An intelligent comparison and selection method for probabilistic graph formation models based on a double-layer optimization architecture
The intelligent selection method for probabilistic graphical stratigraphic models using a two-layer optimization architecture solves the data-driven selection problem of probabilistic graphical stratigraphic models, achieving intelligent and accurate stratigraphic modeling while avoiding subjective experience errors.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2026-03-02
- Publication Date
- 2026-07-14
Smart Images
Figure CN122391527A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of geotechnical engineering and geological engineering, specifically to an intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture. Background Technology
[0002] Stratigraphic modeling based on sparse borehole data is a crucial component of engineering surveying and design. Accurate prediction of stratigraphic distribution can effectively identify unique geological structures (such as weak interlayers and fault-induced water inrushes), which are the root causes of engineering accidents such as uneven settlement and groundwater intrusion. Therefore, accurately obtaining stratigraphic distribution through stratigraphic modeling not only significantly enhances the scientific rigor and relevance of surveying and design but is also a key factor in optimizing construction plans, rationally estimating project costs, and ensuring the safe operation of engineering projects.
[0003] In recent years, geostatistical methods for stratigraphic modeling have been widely applied, with Kriging being the most representative geostatistical method. Kriging quantifies the spatial correlation of strata by fitting a variogram to borehole data, and is a commonly used interpolation method for stratigraphic modeling. However, the fitting accuracy of the variogram is highly dependent on the borehole data density. In practical engineering, due to the high cost of drilling, borehole data is often extremely sparse, making it difficult for Kriging interpolation to accurately capture complex stratigraphic distributions. Therefore, introducing machine learning methods with strong feature extraction and nonlinear learning capabilities can effectively improve the accuracy of stratigraphic modeling.
[0004] In machine learning methods, probabilistic graphical models can reflect the spatial topological relationships of geological structures as conditional probability distributions at a certain location. Based on this mechanism, probabilistic graphical models are widely used in the field of stratigraphic modeling. Probabilistic graphical models are often divided into directed probabilistic graphical models and undirected probabilistic graphical models, with directed probabilistic graphical models also known as Bayesian networks.
[0005] Currently, some scholars have successfully applied probabilistic graphical stratigraphic models to the field of stratigraphic modeling. However, the current modeling methods based on probabilistic graphical stratigraphic models still have the following problems: 1) lack of intelligent comparison methods for probabilistic graphical stratigraphic models driven by site data; 2) lack of spatial initialization methods for stratigraphic modeling based on exploration data mileage and elevation; 3) lack of evaluation indicators that can reflect the accuracy of stratigraphic predictions in each stratigraphic modeling result.
[0006] Therefore, there is an urgent need for an intelligent method that can achieve data-driven model selection, avoiding errors caused by relying on subjective experience when selecting probabilistic stratigraphic models. Summary of the Invention
[0007] To address the problems existing in the prior art, this invention provides a probabilistic graphical stratigraphic model intelligent selection method based on a two-layer optimization architecture. This method enables data-driven intelligent selection of probabilistic graphical stratigraphic models and is applicable to the selection of probabilistic graphical stratigraphic models in stratigraphic simulation in engineering application scenarios. It effectively avoids prediction errors caused by selecting stratigraphic models based on subjective experience and solves the problems mentioned in the background art.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent selection of probabilistic graphical stratigraphic models based on a two-layer optimization architecture, comprising the following steps: S1. Construct a formation modeling space initialization method based on Kriging. Initialize the formation modeling space according to the method and the borehole data of the work site. Grid the formation modeling space and fill the corresponding grid with the borehole data. S2. Based on the initial formation modeling space and borehole data in step S1, construct probabilistic graphical formation models on the formation modeling space, including: 4-neighborhood system Markov random field, 8-neighborhood system Markov random field, and coupled Markov chain. S3. Based on the formation modeling space and borehole data in step S1, construct a dataset partitioning and usage method for the two-layer optimization architecture. The two-layer optimization architecture consists of an outer layer nested with an inner layer. Based on cross-validation, the dataset composed of borehole data is divided into an outer test set and an outer training set, where the outer training set is passed into the inner layer for parameter learning. S4. In the inner layer, the outer training set divided in step S3 is used as known data, and parameter learning is performed based on the probabilistic map stratigraphic model constructed in step S2 and the corresponding parameter learning method. S5. In the outer layer, based on the optimal parameters learned in step S4, the outer layer training set in step S3 is used as known data to perform formation simulation based on the probabilistic map formation model constructed in step S2. A formation weighted accuracy calculation method is proposed, and the formation weighted accuracy is calculated based on the formation simulation results and the borehole data of the outer layer test set. S6. Repeat steps S4 and S5 until the cross-validation cycle is completed. Calculate the average stratigraphic weighted accuracy of each probability map stratigraphic model in the cross-validation and use it as an evaluation index to measure the model's generalization ability, thus realizing intelligent model selection.
