A method for reconstructing continuous missing gamma well logs

Abstract

The present application relates to a kind of continuous missing gamma well logging curve reconstruction methods, steps are as follows: multi-source data acquisition and uniform depth sampling;Anti-leakage division training well and target well;Continuous missing interval synchronous masking;Search prior generation;In-well anchoring correction;Multi-modal local residual correction;Complete curve output.Improve the utilization value of continuous missing well logging data.The method search library is limited to be constructed only by training well, the real natural gamma in the continuous missing interval of target well does not participate in any training, search or preprocessing process, significantly reduces the risk of information leakage;Ensure that the reconstruction is completed without contacting the real gamma curve in the continuous missing interval of target well.

Description

Technical Field

[0001] This invention relates to the fields of intelligent processing of oil and gas reservoir logging data, reconstruction of continuously missing logging curves, machine learning retrieval enhancement modeling, and multimodal information fusion technology, and particularly to a method for reconstructing natural gamma (GR) logging curves to prevent information leakage in scenarios with continuous missing data over long depth intervals. Background Technology

[0002] Natural gamma ray (GR) logging curves can reflect formation clay content, lithological assemblage, and formation variations, serving as crucial foundational data for sandstone and mudstone identification, stratigraphic division, reservoir heterogeneity analysis, and hydrocarbon reservoir interpretation. However, during actual well logging acquisition and historical data processing, factors such as instrument malfunctions, changes in wellbore conditions, incomplete logging series, acquisition cost limitations, and differences in data archiving can cause partial, continuous, or even complete missing sections of the gamma ray logging curve for target wells.

[0003] Existing methods for reconstructing missing logging curves mainly include traditional machine learning regression, deep sequence models, generative models, and multi-source information fusion models. Traditional machine learning methods have low computational cost, but their expressive power is insufficient when dealing with complex nonlinear responses and strongly nonstationary logging sequences. Deep sequence models can learn the statistical mapping relationship between logging curves, but when the gamma curve is completely missing in a long depth range, there is a lack of direct observation anchor points within the missing segment, which easily leads to problems such as overall trend drift, weakened peak and valley responses, and over-smoothing.

[0004] Generative models can learn the potential distribution of logging curves and generate candidate curves, but the generated results are difficult to precisely constrain the geological background, amplitude scale, and changes in adjacent layers at a specific depth of the target well. When there are differences in sedimentary background, sand layer development, and logging response amplitude across wells, directly transferring training well information may cause amplitude shifts and local morphological distortions.

[0005] Some existing methods attempt to incorporate multimodal data such as well logging images, formation information, and sand layer information. However, if multi-source information is directly stitched into the regression model, the model struggles to distinguish the roles of numerical curves, image structures, and geological labels, potentially introducing modal noise. More importantly, if real gamma information of the missing segments of the target well is used during normalization, search library construction, candidate sample generation, or image processing, information leakage will occur, leading to an overestimation of the model's generalization ability in blind well applications.

[0006] In summary, the existing technologies still have the following problems: First, there is a lack of direct observation constraints within long continuous missing segments; second, the amplitude scales of training well priors and target wells are inconsistent; third, local peaks and valleys, abrupt changes, and high gradient intervals are difficult to recover stably; fourth, multimodal information lacks a controlled usage mechanism for missing intervals; and fifth, there is a lack of a comprehensive leak prevention technology solution covering the entire process from training well normalization, retrieval library construction, image masking to target well prediction. Summary of the Invention

[0007] This invention aims to address the aforementioned problems by providing a method for reconstructing continuously missing gamma logging curves based on leakage prevention retrieval and enhanced multimodal learning. This method solves the problems of trend drift, cross-well amplitude shift, local morphological distortion, and high information leakage risk that easily occur in the reconstruction of long continuous missing segments by existing methods.

[0008] This invention is applicable to engineering scenarios such as oil and gas field development, reservoir fine interpretation, sandstone and mudstone identification, formation correlation and historical logging data supplementation. It is particularly suitable for situations where the gamma curve of the target well is completely missing in a long depth range, but other logging curves and auxiliary geological information are still available.

