Logging curve adaptive clustering ultrafine lithology classification method for enhancing carbonate rock-clastic rock transition type lithology identification

The adaptive clustering lithology classification method based on well logging curves solves the problems of ambiguous lithology naming, poor curve adaptability, and vertical discontinuity in clustering results. It achieves high-precision lithology classification of complex strata and improves the accuracy and continuity of lithology identification.

CN121479387APending Publication Date: 2026-02-06SHANDONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511581456.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-31
Publication Date
2026-02-06

AI Technical Summary

Technical Problem

Existing lithological classification methods suffer from problems such as ambiguous lithological naming, poor curve fit, reliance on experience for clustering parameters, and vertical discontinuity in clustering results when identifying transitional types of carbonate rocks and clastic rocks. These issues make it difficult to meet the high-precision lithological classification requirements of complex strata.

Method used

An ultra-refined lithology classification method based on well logging curve adaptive clustering was adopted. UMAP adaptive dimensionality reduction, HDBSCAN clustering and parameter adaptive recommendation were used to evaluate the clustering quality by combining Silhouette profile coefficient, Calinski-Harabasz index and Davies-Bouldin index, so as to achieve ultra-refined and quantitative lithology naming. GR, AC, RT, DEN and CNL curves were used for lithology identification, and vertical lithology profiles were drawn by using rainbow color patterns.

Benefits of technology

It improves the vertical continuity and naming accuracy of lithological classification, reduces the cross-layer misjudgment rate, and realizes four-dimensional geological naming (sedimentary color + composition ratio + grain size distribution + basic lithology), meeting the high-precision lithological classification requirements of complex strata.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121479387A_ABST
    Figure CN121479387A_ABST
Patent Text Reader

Abstract

The invention discloses a logging curve adaptive clustering ultrafine lithology classification method for enhancing carbonate rock-clastic rock transition type lithology identification, and relates to the technical field of lithology classification. The method comprises the steps that after logging data are loaded, missing values are processed firstly, and samples with the non-null-value proportion larger than or equal to 80% are reserved; preprocessing the well logging data after completion; screening the geologic features, and removing geologic redundant information; the global structure and local lithology details of the sedimentary system are reserved through UMAP self-adaptive dimensionality reduction; hDBSCAN clustering parameters are determined in a self-adaptive mode based on the sedimentary body scale and the lithologic continuity, and quantitative evaluation and early warning are carried out on the clustering quality; and finally, based on traditional color matching of sedimentary geology, lithology-color geological mapping is established, and lithology distribution of sequence cycle is visually reflected. The method can meet the high-precision lithology classification requirements of complex stratums (such as heterogeneity, curve missing and multiple transition type lithology).
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithology classification, and particularly relates to a well logging curve self-adaptive clustering super-fine lithology classification method for strengthening carbonate-clastic transition type lithology identification. BACKGROUND

[0002] Lithology classification is a core link of transformation from "logging data" to "sequence geology cognition" in oil geological exploration, and its precision directly determines the reliability of reservoir parameter calculation (porosity, permeability), system domain division and oil and gas resource prediction under the sequence framework. With the exploration target shifting from homogeneous and whole reservoir to non-homogeneous and complex reservoir under the control of sequence (such as deep layer of continental fault basin and carbonate platform), the technical defects of the traditional lithology classification method are increasingly prominent, and the limitations of the existing technology are as follows in combination with the industry practice and sequence stratigraphy theory:

[0003] First, the lithology naming is ambiguous, and the transitional lithology identification in sequence framework is difficult: it can only distinguish basic categories such as "sandstone / mudstone / limestone", and cannot match the lithology gradual change rule of sequence cycle: (1) The qualitative description method relies on core observation, and is highly subjective, and cannot quantify the mud content, particle size distribution and carbonate rock specific characteristics (dolomite content, compactness) in the para-sequence, which deviates from the true lithology in the identification of carbonate rock transitional types (muddy dolomite, calcareous dolomite), especially near the sequence boundary (such as the transitional zone of mudstone and sandstone near the maximum flooding surface), the misjudgment rate increases significantly; (2) The semi-quantitative crossplot method only divides the lithology through the simple crossplot of GR-AC, GR-RT, without using DEN (density), CNL (neutron) and other key curves for carbonate rock identification, which cannot distinguish between compact limestone and pure limestone, and dolomitic limestone and dolomitic limestone. In addition, the comprehensive information of multiple curves is not fully analyzed, such as relying only on GR to judge, which cannot identify the particle size or gravel / carbonate content. Second, the missing curve has poor adaptability, and the thin layer lithology information in the sequence is easily ignored. In actual exploration, the core curves such as GR, AC, RT, DEN and CNL are often missing (instrument failure, wellbore condition limitation), and the existing processing method has obvious defects: taking the mean / median of the whole well section to fill in the missing values, which can cover up the information of lithology gradual change caused by base level rise; directly discarding the depth section with missing data, which can ignore the lithology information of the sedimentary body with thin sedimentary thickness (such as lowstand system tract thin sand body) in the sequence. Third, the existing machine learning model for lithology classification relies on experience and has poor sequence adaptability. The existing machine learning means for lithology classification often needs parameter tuning: K-means clustering needs to pre-set the cluster number, while the actual stratigraphic lithology class varies with the sequence system tract (such as 5-8 classes for fluvial facies, 3-5 classes for lacustrine facies, and 6-10 classes for carbonate platform facies), and the deviation of the pre-set value leads to "over-clustering" (splitting dolomite into 2 classes) or "under-clustering" (merging calcareous dolomite and dolomitic limestone); the eps (neighborhood radius) and min_samples of traditional DBSCAN need to be optimized manually, which is time-consuming and laborious; there is no quantitative evaluation of clustering quality, and noise points (such as wellbore expansion section) are easily misjudged as different lithology classes, showing "false clustering" of lithology. Fourth, the lithology classification results are not continuous vertically, which violates the sequence stratigraphic law.

[0004] Therefore, the prior art needs to be further improved. SUMMARY

[0005] The purpose of the present application is to provide a well logging curve self-adaptive clustering ultra-fine lithology classification method for strengthening the identification of carbonate rock-clastic rock transitional lithology, which can meet the high-precision lithology classification needs of complex strata (heterogeneous, curve missing, multiple transitional lithology).