[0009] Preferably, in step S1, based on the borehole mileage and elevation range, the upper and lower boundaries of the formation modeling space are obtained using the borehole top and bottom elevations based on Kriging interpolation, and the formation modeling space is determined; the interval of the grid in the length and depth directions is preset, and the formation modeling space is gridded based on this interval size; based on the borehole mileage location and the elevation range of each formation, the formation data of the borehole is filled into the corresponding grid.
[0010] Preferably, the formation data of the borehole includes soil types / lithologies such as silty clay, silt, fine sand, silt, coarse sand, gravelly soil, tuff, and argillaceous sandstone.
[0011] Preferably, step S2 specifically includes the following: S21. Markov Random Field Construction: Based on the neighborhood system, construct the conditional probability distribution of soil type / lithology. According to the Hammersly-Clifford theorem, the conditional probability distribution at each grid is as follows: Where exp is the exponential function; Z is the normalization constant; The energy function is calculated based on the grid state, the state of the neighboring system, and the weights in each direction. The weights in each direction for the four-neighboring system are... The weights of each direction in the eight-neighborhood system are: ; S22. Construction of Coupled Markov Chains: Constructing a conditional probability distribution based on Markov property soil type / lithology at each grid: ;in The vertical transition probability matrix VTPM is calculated using the vertical transition count matrix obtained from known borehole data. The horizontal transition probability matrix HTPM is calculated by multiplying the main diagonal elements of the vertical transition counting matrix by K, with reference to the Walter phase law, where K is the factor that expands the main diagonal elements. For soil type / lithology quantity, The number of grid divisions in the longitudinal direction for modeling the geological strata. For the soil type / lithology of this grid, , , These represent the soil type / lithology of the top, left, and rightmost grids, respectively; the denominator in this formula is a normalization term. f For any soil type / lithology.
[0012] Preferably, in step S3, the borehole dataset is divided into an outer test set and an outer training set based on k-fold cross-validation. That is, the borehole dataset is divided into k mutually exclusive subsets of equal size. In each round of cross-validation, one subset is used as the outer test set for subsequent evaluation of the model's generalization ability in the outer layer; the remaining k-1 subsets are used as the outer training set. In each round of cross-validation, the outer training set is passed into the inner layer, and the outer training set is used as known data for parameter learning in the inner layer.
[0013] Preferably, in step S4, the parameter learning method is used to learn parameters for three probabilistic map stratigraphic models, wherein the parameter learning method for the 4-neighborhood system and the 8-neighborhood system Markov random field is the same. Markov random field parameter learning methods include: using borehole data from the outer training set as known data, introducing Bayesian optimization, and using the likelihood function value as an indicator for parameter learning; the parameter learning objects for 4-neighborhood systems and 8-neighborhood systems are respectively... , ; The coupled Markov chain parameter learning method includes: the parameter learning object of the coupled Markov chain is the K value in the Walter phase law; under the given K value, VTPM and HTPM are calculated and the formation is simulated; HTPM' is calculated based on the formation simulation results; the coupled Markov chain uses the mean square error of HTPM and HTPM' as the parameter evaluation index, and the parameter learning is performed by traversing the parameter value space.