[0009] The technical solution of this invention is as follows:

[0010] A method for reconstructing continuously missing gamma logging curves, the method is as follows: Multi-source data acquisition and unified depth sampling: Acquire conventional logging curves, logging images, formation information and sand layer information from multiple wells; among them, conventional logging curves include natural gamma curves and conditional curves used for auxiliary reconstruction, and are resampled at a unified depth interval; Leakage prevention is divided into training wells and target wells: training wells and target wells are divided on a well-by-well basis; Synchronous masking of continuous missing intervals: Set continuous missing intervals in the natural gamma curve of the target well and mask the natural gamma values ​​of the intervals, while masking the natural gamma orbits within the corresponding depth range in the logging image. Retrieval Prior Generation: Candidate depth windows are extracted from the training wells to construct the retrieval library, the similarity between continuous missing intervals and candidate depth windows is calculated, and candidate samples are selected to generate retrieval priors; In-well anchoring correction: The scale correction parameter and offset correction parameter are estimated by using the observed natural gamma value outside the continuous missing interval to perform anchoring correction on the retrieval prior and obtain the anchoring prior; Multimodal local residual correction: Input the condition curve of the target well, the masked logging image, formation information and sand layer information, and the anchoring prior into the local residual correction function to predict the residual term in the continuous missing interval. Add the anchoring prior to the residual term to obtain the reconstructed natural gamma value in the continuous missing interval. Complete curve output: The continuous missing intervals are identified by the missing mask, the observed natural gamma values ​​of the target well are retained, and the continuous missing intervals are replaced with the reconstructed natural gamma values, thus outputting the complete natural gamma curve.

[0011] The condition curves include at least one of the following: acoustic time difference curve, natural potential curve, deep induced resistivity curve, and medium induced resistivity curve.

[0012] The retrieval library consists only of candidate depth windows from training wells; real natural gammas of consecutive missing intervals in target wells are not included in the retrieval library.

[0013] The consecutive missing intervals are identified by a missing mask, and the specific process is as follows: In the formula: The starting depth of the consecutive missing interval; For the first Depth of each sampling point; The length of the consecutive missing interval; For the first Missing mask value for each sampling point; when When =1, it means the first... When each sampling point belongs to a continuous missing interval, When =0, it means the first... Each sampling point represents an observed interval.

[0014] The specific process for generating the retrieval prior is as follows: ; In the formula: For the first Prior values ​​for each sampling point; For the first The candidate sample at the th The gamma value of each sampling point; The maximum number of candidate samples; This is the normalized similarity weight.

[0015] The specific calculation process for the anchoring prior is as follows: In the formula: For the first Anchoring priors for each sampling point; These are the scale correction parameters; These are the offset correction parameters.

[0016] The specific calculation process for reconstructing the natural gamma value is as follows: ; In the formula: For the first The residual term for each sampling point; This is a local residual correction function; , , , The order is number 1 Condition curves, well logging images, formation information, and sand layer information for each sampling point; To reconstruct the natural gamma value.

[0017] The formula for fusing the complete natural gamma curve is: In the formula: A complete natural gamma curve; Missing mask; The target well has been observed to have natural gamma values; To anchor priors; For residual terms; This indicates element-wise multiplication.

[0018] It also includes normalizing the training wells and target wells before generating the prior knowledge.

[0019] The technical advantages of this invention are as follows:

[0020] 1. The retrieval library of this method is limited to be constructed only from training wells. The real natural gamma within the continuous missing interval of the target well does not participate in any training, retrieval or preprocessing process, which significantly reduces the risk of information leakage; and ensures that the reconstruction is completed without touching the real gamma curve of the continuous missing interval of the target well.

[0021] 2. By limiting the retrieval library through training wells, external morphological references are provided for the continuous missing intervals of the target wells, which alleviates the trend drift and oversmoothing problems caused by the lack of observation constraints within the continuous missing intervals in the end-to-end regression method.

[0022] 3. By using the observed natural gamma values ​​outside the continuous missing interval of the target well for scale and offset anchoring, the prior of training well retrieval can be adapted to the local amplitude and trend of the target well, reducing the amplitude offset in cross-well migration.

[0023] 4. By focusing on high-error regions and correcting local residuals, the reconstruction stability of local peaks and valleys, abrupt responses, high-gradient intervals, and complex sandstone-mudstone transition zones is improved.

[0024] 5. A controlled usage mechanism is adopted for well logging images, formation information, and sand layer information, so that multimodal information mainly serves candidate screening and residual refinement, without replacing the dominant constraints of numerical curves, thus balancing numerical accuracy, geological rationality, and engineering interpretability.