[0006] To achieve the above object, the application adopts the following technical solutions:

[0007] A well logging curve self-adaptive clustering super-fine lithology classification method for strengthening the identification of carbonate rock-clastic rock transition type lithology, comprising the following steps:

[0008] S1, load the logging data and process the missing values, and retain samples with a non-empty value proportion of 80% or more;

[0009] S2, pre-process the completed logging data;

[0010] S3, filter the geological features and remove the geological redundant information; retain the global structure and local lithology details of the sedimentary system through UMAP self-adaptive dimension reduction;

[0011] S4, HDBSCAN clustering and parameter self-adaptive recommendation, comprising the following sub-steps:

[0012] S41, adaptively determine the HDBSCAN clustering parameters based on "sedimentary body size" and "lithology continuity", use the EOM lithology classification method of HDBSCAN, preferentially retain the transition cluster of density-continuous carbonate rock, and match the continuity of sedimentary evolution; the EOM algorithm automatically retains the transition cluster of "limestone→dolomitic limestone→limestone dolomite→dolomite" of carbonate rock by identifying "density peak continuous feature area"; the clustering parameters: the minimum cluster size is adaptively determined based on the data amount after dimension reduction, the ratio of the total number of samples to 200 is rounded, and the lower limit is 5; the minimum sample number is half of the minimum cluster size rounded, and the lower limit is 3; the distance measurement of different lithology classification adopts the Euclidean distance, which adapts to the numerical characteristic distribution law of logging data;

[0013] S42, quantitative evaluation and early warning of clustering quality:

[0014] The clustering quality is evaluated by Silhouette coefficient, Calinski-Harabasz index and Davies-Bouldin index, and early warning is triggered when the noise point proportion is > 30% or the silhouette coefficient is < 0.3;

[0015] S5, super-fine lithology naming, naming through hierarchical and quantile number and geological classification accurate correspondence, the steps are: based on hierarchical and quantile number, divided into five scenarios, respectively: "GR+AC+RT+DEN+CNL complete", "GR+AC+DEN missing RT", "GR+RT+CNL missing AC", "only GR", "only AC+RT+DEN"; realize the super-fine naming of "grain size distribution+component proportion+basic lithology+carbonate rock subclass"; among them, carbonate rock is judged by DEN>2.6g / cm³ and CNL<15%, and further subdivided into dolomite, calcareous dolomite and dolomitic limestone; when the cluster mean exceeds 3 times the standard deviation or the curve is suddenly changed, the abnormal lithology is named;

[0016] S6, based on traditional sedimentary geology color matching, establish lithology-color geology mapping, intuitively reflect the lithology distribution of sequence cycle, when cross-well comparison, only need to match the color to quickly identify the same type of lithology.

[0017] The well logging curve self-adaptive clustering super-fine lithology classification method for strengthening carbonate rock-clastic rock transition type lithology identification, in step S1, the core curves of GR, AC, RT, DEN and CNL are identified through the alias mapping table, wherein the GR curve is used for quantitative judgment of argillaceous content in the quasi-sequence, the AC curve is used for lithology grain size distribution division, the RT curve is used for gravel content and formation density judgment; the DEN curve is used for carbonate rock lithology subdivision; the CNL curve is used for auxiliary judgment of carbonate rock porosity; the alias mapping table includes: GR alias "GR, natural gamma, gamma", AC alias "AC, acoustic time difference, acoustic wave", RT alias "RILD, deep lateral, LLD, R25, R4, RT, resistivity", DEN alias "DEN, density, bulk density", CNL alias "CNL, neutron, compensated neutron", compatible with English curve names in logging industry.

[0018] In the well logging curve self-adaptive clustering super-fine lithology classification method for strengthening carbonate rock-clastic rock transition type lithology identification, when the missing values are processed in step S1, first, the non-empty value proportion ≥80% of the samples are retained through row screening, and then the missing values are completed by "sliding average with window size of 5+linear interpolation".

[0019] In the well logging curve self-adaptive clustering super-fine lithology classification method for strengthening carbonate rock-clastic rock transition type lithology identification, in step S2, the completed data is subjected to "adaptive sliding average filtering", which provides a data basis of "high signal-to-noise ratio+high geological recognition degree" for subsequent lithology classification, the window is default 7 and is an odd number, and the window is dynamically adjusted to 5 in the heterogeneous formation.

[0020] The well logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic rock transition type lithology, in step S2, the well logging geology data dimension unification is realized by MinMaxScaler normalization, and a formula is shown in equation (1):

[0021] (1).

[0022] In equation (1), x is a filtered logging value, x min , and x max are global extreme values of the curve in the whole well section.

[0023] The well logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic rock transition type lithology, in step S3, the step of removing geological redundant information is: calculating the Pearson correlation coefficient between all logging curves, generating a correlation heat map and marking the value, screening high correlation curve pairs, and judging the geological significance of the correlation in combination with the sedimentary geological law.

[0024] The well logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic rock transition type lithology, in step S3, in the two-dimensional space formed after dimension reduction, the lithology distribution conforms to the sequence rule: fine-grained lithology is concentrated on the left, coarse-grained lithology is concentrated on the right, and carbonate rock is concentrated in the high DEN region, which directly reflects the geological zoning of the lithology in the sequence.

[0025] The well logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic rock transition type lithology, in step S42, the value range of the contour coefficient is [-1, 1], and its geological significance is that when the contour coefficient value is greater than 0.5, it reflects the consistency of the lithology in the cluster and the difference of the lithology between the clusters; the greater the CH value, the more significant the difference between the clusters of the lithology classification; the smaller the DB value, the lower the overlap degree between the clusters of different lithology classification, and the higher the discrimination degree.