[0014] Preferably, in step S5, the method for calculating the formation weighted accuracy specifically includes: in each round of cross-validation, the outer training set is used as known data, and each probabilistic map formation model performs formation simulation based on the known data and the optimal parameters learned from the inner layer. The outer test set is used as real data, and the formation weighted accuracy at the test set location is calculated. Specifically, the prediction accuracy is first calculated for each formation. Then, a weighted average of the prediction accuracy rates for each stratum is calculated, yielding the stratum-weighted accuracy rate. Engineers can assign weights to different strata based on their specific distribution.
[0015] The beneficial effects of this invention are: the method of this invention realizes intelligent comparison and selection of probabilistic map stratigraphic models driven by site survey data, effectively avoiding prediction errors caused by selecting stratigraphic models based on subjective experience. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the process steps of the intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture in an embodiment of the present invention; Figure 2 This is a schematic diagram of a typical virtual two-dimensional geological profile selected in an embodiment of the present invention, where different gray levels represent different soil types. (a) is the selected virtual two-dimensional geological profile, and (b) is the initialization result of the stratum modeling space. Figure 3 A schematic diagram of the formation simulation results of the 8-neighborhood Markov random field based on the virtual borehole and optimal parameters in the embodiment; Figure 4 This is a schematic diagram of the formation simulation results of the Markov random field for the 4-neighborhood system based on the virtual borehole and optimal parameters in the embodiment; Figure 5 This is a schematic diagram of the formation simulation results based on the virtual borehole and optimal parameters in the embodiment, used to couple the Markov chain. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] This invention provides an intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture under a typical virtual two-dimensional geological profile. The virtual two-dimensional geological profile in this embodiment conforms to the actual stratigraphic distribution characteristics. The virtual two-dimensional geological profile used in this embodiment is an illustrative example and does not constitute a limitation on the scope of patent protection. The specific method steps are as follows: Figure 1 As shown, it includes the following: Step S1: Construct a Kriging-based method for initializing the stratigraphic modeling space. Initialize the stratigraphic modeling space according to this method and the site survey data (drilling data). Grid the stratigraphic modeling space and fill the corresponding grid with the drilling data. Based on the borehole mileage and elevation range, the upper and lower boundaries of the stratigraphic modeling space are obtained using the borehole top and bottom elevations, based on Kriging interpolation, thus defining the stratigraphic modeling space. The stratigraphic data from the boreholes specifically includes, but is not limited to, soil types / lithologies such as silty clay, silt, fine sand, silt, coarse sand, gravelly soil, tuff, and argillaceous sandstone. This part implements automated stratigraphic modeling space construction, stratigraphic modeling space meshing, and borehole stratigraphic data filling.
[0019] The virtual two-dimensional geological profile used in this embodiment is as follows: Figure 2 As shown in (a), five sets of virtual boreholes are extracted from the positions indicated by the dashed lines. The mileage range of the five sets of virtual boreholes is 0m-20m, and the elevation range is -10m-0m. Using the top and bottom elevations of the virtual boreholes, the upper and lower boundaries of the stratum modeling space are obtained based on Kriging interpolation. Since the top elevation of the boreholes in this virtual case is 0m and the bottom elevation is -10m, the upper and lower boundaries of the stratum modeling space are straight lines with elevations of 0m and -10m. If there are fluctuations in the top and bottom elevations of the boreholes, the upper and lower boundaries of the stratum modeling space are curves obtained by Kriging interpolation. The stratum modeling space in this virtual case is a rectangular space with a length of 20m and a height of 10m, containing three soil types: "clay", "silt", and "coarse sand". Then, the stratum modeling space is gridded according to a rectangular grid with a length of 1m and a depth of 0.1m, and the five sets of virtual boreholes are filled into the gridded stratum modeling space, as shown below. Figure 2 As shown in (b).
[0020] Step S2: Based on the initial formation modeling space and borehole data from Step S1, construct probabilistic graphical formation models on the formation modeling space, including: 4-neighborhood system Markov random field, 8-neighborhood system Markov random field, and coupled Markov chain.