[0025] 6. This method is applicable to the completion of historical data in oil and gas fields, repair of missing logging curves, restoration of reservoir interpretation continuity, and auxiliary interpretation in blind well scenarios, thereby enhancing the utilization value of continuously missing logging data. Attached Figure Description

[0026] Figure 1 This is a flowchart of the present invention.

[0027] Figure 2 A diagram illustrating data partitioning to prevent information leakage.

[0028] Figure 3 This is a diagram of the architecture of the present invention. Detailed Implementation

[0029] Example 1: A method for reconstructing continuously missing gamma logging curves, the method is as follows: Multi-source data acquisition and unified depth sampling: Acquire conventional logging curves, logging images, formation information and sand layer information from multiple wells; among them, conventional logging curves include natural gamma curves and conditional curves used for auxiliary reconstruction, and are resampled at a unified depth interval; Leakage prevention is divided into training wells and target wells: training wells and target wells are divided on a well-by-well basis; Synchronous masking of continuous missing intervals: Set continuous missing intervals in the natural gamma curve of the target well and mask the natural gamma values ​​of the intervals, while masking the natural gamma orbits within the corresponding depth range in the logging image. Retrieval Prior Generation: Candidate depth windows are extracted from the training wells to construct the retrieval library, the similarity between continuous missing intervals and candidate depth windows is calculated, and candidate samples are selected to generate retrieval priors; In-well anchoring correction: The scale correction parameter and offset correction parameter are estimated by using the observed natural gamma value outside the continuous missing interval to perform anchoring correction on the retrieval prior and obtain the anchoring prior; Multimodal local residual correction: Input the condition curve of the target well, the masked logging image, formation information and sand layer information, and the anchoring prior into the local residual correction function to predict the residual term in the continuous missing interval. Add the anchoring prior to the residual term to obtain the reconstructed natural gamma value in the continuous missing interval. Complete curve output: The continuous missing intervals are identified by the missing mask, the observed natural gamma values ​​of the target well are retained, and the continuous missing intervals are replaced with the reconstructed natural gamma values, thus outputting the complete natural gamma curve.

[0030] Example 2, based on Example 1, the condition curve includes at least one of the following: acoustic time difference curve, natural potential curve, deep induced resistivity curve, and medium induced resistivity curve; The retrieval library consists only of candidate depth windows from training wells; real natural gammas of consecutive missing intervals in target wells are not included in the retrieval library.

[0031] Example 3, based on Example 2, The consecutive missing intervals are identified by a missing mask, and the specific process is as follows: In the formula: The starting depth of the consecutive missing interval; For the first Depth of each sampling point; The length of the consecutive missing interval; For the first Missing mask value for each sampling point; when When =1, it means the first... When each sampling point belongs to a continuous missing interval, When =0, it means the first... Each sampling point represents the observed interval; The specific process for generating the retrieval prior is as follows: ; In the formula: For the first Prior values ​​for each sampling point; For the first The candidate sample at the th The gamma value of each sampling point; The maximum number of candidate samples; This is the normalized similarity weight. The specific calculation process for the anchoring prior is as follows: In the formula: For the first Anchoring priors for each sampling point; These are the scale correction parameters; These are the offset correction parameters. The specific calculation process for reconstructing the natural gamma value is as follows: ; In the formula: For the first The residual term for each sampling point; This is a local residual correction function; , , , The order is number 1 Condition curves, well logging images, formation information, and sand layer information for each sampling point; To reconstruct the natural gamma value. The formula for fusing the complete natural gamma curve is: In the formula: A complete natural gamma curve; Missing mask; The target well has been observed to have natural gamma values; To anchor priors; For residual terms; This indicates element-wise multiplication.

[0032] Example 4, based on Example 3, further includes normalizing the training wells and target wells before generating the prior knowledge.

[0033] Specific experimental case study – A method for reconstructing continuously missing gamma logging curves, the method is as follows:

[0034] Step 1, Multi-source data acquisition and unified depth sampling: Acquire conventional logging curves, logging images, formation information and sand layer information from multiple wells; wherein, the conventional logging curves include natural gamma curves and condition curves used to assist reconstruction (at least one of sonic transit time curves, spontaneous potential curves, deep induced resistivity curves and medium induced resistivity curves), and resample at a unified depth interval (0.125 m) so that the condition curves, logging images, formation information and sand layer information correspond under the same depth index.