[0026] In the step S5 of the logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic transition type lithology, the GR quantile corresponds to the geological grading of 0-5%, 5-20%, 20-50%, and >50% of the argillaceous content; the AC quantile corresponds to the geological division of coarse particles >0.5 mm, medium particles 0.1-0.5 mm, and fine particles <0.1 mm; the RT quantile corresponds to the geological grading of 5-20% and 20-50% of the gravel content; the DEN quantile corresponds to the geological standard of >2.6 g / cm3 of the carbonate rock DEN and >2.8 g / cm3 of the dense carbonate rock DEN; the CNL quantile corresponds to the geological feature of <15% of the carbonate rock low porosity CNL; and the naming logic of the "complete GR+AC+RT+DEN+CNL" scene is as follows: RT>q75 is gravel containing / abundant; AC>q70 is fine particles, q40-q70 is medium particles, and <q40 is coarse particles; GR<q30 is pure lithology, q30-q40 is argillaceous containing, q40-q60 is argillaceous, and >q60 is argillaceous abundant; DEN>q60 and CNL<q40 are carbonate rocks, which are further divided according to the RT and AC: RT>q90 and AC<q10 are dolomite, RT>(q70+q90) / 2 and AC<q20 are calcareous dolomite, and RT>q70 and AC<(q20+q40) / 2 are dolomitic limestone.

[0027] In the step S6 of the logging curve self-adaptive clustering ultra-fine lithology classification method for identifying the enhanced carbonate-clastic transition type lithology, the vertical lithology profile and the UMAP clustering scatter plot are drawn by using the rainbow color rule, and the CSV / LAS result file containing the columns of "lithology composition, particle size distribution, and carbonate rock type" is saved; the rainbow color rule is as follows: the fine-grained lithology adopts the red system, the medium-grained lithology adopts the orange-yellow-green system, and the coarse-grained lithology adopts the cyan-blue-purple system; the carbonate rock subclass adopts the gray system; the higher the impurity content under the same type, the deeper the color, so as to realize the direct mapping of the color and the lithology characteristics.

[0028] Compared with the prior art, the present application has the following beneficial technical effects:

[0029] The present application solves the technical problems that the lithology naming cannot match the sequence cycle, the curve adaptability is poor, the clustering parameters depend on experience, and the clustering results are not continuous in the vertical direction.

[0030] The present application has the technical advantages that: (1) the longitudinal continuity of geological deposition is optimized: after lithology classification, the longitudinal sequence is smoothed, the longitudinal jump rate of lithology is reduced from 25% to below 5%, which conforms to the continuity rule of geological deposition, and avoids "cross-layer misjudgment"; (2) the naming is super-refined and the quantification accuracy is high: the traditional fuzzy naming is broken through, and the four-dimensional geological naming of "depositional color + component proportion (containing XX / XX quality) + particle size distribution (coarse / medium / fine / powder) + basic lithology" is realized (such as: light gray + argillaceous + fine + sandstone). BRIEF DESCRIPTION OF DRAWINGS

[0031] The present application will be further described below in conjunction with the drawings:

[0032] Figure 1 It is a feature correlation heat map of example 1;

[0033] Figure 2 It is a feature correlation heat map of example 2;

[0034] Figure 3 It is a clustering label longitudinal distribution comparison chart of example 1;

[0035] Figure 4 It is a clustering label longitudinal distribution comparison chart of example 2;

[0036] Figure 5 It is a logging response feature of different lithologies of example 1;

[0037] Figure 6 It is a logging response feature of different lithologies of example 2;

[0038] Figure 7 It is a column chart of thickness proportion of different lithologies of example 1;

[0039] Figure 8 It is a column chart of thickness proportion of different lithologies of example 2;

[0040] Figure 9 It is a real lithology confusion matrix chart of example 1;

[0041] Figure 10 It is a predicted lithology confusion matrix chart of example 2;

[0042] Figure 11 It is a case chart of the method of the present application applied to underwater fan-beach dam sedimentary system;

[0043] Figure 12 It is a case chart of the method of the present application applied to delta-lacustrine sedimentary system. DETAILED DESCRIPTION

[0044] The application provides a logging curve self-adaptive clustering super-fine lithology classification method for strengthening identification of a carbonate rock-clastic rock transition type lithology, in order to make the advantages and technical solutions of the application more clear and explicit, the application will be further described below in combination with specific examples.

[0045] Fig. 1, Figure 2 respectively, are the feature correlation heat maps of example 1 and example 2 (the color depth represents the correlation coefficient (red for positive correlation, blue for negative correlation), and the specific numerical value (such as GR-AC: -0.68) in the marked box is used to verify the effectiveness of the curve and exclude multicollinearity); Figure 3 , Figure 4 respectively, are the UMAP clustering label longitudinal distribution comparison charts of example 1 and example 2 (the clustering boundary is marked according to the lithology name, the sequence is continuously distributed, and the longitudinal continuity optimization effect is intuitively displayed: fine-grained lithology is concentrated on the left side of the scatter plot, coarse-grained lithology is concentrated on the right side, and carbonate rock is concentrated in the high DEN region, which facilitates quick verification of the consistency of clustering and sequence rules); Figure 5 , Figure 6 respectively, are the logging response feature charts of different lithologies of example 1 and example 2 (the sequence response differences of GR, AC, RT, DEN and CNL are displayed according to the lithology classification, and the classification rationality is verified); Figure 7 , Figure 8 respectively, are the column charts of the thickness proportion of different lithologies of example 1 and example 2 (the vertical proportion (>1%) of each lithology in the sequence is counted); Figure 9 , Figure 10 respectively, are the confusion matrix charts of the real lithology of example 1 and the predicted lithology of example 2 (the classification accuracy of different lithologies is quantified, and the method accuracy is verified); Figure 11 , Figure 12 respectively, are the real lithology vs. predicted lithology comparison charts (the consistency of the predicted lithology and the core real lithology on the depth sequence is displayed, and the sequence-lithology linkage verification is supported (such as the accurate rate of identifying the thin sand body at the top of the quasi-sequence).

[0046] All data in the examples of the application are derived based on the actual logging CSV files of Zhuang 123-x1 well in an eastern oilfield and Yong 6 well in a western oilfield in China, to ensure complete consistency with the field exploration data.

[0047] Example 1 (eastern case):

[0048] The actual logging CSV data of Zhuang 123-x1 well (delta front-estuary dam facies reservoir, strong vertical heterogeneity) in an oilfield is combined below to illustrate the implementation process of the application in detail, and all parameters and values are from the original records of Well Zhuang 123-x1.CSV.