[0021] Based on the initial formation modeling space and borehole data from step S1, a probabilistic map formation model is constructed. The Markov random field construction process is as follows: Conditional probability distributions regarding soil type / lithology are constructed based on the neighborhood system. According to the Hammersly-Clifford theorem, the conditional probability distribution at each grid is: Where Z is the normalization constant; The energy function is calculated based on the grid state, the state of the neighboring system, and the weights in each direction. The weights in each direction for the four-neighboring system are... The weights of each direction in the eight-neighborhood system are: The construction process of the coupled Markov chain is as follows: at each grid, a conditional probability distribution based on Markov property soil type / lithology is constructed: ;in The Vertical Transition Probability Matrix (VTPM) is calculated from the vertical transition count matrix obtained by statistically analyzing known borehole data. The horizontal transition probability matrix (HTPM) is calculated by multiplying the main diagonal elements of the vertical transition counting matrix by K, referencing the Walter phase law.
[0022] Step S3: The dual-layer optimization architecture consists of an outer layer nested within an inner layer. Based on the formation modeling space and borehole data from step S1, a method for dividing and using the dataset for the dual-layer optimization architecture is constructed. The dataset composed of borehole data is divided into an outer test set and an outer training set based on cross-validation. The outer training set is then passed into the inner layer for parameter learning.
[0023] In the outer layer, the borehole dataset is divided into an outer test set and an outer training set based on cross-validation. Taking k-fold cross-validation as an example, the borehole dataset is divided into k mutually exclusive subsets of equal size. In each round of cross-validation, one subset is used as the outer test set for subsequent evaluation of the model's generalization ability in the outer layer, while the remaining k-1 subsets are used as the outer training set. In each round of cross-validation, the outer training set is fed into the inner layer, where the outer training set is used as known data for parameter learning.
[0024] In this embodiment, based on the virtual borehole data, the borehole dataset is divided into five mutually exclusive subsets using leave-one-out cross-validation, where each subset contains one virtual borehole. In leave-one-out cross-validation, one virtual borehole is sequentially set as the outer test set, and the remaining four virtual boreholes are used as the outer training set. Subsequently, in each round of cross-validation, the outer training set is passed to the inner layer, where the outer training set is used as known data for parameter learning.
[0025] Step S4: In the inner layer, the outer layer training set divided in step S3 is used as known data, and parameter learning is performed based on the probabilistic map stratigraphic model constructed in step S2 and the corresponding parameter learning method.
[0026] The Markov random field parameter learning method involves using the borehole data from the outer training set as known data, incorporating Bayesian optimization, and using the likelihood function value as the metric for parameter learning. The Markov random field parameter learning objects for the 4-neighborhood and 8-neighborhood systems are respectively... , The parameter learning object of the coupled Markov chain is the K value in the Walter phase law. Given the K value, VTPM and HTPM are calculated, and formation simulation is performed. Based on the formation simulation results, HTPM' is calculated. The coupled Markov chain uses the mean square error of HTPM and HTPM' as the parameter evaluation index, and the parameters are learned by traversing the parameter value space.
[0027] In a 4-neighborhood Markov random field system, the parameter values range as follows: , The optimal parameters in the first round of cross-validation are: , The optimal parameters in the second round of cross-validation are: , The optimal parameters in the third round of cross-validation are: , The optimal parameters in the fourth round of cross-validation are: , The optimal parameters in the fifth round of cross-validation are: , .
[0028] In an 8-neighborhood Markov random field system, the parameter values range as follows: , , , The optimal parameters in the first round of cross-validation are: , , , The optimal parameters in the second round of cross-validation are: , , , The optimal parameters in the third round of cross-validation are: , , , The optimal parameters in the fourth round of cross-validation are: , , , The optimal parameters in the fifth round of cross-validation are: , , , .
[0029] In a coupled Markov chain, the parameter values can take the following ranges: The optimal parameters in the first round of cross-validation are: The optimal parameters in the second round of cross-validation are: The optimal parameters in the third round of cross-validation are: The optimal parameters in the fourth round of cross-validation are: The optimal parameters in the fifth round of cross-validation are: .
[0030] Step S5: In the outer layer, based on the optimal parameters learned in step S4, the outer layer training set in step S3 is used as known data to perform formation simulation based on the probabilistic map formation model constructed in step S2. A formation weighted accuracy calculation method is proposed, and the formation weighted accuracy is calculated based on the formation simulation results and the borehole data of the outer layer test set.