[0035] Step 2, Leakage Prevention: Divide the training wells and target wells into training wells and target wells, using wells as the unit.

[0036] Step 3, Synchronous Masking of Continuous Missing Intervals: A continuous missing interval is defined in the natural gamma curve of the target well and identified by a missing mask. The natural gamma values ​​of this interval are masked, and the corresponding natural gamma trajectories within the depth range of the logging image are also masked to prevent the logging image from exposing the predicted GR morphology. The specific process of the missing mask identification is as follows: ;when When =1, it means the first... When each sampling point belongs to a continuous missing interval, When =0, it means the first... Each sampling point represents an observed interval.

[0037] Step 4, Normalization: Perform outlier checks and scale normalization on the conventional logging curves of the training set. The normalization formula is as follows: In the formula: For the first Original well logging values ​​from each sampling point; For the first Normalized well logging values ​​from each sampling point; This represents the global minimum value of the training well on a given conventional logging curve. This represents the global maximum value of the training well on a given conventional logging curve. It is a small constant to prevent the denominator from being zero. The normalized scale obtained from the training wells is applied to both the training and target wells. The true gamma ray statistic (GR) within the missing intervals of the target wells is not included in any statistics to ensure that the true gamma ray statistic information of the consecutive missing intervals of the target wells is not indirectly used. Since the true gamma statistics of consecutive missing intervals (which have been masked) do not participate... and Therefore, the subsequent condition curve features, retrieval distance, candidate weights and residual terms obtained based on this scale will not indirectly use the statistical information of the true GR of the missing segment of the target well; thus realizing the scale conversion between the training well and the target well.

[0038] Step 5, retrieve prior knowledge generation: (1) Parameter extraction: Extract the statistical and morphological features of the training well condition curve, the training well boundary context features corresponding to the continuous missing interval of the target well, the training well formation label and the training well sand layer label features, and then construct the statistical and morphological features of the training well, the training well boundary context features corresponding to the continuous missing interval of the target well and the training well geological label features. Statistical and morphological features of the known condition curves of the target well, contextual features of the boundary of the continuous missing interval of the target well, and stratigraphic and sand layer label features are extracted, and then the statistical and morphological features, contextual features of the continuous missing interval of the target well, and geological label features of the target well are constructed. (2) In addition, to characterize the local response relationship between the conditional curve and the natural gamma curve, the Pearson correlation coefficient is calculated: In the formula: For each condition curve in The value of each sampling point; The mean of the condition curve within the set of valid observation sampling points; For the natural gamma curve in The true natural gamma value of each sampling point; To represent the mean of the natural gamma curve within the set of valid observation sampling points; This represents the total number of sampling points; Here are the Pearson correlation coefficients for each condition curve; (3) Combine the Pearson correlation coefficients of each condition curve to form a local correlation feature. , ;in, is the Pearson correlation coefficient of the acoustic time-of-flight curve; is the Pearson correlation coefficient of the spontaneous potential curve; The Pearson correlation coefficient for deep induced resistivity; is the Pearson correlation coefficient for the induced resistivity; (4) Construct retrieval feature vectors for training wells and target wells respectively. : The statistical and morphological characteristics of the target well's condition curve, the contextual characteristics of the target well's continuously missing intervals, the target well's formation labels, the target well's sand layer label characteristics, and the target well's local correlation characteristics are concatenated to form the target well retrieval feature vector. ; The statistical and morphological characteristics of the training well condition curves, and the training wells corresponding to the continuous missing intervals of the target well (the first...) The training well retrieval feature vector is composed of boundary context features (each candidate depth window of training wells), formation labels, sand layer labels, and local correlation features of the training wells. ; (5) Compare the target wells to retrieve feature vectors and training well retrieval feature vector : Initial candidate distance between the two ; It reflects the comprehensive differences between training wells and target wells in terms of statistical and morphological characteristics, contextual characteristics, geological background, and local response relationships; Then calculate the first Initial similarity weights for candidate depth windows of training wells : ,and In the formula: The temperature coefficient is used; subsequently, the GR curve segments of all candidate depth windows of training wells are weighted to form the retrieval prior. , .