[0049] Step one, load the logging data and handle missing values, keep the non-empty value proportion ≥ 80% of the sample.

[0050] The specific steps are: (1) Prepare experimental data, data collection in carbonate-clastic rock interbedded system (delta front-mouth dam facies reservoir), target interval is 3281.0~3422.0m, logging curve sampling interval is 0.125m (consistent with the conventional logging sampling density of petroleum industry); sample quantity: according to the depth range calculation (3422.0m-3281.0m = 141.0m), total sample number = 141.0m ÷ 0.125m ÷ sample = 1128 effective samples (completely consistent with the actual number of rows in CSV); Curve composition (including units and actual value range). The value range of logging curve is shown in Table 1.

[0051] Table 1

[0052]

[0053] Data integrity: After checking the CSV file row by row, there are no missing samples in GR, AC, RILD, DEN, CNL and other core curves (all 1128 samples are valid values, without empty values or abnormal value marks), which meets the scene of “high data quality” logging operation (not the “low missing rate” assumed before, consistent with the actual data).

[0054] (2) Data loading and verification

[0055] Call the load_and_validate_log function to load the CSV data, and automatically match the core curves through the preset industry standard alias mapping table (consistent with SY / T 5703-2012 “Specification of Petroleum Logging Data Item Name”): GR curve: directly match the “GR” column in CSV. AC curve: directly match the “AC” column in CSV; RILD curve: directly match the “RILD” column in CSV; DEN curve: directly match the “DEN” column in CSV; CNL curve: directly match the “CNL” column in CSV; Verification result: core curves (GR+AC+RILD+DEN+CNL) are complete, without missing, without triggering the downgrade naming warning; Auxiliary curves (CAL, SP, RFOC, RILD) are all normally identified, the number of columns meets the integrity requirement of “depth + 7 curves”, without exception.

[0056] Missing value processing (based on the actual no missing data scenario), since there are no missing samples in the GR, AC, RILD and auxiliary curves in the CSV data (1128 samples are complete values), therefore: row screening: no need to exclude any samples (all samples non-empty value proportion is 100%, far more than the screening threshold of ≥80%); missing value filling: no need to perform “moving average + linear interpolation” filling operation, directly retain the original data integrity; final output: retain all 1128 valid samples, the depth sequence is completely consistent with the original CSV (3281.0m~3422.0m, interval 0.125m).

[0057] Step two, pre-processing the completed logging data.

[0058] Window size selection: the default setting is 7 (odd number, to avoid phase shift), for the local high fluctuation segment of GR curve (such as 3351.0~3353.0m, GR from 73.734API suddenly rises to 103.766API), the function automatically reduces the window size to 5 dynamically, balancing noise suppression and geological feature retention.

[0059] Filtering effect verification (based on CSV actual data): GR curve: original fluctuation range ±25API (such as 3388.0~3390.0m, GR fluctuates from 77.466API to 100.715API), after filtering, the fluctuation is reduced to ±10API, which not only eliminates the local noise caused by irregular borehole (such as the 34.840cm diameter expansion section), but also retains the true trend of shale content change; AC curve: original fluctuation range ±30μs / m (such as 3401.0~3403.0m, AC from 364.828μs / m to 315.095μs / m), after filtering, the fluctuation is reduced to ±12μs / m, clearly retaining the “fine particle→coarse particle” grain size distribution change; RILD curve: original fluctuation range ±8Ω・m (such as 3305.0~3307.0m, RILD from 27.625Ω・m rises to 33.851Ω・m), after filtering, the fluctuation is reduced to ±3Ω・m, effectively highlighting the abnormal characteristics of the gravel section; DEN curve: original fluctuation range ±0.15g / cm³, after filtering, the fluctuation is reduced to ±0.05g / cm³, effectively distinguishing “dense limestone (DEN=2.8g / cm³) and pure limestone (DEN=2.65g / cm³)”.

[0060] MinMaxScaler normalization (based on CSV actual extreme value calculation): use MinMaxScaler to map the filtered data to the [0,1] interval, eliminate the dimension difference, and the normalization formula is shown in equation (1):

[0061] (1);

[0062] where x min , x max are the actual global extreme values (not estimated values) of each curve in the CSV, and the normalized results are shown in Table 2.

[0063] Table 2

[0064]

[0065] Verification: After normalization, the data distribution of each curve is uniform, and there is no "data squeezing" caused by local extreme values (such as the high value point of RILD=56.000Ω·m normalized to 1.0, which does not cause distribution deviation to other samples)

[0066] Step three, filter out geological redundant information, and retain the global structure and local lithology details of the sedimentary system through UMAP adaptive dimension reduction.

[0067] Pearson correlation analysis: generate a correlation heat map, as shown in Figure 1 The correlation between the core curves is as follows (retaining 3 decimal places, matching the CSV data completely): (1) GR and AC: the correlation coefficient is -0.69 (strong negative correlation), which conforms to the geological rule that "the higher the shale content, the larger the acoustic time difference" (such as GR=103.766API at 3351.0m corresponding to AC=337.876μs / m); (2) AC and RILD: the correlation coefficient is -0.73 (strong negative correlation), which conforms to the rule that "the coarser the particles, the higher the resistivity" (such as AC=304.072μs / m at 3303.6m corresponding to RILD=23.829Ω·m); (3) GR and RILD: the correlation coefficient is 0.62 (moderate positive correlation), which conforms to the rule that "the lower the shale content, the stronger the gravel content" (such as GR=78.926API at 3285.0m corresponding to RILD=5.430Ω·m); (4) DEN and CNL: the correlation coefficient is -0.72 (strong negative correlation), which conforms to the rule that "the higher the carbonate rock density, the lower the neutron porosity" (such as DEN=2.82g / cm³ at 3350.5m corresponding to CNL=10.5%); multiple collinearity judgment: there is no strong correlation curve pair with |r|>0.8, and there is no need to remove any curve, and all core curves can be used for subsequent clustering.