[0031] In this embodiment, each round of cross-validation uses the outer training set consisting of four virtual boreholes as known data. Based on the optimal parameters learned in S4, formation simulation is performed, and the formation simulation results are calculated at the location of the outer test set. Here, it is assumed that the silt layer in the virtual case may exhibit pinch-out characteristics; accurate identification of the silt layer is a key aspect of formation modeling. Therefore, the weights of the three formations are: clay 1, silt 2, and coarse sand 1.
[0032] Step S6: Repeat steps S4 and S5 until the cross-validation cycle is completed. Calculate the average stratigraphic weighted accuracy of each probability map stratigraphic model in the cross-validation, and use this as an evaluation index to measure the model's generalization ability, thereby realizing intelligent model selection.
[0033] The stratigraphic-weighted accuracy of the 4-neighborhood Markov random field in each round of cross-validation is: The average formation-weighted accuracy (model evaluation index) is 0.848.
[0034] The stratigraphic-weighted accuracy of the 8-neighborhood Markov random field system in each round of cross-validation is: The average formation-weighted accuracy (model evaluation index) is 0.918.
[0035] The formation-weighted accuracy of the coupled Markov chain in each round of cross-validation is: The average formation-weighted accuracy (model evaluation index) is 0.874.
[0036] The calculation results show that, in this embodiment, the 8-neighborhood Markov random field has the strongest model generalization ability, followed by the coupled Markov chain, and the 4-neighborhood Markov random field has the weakest. Therefore, the 8-neighborhood Markov random field is selected as the stratigraphic model for the subsequent stratigraphic modeling process in this virtual two-dimensional geological profile.
[0037] To further verify the effectiveness of the proposed intelligent selection method for probabilistic graphical formation models based on a two-layer optimization architecture, we use 8-neighborhood Markov random fields, 4-neighborhood Markov random fields, and coupled Markov chains to perform parameter learning and formation simulation based on five sets of borehole data.
[0038] The parameter learning results for the 8-neighborhood Markov random field system are as follows: , , , The formation simulation results based on the optimal parameters are as follows: Figure 3 As shown. The parameter learning results for the 4-neighborhood system Markov random field are as follows. , The formation simulation results based on the optimal parameters are shown in the figure. The parameter learning result of the coupled Markov chain is K=4.5, and the formation simulation results based on the optimal parameters are as follows. Figure 5 As shown.
[0039] according to Figures 2-5 It can be seen that the 8-neighborhood Markov random field can effectively identify the dip angle and stratigraphic distribution based on virtual borehole data. The figure clearly shows that the 8-neighborhood Markov random field has better stratigraphic identification ability than the coupled Markov chain, which in turn outperforms the 4-neighborhood Markov random field. This result fully demonstrates the rationality and effectiveness of the proposed method.
[0040] The results show that the method proposed in this application reasonably and effectively realizes data-driven intelligent model selection. The proposed method effectively avoids errors caused by subjective experience, providing engineers with a reasonable and objective reference when using probabilistic chart stratigraphic models.
[0041] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0042] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.
[0043] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0044] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."
[0045] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligent selection of probabilistic graphical stratigraphic models based on a two-layer optimization architecture, characterized in that, Includes the following steps: S1. Construct a formation modeling space initialization method based on Kriging. Initialize the formation modeling space according to the method and the borehole data of the work site. Grid the formation modeling space and fill the corresponding grid with the borehole data. S2. Based on the initial formation modeling space and borehole data in step S1, construct probabilistic graphical formation models on the formation modeling space, including: 4-neighborhood system Markov random field, 8-neighborhood system Markov random field, and coupled Markov chain. S3. Based on the formation modeling space and borehole data in step S1, construct a dataset partitioning and usage method for the two-layer optimization architecture. The two-layer optimization architecture consists of an outer layer nested with an inner layer. Based on cross-validation, the dataset composed of borehole data is divided into an outer test set and an outer training set, where the outer training set is passed into the inner layer for parameter learning. S4. In the inner layer, the outer training set divided in step S3 is used as known data, and parameter learning is performed based on the probabilistic map stratigraphic model constructed in step S2 and the corresponding parameter learning method. S5. In the outer layer, based on the optimal parameters learned in step S4, the outer layer training set in step S3 is used as known data to perform formation simulation based on the probabilistic map formation model constructed in step S2. A formation weighted accuracy calculation method is proposed, and the formation weighted accuracy is calculated based on the formation simulation results and the borehole data of the outer layer test set. S6. Repeat steps S4 and S5 until the cross-validation cycle is completed. Calculate the average stratigraphic weighted accuracy of each probability map stratigraphic model in the cross-validation and use it as an evaluation index to measure the model's generalization ability, thus realizing intelligent model selection.
