A 3D Formation Modeling Method for Drilling Strategy Optimization

Abstract

This invention discloses a three-dimensional stratigraphic modeling method for borehole strategy optimization, comprising the following steps: integrating discrete standard penetration test data, continuous static cone penetration test data, predicted unsampled location data, and existing stratigraphic data into multi-confidence geological exploration data; generating virtual boreholes with stratigraphic rationality using the GA-Stacking machine learning algorithm based on the multi-confidence geological exploration data; performing stratigraphic attribute modeling and interface reconstruction using the radial basis function implicit modeling method based on the original exploration data and virtual borehole data; and conducting uncertainty assessment and borehole layout optimization based on the constructed stratigraphic model. This invention addresses the impact of the sparsity of the original exploration data and the randomness of the modeling process on the accuracy of stratigraphic modeling by proposing an ensemble learning implicit stratigraphic modeling method based on multi-source sparse geological exploration data, providing an effective approach for refined stratigraphic modeling and a scientific basis for geological exploration borehole layout.

Description

Technical Field

[0001] This invention relates to the field of formation modeling technology, and in particular to a three-dimensional formation modeling method for borehole strategy optimization. Background Technology

[0002] The scarcity of urban land resources and advancements in underground construction technology have driven the rapid development and utilization of underground space. Compared to surface engineering, underground engineering is constructed within rock and soil masses, characterized by higher costs, more complex construction environments, and higher quality requirements. Since the rock and soil masses are buried underground, engineers cannot directly observe their distribution and properties. Conducting engineering geological exploration to determine the composition and engineering properties of each layer of rock and soil, and establishing accurate stratigraphic models, is the foundation for underground engineering planning and design, and also a guarantee of construction safety.

[0003] Stratigraphic modeling is often constrained by many factors: (1) the uncertainty of stratigraphic properties. Soil and rock masses have undergone historical weathering, transportation, deposition, and metamorphism, and have been subjected to various natural and human activities, forming a natural temperament of "a thousand soils, a thousand different kinds," and their composition and mechanical parameters have strong spatial variability; (2) geological exploration data, as the original input parameters for stratigraphic division, has strong sparsity and discontinuity. The sample size of soil and rock masses obtained by geological exploration is usually less than one millionth of the volume of the area affected by underground engineering construction. The uncertainty of stratigraphic properties and the sparsity of geological exploration data lead to great uncertainty in the stratigraphic modeling process.

[0004] Traditional methods for constructing stratigraphic models primarily involve engineers manually connecting adjacent stratigraphic boundaries based on collected geological exploration data to create stratigraphic profiles for engineering design, and assigning corresponding property parameters to each stratum. This method requires extensive human-computer interaction and is limited by the sparse and widely spaced exploration points; a single linear connection cannot accurately reflect the actual distribution of strata in adjacent areas. To address these issues, researchers have proposed implicit modeling methods. These methods, based on geostatistics and spatial interpolation theory, automatically and quickly construct implicit surface models of stratigraphy based on geological exploration data, without frequent manual intervention. However, the performance of implicit modeling is susceptible to the distribution of the original data. When the exploration intervals are too large, traditional interpolation methods struggle to accurately capture local variations in stratigraphic details, resulting in low local modeling accuracy.

[0005] In recent years, machine learning methods have been increasingly widely applied in geotechnical engineering. Leveraging their powerful nonlinear fitting capabilities, they can uncover latent patterns in data, integrate and analyze complex correlations between engineering parameters, effectively improving the limitations of information processing and analytical capabilities in traditional stratigraphic division, and achieving rapid and accurate stratigraphic division. Several inventions have demonstrated the effectiveness of machine learning methods in stratigraphic modeling. Kumar et al. used a fuzzy multilayer perceptron (F-MLP) model, combining borehole data with artificially synthesized data, to achieve "data-driven" interpolation of stratigraphic categories between boreholes. Shi and Wang, based on limited borehole data and training images, proposed a data-driven extreme gradient boosting (XGB) model for interpolating stratigraphic profiles. Abedi et al. applied Support Vector Machine (SVM) to ore deposit exploration modeling; Guo Jiateng et al. used SVM and Back Propagation Neural Network (BP) to achieve implicit 3D geological modeling; Wang Hao et al. used BP neural networks to predict mineral distribution around ore deposits; Zhu Junsheng et al. improved the accuracy of thin-layer strata prediction using an improved K-Nearest Neighbor (KNN) algorithm. Although these machine learning methods have played a role in assisting stratigraphic modeling, how to effectively improve the generalization ability and classification accuracy of the models remains one of the urgent problems to be solved in stratigraphic classification modeling. Summary of the Invention