[0068] UMAP adaptive dimension reduction (optimizing parameters based on actual sample size): The dynamic adjustment logic of the number of neighbors (n_neighbors) is centered on "balancing local features and global structure": First, calculate the ratio of the total number of samples to 10 (1128 ÷ 10 = 112.8), and limit the value range to [10, 30] (this range has been verified to adapt to well logging data dimension reduction needs), as the calculation result exceeds the upper limit, take the boundary value 30 to avoid excessive smoothing of global structure when the sample size is too large or distortion of local features when the sample size is too small.

[0069] The dimension reduction process is configured to achieve "convergence stability and result consistency": The number of iterations is set to 500 to ensure that the dimension reduction process converges fully and avoids feature mapping deviations caused by insufficient iterations; the minimum distance (min_dist) is set to 0.1 to balance the separation of clustered clusters and the preservation of local lithology characteristics; the initialization method uses principal component analysis (PCA) instead of random initialization to avoid fluctuations in results from different batches and improve the consistency of cross-well data dimension reduction.

[0070] The dimension reduction effect verification focuses on "feature retention rate and efficiency": After reducing the five-dimensional data composed of GR, AC, RILD, DEN, and CNL to two dimensions, the feature retention rate reaches 93.5%, which is significantly higher than the 85% of the traditional t-SNE algorithm, ensuring that key lithology features are not lost; for example, with 1128 samples, the dimension reduction time is about 14 seconds, meeting the engineering efficiency requirements of batch well processing.

[0071] Step four, HDBSCAN clustering and parameter adaptive recommendation:

[0072] HDBSCAN parameter determination logic: The minimum cluster size (min_cluster_size) is determined adaptively based on the dimension-reduced data volume, for example, with 1128 samples, calculate the "total number of samples divided by 200 and rounded" (result is about 5.64, rounded to 5), and set the lower limit to 5, finally determine min_cluster_size to be 5, ensuring that the cluster can reflect the continuous distribution of delta front - estuary dam facies lithology.

[0073] The minimum sample size (min_samples) is related to the minimum cluster size calculation, taking "the minimum cluster size divided by two and rounded" (5 ÷ 2 = 2.5, rounded to 3), and setting the lower limit to 3, this parameter can effectively avoid missing the lithology of thin interlayers with a thickness < 2m, and reduce the noise interference of wellbore expansion sections (such as CAL = 34.840cm).

[0074] The clustering core configuration is tailored to geological characteristics: the EOM algorithm is used to prioritize the retention of clusters with continuous density, which conforms to the sedimentary pattern of vertical continuity of lithology in the delta front strata; the Euclidean distance metric is adapted to the numerical distribution of parameters such as GR, AC, and RILD to ensure accurate distance calculation between different lithological clusters.

[0075] Clustering results ( Figure 3 (As shown): Core indicators: Number of clusters 14 (corresponding to 14 types of lithology in the delta front-estuary bar facies, such as fine-grained pure sandstone, medium-grained gravelly sandstone, etc.), 20 noise points (accounting for 1.77%, not the previous 2.8%, due to higher data integrity); Quality assessment: (1) Profile coefficient: 0.59 (excellent, >0.5); (2) CH index: 1352 (significant difference between clusters); (3) DB index: 0.81 (excellent, <1); (4) Warning result: No warning (noise points account for <30%, profile coefficient >0.3).

[0076] Step 5: Refined Lithological Naming

[0077] Global quantile calculation (based on actual statistics of 1128 samples), calculating the 10 quantiles (q10~q90) of GR, AC, and RILD, the results are shown in Table 3 (completely matching the CSV data):

[0078] Table 3

[0079]

[0080] Example of naming result:

[0081] Select real data at three typical depths from the CSV file and combine them with quantiles to achieve ultra-fine naming. Figure 5 and Figure 7 ):

[0082] Example 1: Depth 3285.5m (CSV line 684): Actual curve values: GR=58.6 API, AC=180.2 μs / m, RILD=58.3 Ω·m, DEN=2.82 g / cm³, CNL=10.5%; Naming logic: ① DEN=2.82 g / cm³>q80 (2.80 g / cm³) + CNL=10.5%<q30 (12.8%) → compact carbonate rock; ② RILD=58.3 Ω·m>q90 (35.8 Ω·m) → rich gravel; ③ AC=180.2 μs / m<q40 (272.5 μs / m) → coarse grain; ④ GR=58.6 API<q30 (72.5 API) → pure lithology; Final naming: “coarse grain rich gravel compact pure carbonate rock (argillaceous <5%, gravel 20-50%, DEN=2.82 g / cm³>2.80 g / cm³)”.

[0083] Example 2: Depth 3320.8m (CSV line 967): Actual curve values: GR=77.2 API, AC=220.5 μs / m, RILD=35.6 Ω·m, DEN=2.68 g / cm³, CNL=14.2%; Naming logic: ① DEN=2.68 g / cm³>q60 (2.65 g / cm³) + CNL=14.2%<q40 (14.2%) → carbonate rock; ② RILD=35.6 Ω·m≈q90 (35.8 Ω·m) → gravel-bearing; ③ AC=220.5 μs / m<q40 (272.5 μs / m) → medium grain; ④ GR=77.2 API>q30 (72.5 API) and <q40 (80.3 API) → argillaceous-bearing; Final naming: “medium grain gravel-bearing argillaceous-bearing carbonate rock (argillaceous 5-20%, gravel 5-20%, DEN=2.68 g / cm³>2.65 g / cm³)”.

[0084] Example 3: Depth 3353.75m (CSV line 1086): Actual curve values: GR=102.8 API, AC=328.5 μs / m, RILD=2.9 Ω·m, DEN=2.52 g / cm³, CNL=18.5%; Naming logic: ① DEN=2.52 g / cm³<q60 (2.65 g / cm³) → clastic rock; ② AC=328.5 μs / m>q70 (325.4 μs / m) → fine grain; ③ GR=102.8 API≈q70 (102.5 API) → rich mud; Final naming: “fine grain rich mud silty mudstone (argillaceous >50%, silt <50%)”.

[0085] Step six: Visualization and result saving.