2. The intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 1, characterized in that: In step S1, based on the mileage and elevation range of the borehole, the top and bottom elevations of the borehole are used to obtain the upper and lower boundaries of the formation modeling space using Kriging interpolation, and the formation modeling space is determined; the interval of the grid in the length and depth directions is preset, and the formation modeling space is gridded based on this interval size. Based on the borehole mileage location and the elevation range of each stratum, the stratum data of the borehole is filled into the corresponding grid.
3. The intelligent selection method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 2, characterized in that: The geological data from the boreholes include soil types / lithologies such as silty clay, silt, fine sand, silt, coarse sand, gravelly soil, tuff, and argillaceous sandstone.
4. The intelligent selection method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 1, characterized in that: Step S2 specifically includes the following: S21. Markov Random Field Construction: Based on the neighborhood system, construct the conditional probability distribution of soil type / lithology. According to the Hammersly-Clifford theorem, the conditional probability distribution at each grid is as follows: Where exp is the exponential function; Z is the normalization constant; The energy function is calculated based on the grid state, the state of the neighboring system, and the weights in each direction. The weights in each direction for the four-neighboring system are... The weights of each direction in the eight-neighborhood system are: ; S22. Construction of Coupled Markov Chains: Constructing a conditional probability distribution based on Markov property soil type / lithology at each grid: ;in The vertical transition probability matrix VTPM is calculated using the vertical transition count matrix obtained from known borehole data. The horizontal transition probability matrix HTPM is calculated by multiplying the main diagonal elements of the vertical transition counting matrix by K, referencing the Walter phase law.
5. The intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 1, characterized in that: In step S3, the borehole dataset is divided into an outer test set and an outer training set based on k-fold cross-validation. That is, the borehole dataset is divided into k mutually exclusive subsets of equal size. In each round of cross-validation, one subset is used as the outer test set to evaluate the model's generalization ability in the outer layer. The remaining k-1 subsets are used as the outer training set. In each round of cross-validation, the outer training set is passed to the inner layer, and the outer training set is used as known data for parameter learning in the inner layer.
6. The intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 1, characterized in that: In step S4, the parameter learning method is used to learn parameters for three probabilistic map stratigraphic models, with the parameter learning method for the 4-neighborhood system and the 8-neighborhood system being the same. Markov random field parameter learning methods include: using borehole data from the outer training set as known data, introducing Bayesian optimization, and using the likelihood function value as an indicator for parameter learning; the parameter learning objects for 4-neighborhood systems and 8-neighborhood systems are respectively... , ; The coupled Markov chain parameter learning method includes: the parameter learning object of the coupled Markov chain is the K value in the Walter phase law; under the given K value, VTPM and HTPM are calculated and the formation is simulated; HTPM' is calculated based on the formation simulation results; the coupled Markov chain uses the mean square error of HTPM and HTPM' as the parameter evaluation index, and the parameter learning is performed by traversing the parameter value space.
7. The intelligent comparison method for probabilistic graphical stratigraphic models based on a two-layer optimization architecture as described in claim 1, characterized in that: In step S5, the method for calculating the formation weighted accuracy specifically includes: in each round of cross-validation, the outer training set is used as known data, and each probabilistic map formation model performs formation simulation based on the known data and the optimal parameters learned from the inner layer. The outer test set is used as real data, and the formation weighted accuracy at the test set location is calculated. Specifically, the prediction accuracy is first calculated for each formation. Then, the weighted average of the prediction accuracy of each stratum is calculated to obtain the stratum weighted accuracy.