[0006] Purpose of the invention: To overcome the shortcomings of the prior art, this invention provides a three-dimensional stratigraphic modeling method for borehole strategy optimization. This method uses a genetic algorithm to construct an optimized ensemble learning model, learns the stratigraphic distribution characteristics based on geological exploration data, generates a stratigraphic classification dataset of virtual boreholes, and establishes a stratigraphic model based on the radial basis function modeling method. On this basis, information entropy is introduced to quantify the uncertainty of the modeling process and optimize the borehole layout scheme.

[0007] Technical solution: The three-dimensional formation modeling method for borehole strategy optimization disclosed in this invention includes the following steps: S1. Integrate discrete standard penetration test data, continuous static cone penetration test data, predicted unsampled location data, and existing stratigraphic data into multi-reliability geological exploration data. S2. Based on multi-belief geological exploration data, the GA-Stacking machine learning algorithm is used to generate virtual boreholes with geological rationality. S3. Based on the original exploration data and virtual borehole data, the radial basis function implicit modeling method is used to model the formation properties and reconstruct the interface. S4. Based on the constructed formation model, perform uncertainty assessment and borehole layout optimization.

[0008] Furthermore, the unsampled location data predicted in S1 is obtained by using a machine learning algorithm with continuous static penetration test data as input features to train the model to predict the stratigraphic type and engineering property parameters of the unsampled locations in the standard penetration test.

[0009] Furthermore, S2 specifically includes: S2-1. Design the GA-Stacking model framework, including the basic Stacking structure and genetic algorithm optimization strategy; S2-2, Model training and parameter optimization for the GA-Stacking model framework; S2-3. The optimized GA-Stacking model is applied to the study area. Based on multi-belief geological data, stratigraphic information in sparse areas is predicted, and virtual boreholes are generated. Each virtual borehole contains three-dimensional coordinates, stratigraphic category labels and corresponding probability distributions to ensure that its stratigraphic sequence conforms to the regional sedimentary patterns.

[0010] Furthermore, when generating virtual boreholes in S2-3, the generated boreholes conform to the stratigraphic regularity in the vertical direction and adapt to the structural trend in the horizontal direction; the stratigraphic statistical variability function is introduced as a spatial correlation constraint to ensure that the generated virtual points maintain reasonable spatial autocorrelation with known points; and the generated results can be screened and corrected by combining expert knowledge rules.

[0011] Furthermore, S3 specifically includes: S3-1, Stratigraphic data preprocessing and control point extraction; The original exploration data and virtual borehole data were systematically analyzed. The stratigraphic data of each exploration point were vertically sequenced according to depth. The elevation of the upper and lower interfaces of continuous stratigraphic units was identified, and a complete dataset of stratigraphic interface control points was constructed. S3-2, Calculation of Boundary and Mesh Generation; The computational boundary of the study area was determined, and a buffer zone was set to eliminate boundary effects. An equal-spacing grid partitioning strategy was adopted, and the grid resolution was set in combination with the range of the study area and the characteristics of the stratigraphic thickness to provide dense computational support points for surface interpolation and ensure the accuracy of interface reconstruction. S3-3, Radial basis function interpolation modeling; The formation interface surface is constructed using the radial basis function algorithm; S3-4, Modeling accuracy verification; The accuracy of the model is quantified by measuring the fit between the thinning test and the profile.

[0012] Furthermore, S3-3 includes: Interpolation function definition; For a given set of stratigraphic interface control points, the interpolation function takes the form: , Where: Φ is the kernel function, λ i Let ||(x,y)-(x) be the coefficients to be determined. i ,y i || is the Euclidean distance norm; Regularization optimization: Regularization techniques are introduced to balance interpolation accuracy and surface smoothness; Model generation: Based on the interpolated stratigraphic interface, a three-dimensional stratigraphic entity model is constructed to realize the visual reconstruction of multiple types of stratigraphy.