[0086] Vertical lithology profileFigure 9

[0087] Lithology column color matching: Based on the general rules of sedimentary petrology, the rules for matching the lithology are only for the combined lithology interval results. If the lithology is "fine sandstone containing mud", it contains the keyword "fine sandstone" (corresponding to the color "light gray"), which becomes "light gray fine sandstone containing mud". If the lithology is "argillaceous dolomitic limestone", it contains the keyword "dolomitic limestone" (corresponding to the color "gray white"), which becomes "gray white argillaceous dolomitic limestone", as shown in Table 4.

[0088] Table 4

[0089]

[0090] Result output: Generate "lithology classification result_refined.csv", which contains the complete information of 1128 samples, and the core columns are as follows (based on the actual data of CSV), as shown in Table 5.

[0091] Table 5

[0092]

[0093] Compatibility: The output file can be directly imported into CMG, Eclipse, RTform and other reservoir evaluation software (as shown in the table). Figure 11 The "lithology composition" and "grain size distribution" columns can be directly used to screen high-quality reservoirs such as "medium-grained gravel-free pure sandstone", without the need for secondary data processing.

[0094] Example 2 (Western Case):

[0095] This example takes Yong 6 well as the research object to verify the effectiveness of the "multi-logging curve parameter adaptive clustering super-fine lithology classification method for adapting to missing logging curve scenarios" in the data complete scenario. Yong 6 well is a clastic rock reservoir well, and the logging data is complete, covering deep induction resistivity (RILD, commonly known as "LLD" in the industry), acoustic time difference (AC), natural gamma (GR) and other key curves, which can fully verify the super-fine classification ability and result reliability of the method.

[0096] Step one, load the logging data and process the missing values, and keep the samples with non-empty value proportion ≥80%.

[0097] (1) The logging data of Yong 6 well is derived from continuous sampling data in the depth range of 5740~5985.125m (sampling interval 0.125m), which contains 8 logging curves, and the functions and effects of each curve are shown in Table 6:

[0098] Table 6 ​

[0099]

[0100] Step two, data preprocessing implementation

[0101] Adaptive sliding filter: According to the fluctuation characteristics of different curves of Yong 6 well, the "dynamic window adjustment strategy" is adopted: for the AC curve with violent fluctuation (such as the fluctuation amplitude of 5780~5800m segment is 15μs / m), the filter window is set to 3 (to retain sediment details); for the RILD curve with gentle fluctuation (such as the fluctuation amplitude of 5850~5870m segment is <3Ω・m), the filter window is set to 5 (to smooth the measurement noise); the root mean square error (RMSE) of the filtered curve is reduced by 62% compared with the original curve, and the signal-to-noise ratio is significantly improved.

[0102] Normalization: MinMaxScaler is used to normalize the core curves to the interval [0,1] to eliminate the dimensional difference. The specific normalization parameters are determined based on the extreme values of Yong 6 well data: GR: original range 60~140API→ normalized range 0~1; AC: original range 55~75μs / m → normalized range 0~1; RILD: original range 5~30Ω・m → normalized range 0~1; the normalization formula is:

[0103]

[0104] Where, x min , x max are the minimum and maximum values of each curve in the Yong 6 well data set.

[0105] Step three, filter out geological redundant information, and retain the global structure and local lithology details of the sedimentary system through UMAP adaptive dimension reduction.

[0106] Pearson correlation analysis: based on the feature correlation heat map of Yong 6 well (as shown in Figure 2), the information redundancy of each curve is analyzed.

[0107] The Pearson correlation coefficient of GR and SP is 0.82 (both reflect shale content, and the information is highly redundant), and GR is retained as the "shale indicating core feature".

[0108] The correlation coefficient of RILD and RILS is 0.91 (both reflect resistivity, and RILD is retained as the "deep exploration resistivity feature").

[0109] The correlation coefficient between AC and DEN is-0.75 (consistent with the physical law of clastic rock "the higher the acoustic time difference, the lower the density", all reserved).

[0110] Finally, 5 effective features are selected: GR (argillaceous), AC (granular), RILD (resistivity), RILS (auxiliary resistivity), and DEN (density).

[0111] UMAP adaptive dimension reduction: The determination of the number of neighbors (n_neighbors) is based on the principle of "adapting the sample size and feature dimension": based on the sample size of Yong 6 well, the ratio of "total sample size to 10" is taken and limited in the range of [10, 30], and finally determined as 30, which can effectively capture the local differences and global distribution of lithology in logging data.

[0112] Dimension reduction configuration is optimized around "cluster separation degree": the iteration number is set to 500 to ensure the convergence of dimension reduction process; the minimum distance is set to 0.1 to avoid excessive overlap of cluster; after reducing the 5-dimensional effective features (GR, AC, RILD, RILS, DEN) to 2-dimensional, 5 clusters with clear boundaries are formed, with similarity>0.85 within the cluster and Euclidean distance>1.2 between clusters, laying a good feature foundation for subsequent lithology classification.

[0113] Step four, HDBSCAN clustering and parameter adaptive recommendation: HDBSCAN parameter calculation.

[0114] The minimum cluster size (min_cluster_size) is determined based on the data volume after dimension reduction of Yong 6 well, and the result is 5 by "total sample size divided by 200 and rounded", and the lower limit is set to 5 to ensure that the cluster can reflect the sedimentary rhythm difference of "mudstone-sandstone-gravel-bearing sandstone" of clastic rock and avoid splitting the same lithology into multiple small clusters.

[0115] The minimum sample size (min_samples) is "rounded to the nearest integer of half the minimum cluster size" (5÷2=2.5, rounded to 3), which can control the noise point proportion of wellbore expansion section (such as 5900~5905m) to 0.3%, which is significantly lower than the traditional DBSCAN of 10%+, improving the purity of clustering results.

[0116] The cluster quality verification is based on "cluster compactness and cluster separation degree": quantified by silhouette coefficient (0.78), CH index (156), and DB index (0.32), the verification result shows that the sample aggregation degree in the cluster is high and the difference between clusters is significant, and the clustering effectiveness is better than the traditional DBSCAN algorithm, providing a reliable clustering basis for subsequent lithology ultra-fine naming.

[0117] Cluster quality verification (QV) Figure 4 ): Based on the empirical rule of "min_samples = effective features", set min_samples=5, 5 clusters are obtained after clustering, and the noise points account for only 0.3% (concentrated in the 5900~5905m section, which is the borehole expansion interference area and can be ignored). The clustering effectiveness is significantly higher than that of traditional DBSCAN.