[0039] Step 6, In-well Anchoring Correction: Due to differences in sedimentary background, sand layer development, and logging response amplitude among different wells, directly transferring the training well retrieval prior to the target well may result in amplitude shift. Therefore, the scale correction parameters and shift correction parameters are estimated using the observed natural gamma values ​​outside the continuous missing interval to perform anchoring correction on the retrieval prior, resulting in the anchored prior. .

[0040] Step 7, Multimodal Local Residual Correction: Input the target well's condition curve, masked logging image, formation information, sand layer information, and anchoring priors into the local residual correction function. The local residual correction function is used to predict residual terms within continuously missing intervals. The specific calculation process of the residual terms is as follows: The residual term is used to compensate only for consecutive missing intervals. The anchored prior is added to the residual term to obtain the reconstructed natural gamma value within the consecutive missing intervals. Specifically, .

[0041] Step 8, Complete Curve Output: Continuous missing intervals are identified using a missing mask. The observed natural gamma values ​​of the target well are retained, and the reconstructed natural gamma values ​​replace the continuous missing intervals, thus outputting the complete natural gamma curve. .

[0042] Example 2, based on Example 1, aims to improve the reconstruction stability of local peaks and valleys, abrupt changes, and high gradient regions. It constructs a joint distance for focusing on high-error regions based on the differences between the training well and the target well in statistical and morphological features, boundary context features, formation labels, sand layer labels, and local correlation features. ; ; ; ; , ; ; In the formula: The difference in conditional curves is calculated using statistical, morphological, and local correlation features. For boundary context differences; Differences in geological background; The difference in high-error-risk intervals is calculated based on local peaks and valleys, abrupt changes, and high-gradient responses; , , and These are their respective weighting coefficients; Represents the set of condition curves; , , , These represent the nth consecutive missing intervals of the target well. The mean, standard deviation, average gradient, and local Pearson correlation coefficient of the condition curves; , , , These represent the mean, standard deviation, average gradient, and local Pearson correlation coefficient of the candidate depth window for the training well, respectively. to These are the corresponding weighting coefficients; and These represent the mean values ​​of the observed GR context outside the upper and lower boundaries of the continuous missing interval of the target well, respectively; and These represent the mean values ​​of the observed GR context outside the upper and lower boundaries of the candidate depth window for training wells, respectively. , These represent the slope changes of the upper and lower boundaries of the continuous missing interval of the target well, respectively; , These represent the slopes of the upper and lower boundaries of the candidate depth window for training wells, respectively. to These are the corresponding weighting coefficients; and These represent the formation labels for the continuous missing interval of the target well and the candidate depth window of the training well, respectively. and These respectively represent sand layer labels or sand layer development types; and These are the corresponding weighting coefficients; Indicates a consecutive missing interval; Indicates the target well number At the sampling point, the first Normalized condition curves; This represents the normalized condition curve for the sampling points corresponding to the candidate depth window of the training well; This represents the first-order difference, used to characterize high gradients and abrupt changes; This represents the second-order difference, used to characterize peaks, valleys, and curvature variations; Indicates the weight of the difficult region; to These are the weighting coefficients.

[0043] This step further transforms the relevant features formed in step 5 into a fine-screening mechanism for difficult intervals, enabling subsequent multimodal local residual correction to prioritize intervals prone to large errors, such as local peaks and valleys, abrupt changes, and high gradients. Furthermore, the target well's relevant features are calculated solely from the observed natural gamma context outside the continuously missing intervals and the available condition curves, formation information, and sand layer information within the continuously missing intervals.

[0044] Evaluation of Results: The complete natural gamma curve was evaluated using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The evaluation range was limited to the continuous missing interval, and the errors were calculated on the original natural gamma scale after inverse normalization. ; In one specific embodiment, 70 rigorously multimodal matched wells were used to construct a 100 m continuous missing interval, corresponding to 800 continuous sampling points. Experimental results show that this method can reduce the reconstruction error of continuous missing GR while maintaining the leak prevention verification process, and improve the local morphological recovery effect in high gradient and high error intervals.

[0045] Specific Application Case – To verify the effectiveness of this method, actual logging data from the Yucha-Daming well area of ​​the Xiasiwan Oilfield were used for the experiment. The experimental data and experimental protocol are listed in tabular form.

