materials are divided by their sum to obtain normalized composition ratios with a sum of 1. The total cementitious material content is then multiplied by the normalized composition ratios of each type of cementitious material to obtain the corresponding content of each type of cementitious material.
8. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, The component parameters include the volume fractions of particles, matrix, and pore fluid, as well as the densities, photoelectric absorption cross sections, and elemental weight fractions of particles, matrix, pore fluid, and various cements. The constraint correction includes: determining the upper limit of the content of various cements based on the total mineral content, elemental mass balance, and fillable porosity; reconstructing the logging response of density, photoelectric absorption cross section, and elements using the content of various cements and component parameters to obtain the logging reconstruction response; comparing the logging reconstruction response with the measured logging response at the same depth and with the same logging parameters to obtain the corresponding deviation; when the absolute value of the deviation exceeds the corresponding set threshold, correcting the content of various cements based on the deviation, under the constraint that the content of various cements is non-negative and does not exceed the corresponding upper limit; otherwise, retaining the original predicted content.
9. The method for quantitative prediction of carbonate rock cement type and content by well logging according to claim 1, characterized in that, Also includes: The prediction confidence is determined by using the depth matching confidence, model divergence, and the forward modeling residual between the reconstructed logging response and the measured logging response; the cementation strength level is determined based on the sum of the contents of various cementing materials after correction and the preset level classification conditions. When the prediction confidence is not lower than the first threshold, the main cementing material type, the single value of the content of each type of cementing material after processing, and the cementing strength level are output; when the prediction confidence is not lower than the second threshold but lower than the first threshold, the cementing material content range is output; when the prediction confidence is lower than the second threshold, the deterministic cementing material type and content result are not output, and the rejection mark and corresponding well section information are output; the second threshold is lower than the first threshold.
10. A quantitative prediction system for the type and content of carbonate rock cements through well logging, characterized in that, include: The data acquisition module is used to acquire the measured logging response and composition parameters of the target carbonate rock well section; The background processing module is used to determine the low-cementation reference response according to the lithofacies, and to construct the cementation sensitive residual based on the difference between the measured logging response and the low-cementation reference response. The joint prediction module is used to input the measured well logging response and cement-sensitive residuals into a pre-established cement prediction model to predict the total cement content and the composition ratio of various cements; the cement prediction model is established based on cement occurrence decoupling labels and well logging data matched with depth and response scales; The content determination module is used to determine the content of various cementitious substances based on the total cementitious substance content and composition ratio. The forward modeling correction module is used to reconstruct the logging response using the contents and component parameters of the various cementing materials to obtain the reconstructed logging response. Based on the deviation between the reconstructed logging response and the measured logging response, the content of various cementing materials is constrained and corrected to obtain the content of various cementing materials after processing. The result generation module is used to take the type corresponding to the maximum value among the various cement contents after processing as the main cement type, and to generate cement type and content prediction results along the well depth.