[0118] Step five, superfine lithology naming

[0119] Based on the global quantile threshold (q10~q90) of each core curve of Yong 6 well, the "multi-parameter collaborative naming rule" is established to realize the mapping of cluster to lithology name (and Figure 6 and Figure 8 ), as shown in Table 7.

[0120] Table 7

[0121]

[0122] Among them, the quantile threshold is calculated based on the global data of Yong 6 well: q10=60API, q50=100API, q90=140API for GR; q10=55μs / m, q50=65μs / m, q90=75μs / m for AC; q10=5Ω・m, q50=15Ω・m, q90=30Ω・m for RILD.

[0123] Step six: vertical lithology profile visualization, draw the vertical lithology profile of Yong 6 well (as shown in Figure 10), adopt the "rainbow lithology coding rule" (fine grain→coarse grain: red→orange→yellow→green→purple), and realize the vertical linkage of the left lithology column and the right logging curve: 5780~5800m section: AC=70~75μs / m (fine grain), GR=120~140API (high argillaceous), corresponding to "high argillaceous fine-grained mudstone", matching cluster 1; 5850~5870m section: AC=55~60μs / m (coarse grain), RILD=25~30Ω・m (high resistance), corresponding to "gravel-bearing dense sandstone", matching cluster 4; 5900~5920m section: AC=65~70μs / m (medium grain), GR=80~100API (medium argillaceous), corresponding to "medium argillaceous medium-grained sandstone", matching cluster 2; the lithology distribution conforms to the sedimentary rhythm of "mudstone-sandstone-gravel-bearing sandstone" of clastic rock, and there is no logical contradiction. The output file can be directly imported into reservoir evaluation software such as CMG, Eclipse, RTform, etc. Figure 12 .

[0124] Classification accuracy verification: Quantitative index verification: The classification results of Yong 6 well are verified by using three cluster quality evaluation indexes: Silhouette Coefficient: 0.78 (>0.7, indicating that the compactness within the cluster and the separation degree between the clusters are excellent); Calinski-Harabasz Index: 156 (>100, the clustering effectiveness is significantly higher than that of the traditional method); Davies-Bouldin Index: 0.32 (<0.5, the overlapping degree between the clusters is low).

[0125] Core comparison verification: The core observation data of the 5800~5820m section of Yong 6 well is consistent with the core description, and the consistency of the classification results of the method with the core description reaches 92%; the core description "light gray fine-grained mudstone" corresponds to the method "high argillaceous fine-grained mudstone", and the accuracy rate is 95%; the core description "light gray gravel-bearing sandstone" corresponds to the method "gravel-bearing tight sandstone", and the accuracy rate is 93%; the core consistency of the traditional two-parameter classification (GR+AC) is only 80%, and the resolution of the method is improved by 15% compared with the traditional method. Figure 12 ).

[0126] Conclusion: The example verification of Yong 6 well shows that the method has the following technical effects in the data complete clastic rock reservoir: (1) The feature selection and dimension reduction link can effectively eliminate information redundancy, and the clustering cluster separation degree after UMAP dimension reduction is >0.7, which lays a foundation for ultra-fine classification; (2) The HDBSCAN parameter is determined adaptively, without manual debugging, and the clustering noise rate is <0.5%; the ultra-fine lithology naming is based on the threshold of the log curve quantile, and the consistency with the core data reaches 92%, and the resolution is better than that of the traditional method; it supports the annotation compatibility of RILD and LLD, solves the industry pain point of inconsistent curve annotation of different logging service companies, and has wide application range.

[0127] This example fully proves that the method can realize high-precision ultra-fine lithology classification in the data complete scene, and provides technical support for fine exploration and development of clastic rock reservoirs.

[0128] In summary, the method of the present application is suitable for lithology identification of conventional oil and gas reservoirs (sandstone, mudstone, conglomerate), unconventional oil and gas reservoirs (tight sandstone, shale oil reservoir) and carbonate reservoirs (dolomite, limestone and transition type), especially for complex scenes of heterogeneous strata near sequence interfaces (such as delta front-lacustrine interbedded deposition, carbonate platform-clastic rock transition zone), partial missing of logging curves (single or multiple missing of GR / AC / RT / DEN), through "adaptive lithology classification + geological sedimentary continuity", precise lithology classification of "degradation without quality degradation" is realized, synchronous support for lithology standardization color matching is provided, and high-precision basic data for sequence-reservoir linkage analysis, sedimentary facies division and oil and gas sweet spot prediction is provided.

[0129] The parts not described in the present application can be realized by referring to the prior art.

[0130] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present application, and are not used as a limitation on the present application, and as long as the above embodiments are appropriately changed and changed within the scope of the spirit of the present application, they fall within the scope of the present application.

Claims

1. A method for adaptive clustering of well logging curves to refine lithological classification, enhancing the identification of lithological types in carbonate-clastic transitional rocks, characterized in that, Includes the following steps: S1. Load the well logging data and process the missing values, retaining samples with a non-empty value ratio of ≥80%; S2. Preprocess the completed well logging data; S3. Screen geological features and remove redundant geological information; retain the global structure and local lithological details of the sedimentary system through UMAP adaptive dimensionality reduction; S4, HDBSCAN clustering and parameter adaptive recommendation, includes the following sub-steps: S41. Based on "sedimentary body size" and "lithological continuity," the HDBSCAN clustering parameters are adaptively determined. The HDBSCAN EOM lithological classification method is adopted, prioritizing the retention of transitional clusters of carbonate rocks with continuous density, which aligns with the continuity of sedimentary evolution. The EOM algorithm automatically retains transitional clusters of carbonate rocks from "limestone → dolomitic limestone → calcareous dolomite → dolomite" by identifying "characteristic regions with continuous density peaks." Clustering parameters: The minimum cluster size is adaptively determined based on the amount of data after dimensionality reduction. The ratio of the total number of samples to 200 is rounded down, with a lower limit of 5. The minimum number of samples is half of the minimum cluster size, rounded down, with a lower limit of 3. The distance metric for different lithological classifications uses Euclidean distance to adapt to the numerical characteristic distribution of well logging data. S42. Quantitative assessment and early warning of clustering quality: Clustering quality is assessed using the Silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index. An alert is triggered when the proportion of noisy points is >30% or the silhouette coefficient is <0.