[0046] Table 1. Example Data and Experimental Protocol .

[0047] Table 2. Experimental results of different methods in the GR reconstruction task with 100 m consecutive missing values. .

[0048] Table 3. Error reduction results of GeoMMR-RAGCR compared to baseline comparison method .

[0049] Table 4. Reconstruction experimental results under different consecutive missing lengths .

Claims

1. A method for reconstructing continuously missing gamma logging curves, characterized in that, The method is as follows: Multi-source data acquisition and unified depth sampling: Acquire conventional logging curves, logging images, formation information and sand layer information from multiple wells; among them, conventional logging curves include natural gamma curves and conditional curves used for auxiliary reconstruction, and are resampled at a unified depth interval; Leakage prevention is divided into training wells and target wells: training wells and target wells are divided on a well-by-well basis; Synchronous masking of continuous missing intervals: Set continuous missing intervals in the natural gamma curve of the target well and mask the natural gamma values ​​of the intervals, while masking the natural gamma orbits within the corresponding depth range in the logging image. Retrieval Prior Generation: Candidate depth windows are extracted from the training wells to construct the retrieval library, the similarity between continuous missing intervals and candidate depth windows is calculated, and candidate samples are selected to generate retrieval priors; In-well anchoring correction: The scale correction parameter and offset correction parameter are estimated by using the observed natural gamma value outside the continuous missing interval to perform anchoring correction on the retrieval prior and obtain the anchoring prior; Multimodal local residual correction: Input the condition curve of the target well, the masked logging image, formation information and sand layer information, and the anchoring prior into the local residual correction function to predict the residual term in the continuous missing interval. Add the anchoring prior to the residual term to obtain the reconstructed natural gamma value in the continuous missing interval. Complete curve output: The continuous missing intervals are identified by the missing mask, the observed natural gamma values ​​of the target well are retained, and the continuous missing intervals are replaced with the reconstructed natural gamma values, thus outputting the complete natural gamma curve.

2. The method for reconstructing continuously missing gamma logging curves according to claim 1, characterized in that, The condition curves include at least one of the following: acoustic time difference curve, natural potential curve, deep induced resistivity curve, and medium induced resistivity curve.

3. The method for reconstructing continuously missing gamma logging curves according to claim 1, characterized in that, The retrieval library consists only of candidate depth windows from training wells; real natural gammas of consecutive missing intervals in target wells are not included in the retrieval library.

4. The method for reconstructing continuously missing gamma logging curves according to claim 1, characterized in that, The consecutive missing intervals are identified by a missing mask, and the specific process is as follows: In the formula: The starting depth of the consecutive missing interval; For the first Depth of each sampling point; The length of the consecutive missing interval; For the first Missing mask value for each sampling point; when When =1, it means the first... When each sampling point belongs to a continuous missing interval, When =0, it means the first... Each sampling point represents an observed interval.

5. The method for reconstructing continuously missing gamma logging curves according to claim 1, characterized in that, The specific process for generating the retrieval prior is as follows: ; In the formula: For the first Prior values ​​for each sampling point; For the first The candidate sample at the th The gamma value of each sampling point; The maximum number of candidate samples; This is the normalized similarity weight.

6. The method for reconstructing continuously missing gamma logging curves according to claim 5, characterized in that, The specific calculation process for the anchoring prior is as follows: In the formula: For the first Anchoring priors for each sampling point; These are the scale correction parameters; These are the offset correction parameters.

7. The method for reconstructing continuously missing gamma logging curves according to claim 6, characterized in that, The specific calculation process for reconstructing the natural gamma value is as follows: ; In the formula: For the first The residual term for each sampling point; This is a local residual correction function; , , , The order is number 1 Condition curves, well logging images, formation information, and sand layer information for each sampling point; To reconstruct the natural gamma value.

8. The method for reconstructing continuously missing gamma logging curves according to claim 7, characterized in that, The formula for fusing the complete natural gamma curve is: In the formula: A complete natural gamma curve; Missing mask; The target well has been observed to have natural gamma values; To anchor priors; For residual terms; This indicates element-wise multiplication.

9. The method for reconstructing continuously missing gamma logging curves according to claim 1, characterized in that, It also includes normalizing the training wells and target wells before generating the prior knowledge.

Images