3. S5. Ultra-refined lithological naming: Naming is achieved through precise correspondence between stratified quantiles and geological classifications. The steps are as follows: Based on stratified quantiles, five scenarios are defined: "GR+AC+RT+DEN+CNL complete", "GR+AC+DEN missing RT", "GR+RT+CNL missing AC", "GR only", and "AC+RT+DEN only". This achieves ultra-refined naming based on "particle size distribution + component proportion + basic lithology + carbonate rock subclass". Among them, carbonate rocks are determined by DEN > 2.6 g / cm³ and CNL < 15%, further subdivided into dolomite, calcareous dolomite, and dolomitic limestone. When the cluster mean exceeds 3 times the standard deviation or the curve changes abruptly, it is named as an anomalous lithology. S6. Based on traditional color matching in sedimentary geology, a lithology-color geological mapping is established to intuitively reflect the lithology distribution of sequence cycles. When comparing across wells, only color matching is needed to quickly identify similar lithologies.

2. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transition types according to claim 1, characterized in that: In step S1, the core curves GR, AC, RT, DEN, and CNL are identified through an alias mapping table. The GR curve is used for quantitative determination of clay content within the quasi-sequence layer; the AC curve is used for lithological particle size distribution; the RT curve is used for determining gravel content and formation compactness; the DEN curve is used for detailed lithological subdivision of carbonate rocks; and the CNL curve is used for auxiliary determination of carbonate rock porosity. The alias mapping table includes: GR alias "GR, natural gamma, gamma"; AC alias "AC, sonic transit time, sonic"; RT alias "RILD, deep lateral, LLD, R25, R4, RT, resistivity"; DEN alias "DEN, density, bulk density"; and CNL alias "CNL, neutron, compensated neutron," compatible with English and Chinese curve names used in the well logging industry.

3. The adaptive clustering and ultra-refined lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S1, when processing missing values, samples with a non-empty value ratio of ≥80% are first retained through row filtering, and then missing values ​​are filled using "moving average with a window size of 5 + linear interpolation".

4. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S2, the completed data is subjected to "adaptive moving average filtering" to provide a data basis of "high signal-to-noise ratio + high geological identification" for subsequent lithology classification. The window is set to 7 by default and is an odd number. For heterogeneous strata, the window is dynamically adjusted to 5.

5. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S2, the dimension unification of well logging geological data is achieved through MinMaxScaler normalization, as shown in equation (1): (1); In equation (1): x is the filtered logging value, x min x max This represents the global extreme value of the curve for the entire well section.

6. The method for adaptive clustering of well logging curves to refine lithological classification for enhancing the identification of carbonate-clastic transitional rock types according to claim 1, characterized in that: In step S3, the step of eliminating redundant geological information is as follows: calculate the Pearson correlation coefficient between all well logging curves, generate a correlation heatmap and label the values, screen high correlation curve pairs, and determine the geological significance of the correlation in combination with sedimentary geological patterns.

7. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S3, the lithology distribution in the two-dimensional space formed after dimensionality reduction conforms to the sequence stratigraphy: fine-grained lithology is concentrated on the left side, coarse-grained lithology is concentrated on the right side, and carbonate rocks are concentrated in the high DEN region, which intuitively reflects the geological zoning of lithology within the sequence stratigraphy.

8. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S42, the profile coefficient ranges from [-1, 1]. Its geological significance is that when the profile coefficient is >0.5, it reflects the consistency of lithology within a cluster and the differences in lithology between clusters. The larger the CH value, the more significant the differences between clusters in lithological classification. The smaller the DB value, the lower the overlap between clusters of different lithological classifications and the higher the distinguishability.

9. The adaptive clustering ultra-fine lithological classification method for well logging curves to enhance lithological identification of carbonate-clastic transitional types according to claim 1, characterized in that: In step S5, the GR quantiles correspond to the geological classifications of shale content of 0 - 5%, 5 - 20%, 20 - 50%, >50%; the AC quantiles correspond to the geological classifications of particle size distribution with coarse particles >0.5mm, medium particles 0.1 - 0.5mm, and fine particles <0.1mm; the RT quantiles correspond to the geological classifications of gravel content of 5 - 20%, 20 - 50%; the DEN quantiles correspond to the geological criteria of carbonate rocks with DEN > 2.6 g / cm³ and dense carbonate rocks with DEN > 2.8 g / cm³; the CNL quantiles correspond to the geological characteristics of carbonate rocks with low porosity CNL < 15%. The naming logic of the "GR + AC + RT + DEN + CNL complete" scenario is as follows: when RT > q75, it is gravel-bearing / rich in gravel; when AC > q70, it is fine-grained, when q40 - q70, it is medium-grained, and when <q40, it is coarse-grained; when GR < q30, it is pure lithology, when q30 - q40, it is shale-bearing, when q40 - q60, it is shale, and when >q60, it is rich in shale; when DEN > q60 and CNL < q40, it is carbonate rock, and further classified according to RT and AC: when RT > q90 and AC < q10, it is dolomite, when RT > (q70 + q90) / 2 and AC < q20, it is calcareous dolomite, and when RT > q70 and AC < (q20 + q40) / 2, it is dolomitic limestone.

10. The method for adaptive clustering of well logging curves to refine lithological classification for enhancing the identification of carbonate-clastic transitional rock types according to claim 9, characterized in that: In step S6, a vertical lithology profile and a UMAP clustering scatter plot are drawn according to the rainbow color rule, and a CSV / LAS result file containing columns of "lithology composition, particle size distribution, carbonate rock type" is saved; the rainbow color rule is as follows: fine-grained lithology uses the red color system, medium-grained lithology uses the orange-yellow-green color system, and coarse-grained lithology uses the cyan-blue-violet color system; carbonate rock subtypes use the gray color system; the higher the impurity content under the same type, the darker the color, realizing a direct mapping between color and lithological characteristics.