[0013] Furthermore, S4 includes: S4-1. Information entropy is used as an uncertainty evaluation index to quantitatively describe the local reliability of the formation model; For each formation unit in a virtual borehole, the entropy value H = −∑k=1mpklogpk, where pk is the probability that the unit belongs to the k-th formation category, and m is the total number of formation categories. The entropy value ranges from 0 to 1. The smaller the entropy value, the closer the formation distribution is to the actual situation, and the lower the uncertainty. The larger the entropy value, the higher the uncertainty, and the formation distribution may have a large error. The three-dimensional distribution of information entropy in the study area is obtained by interpolation. S4-2. Based on the uncertainty quantification results, formulate a borehole layout strategy; Calculate the average entropy value of virtual boreholes in each region, and determine the region with the largest average entropy value as the first new borehole location. After a new borehole is added, set the information entropy of that location to 0, indicating that the stratigraphic information has been determined through exploration. Recalculate the average entropy value of the remaining regions to determine the next new borehole location. Iterate the above process until the uncertainty is reduced to an acceptable range for the project. S4-3. Verify the optimization effect.

[0014] Beneficial effects: Compared with the prior art, the significant advantages of this invention are: 1. Process discrete, non-uniform, and different-sourced raw geological exploration data into continuous and uniform fused data, thereby improving modeling accuracy and reducing errors at the data level.

[0015] 2. The model demonstrates good accuracy and stability in stratigraphic classification and prediction. It exhibits higher classification accuracy compared to single machine learning models.

[0016] 3. The virtual borehole data generated by the model can serve as an effective constraint for the implicit 3D modeling of the formation, and the constructed formation model can more accurately depict the local variation characteristics of the formation.

[0017] 4. Information entropy theory is introduced to quantify the uncertainty of the model. The magnitude of uncertainty is used as the basis for selecting new boreholes, providing an effective reference for the borehole layout plan in the later stages of the project. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the multi-confidence geological data fusion of the present invention; Figure 2 This is a schematic diagram of the implicit modeling framework for radial basis functions of the present invention; Figure 3 This is a virtual borehole distribution diagram of the present invention; Figure 4 This is a schematic diagram of the modeling results of the present invention; Figure 5 This is a schematic diagram illustrating the uncertainty distribution of the present invention; Figure 6 This is a schematic diagram showing the newly added drilling locations in this invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] A three-dimensional formation modeling method for borehole strategy optimization includes the following steps: S1. Fusion of multi-reliability geological exploration data; The main methods of geotechnical engineering geological exploration are field drilling and testing. Testing includes the Standard Penetration Test (SPT), the Cone Penetration Test (CPT), and laboratory tests. SPT is conducted based on exploratory boreholes to obtain soil and rock samples at different depths, and then laboratory tests are used to determine the composition and engineering properties of the soil and rock. Its core advantage lies in the high visibility of soil samples and the high reliability of data, making it a high-reliability data source. However, its drawbacks include sparse measurement points and significant data dispersion (typically a soil sample is obtained every 2-3 m), high testing costs, and the potential for disturbances during sampling and testing to affect the final results. Accurate stratigraphic division is difficult to achieve relying solely on SPT and laboratory test data. CPT's main advantages are its speed, economy, and dense data acquisition (sampling interval of 0.1 m), making it superior to the Standard Penetration Test in terms of continuous stratigraphic profile drawing and cost-effectiveness. Its limitations include the inability to obtain soil samples, lack of intuitive verification, and its greater suitability for fine-grained soils (such as silt and clay). Therefore, in practical engineering, CPT data is often regarded as an exploration source with relatively low reliability.

[0021] Machine learning algorithms are employed to fill the "spatial gaps" in SPT data: Continuous, dense CPT data (such as cone tip drag and sidewall friction) are used as input features to train a model that predicts stratigraphic categories and engineering properties at unsampled SPT locations. This data-driven process establishes a correlation between low-confidence dense data and high-confidence sparse data, achieving "filling the gaps with density." Discrete SPT data, continuous CPT data, predicted unsampled data, and existing stratigraphic data are integrated into multi-confidence geological survey data, ultimately forming a continuous sampling point dataset with three-dimensional coordinates (x, y, z) and stratigraphic category labels. The integrated data retains the high reliability of SPT while leveraging the density of CPT to address data discrepancies, reducing modeling uncertainty from the outset. Figure 1 As shown, this method effectively addresses the discrete distribution of borehole records while taking into account existing geological conditions, thereby improving modeling accuracy at the data level and reducing model uncertainty.

[0022] S2 and GA-Stacking algorithms generate virtual boreholes to construct reasonable supplementary constraints for the formation; GA-Stacking Model Framework Design This model combines genetic algorithms (GA) with stacking ensemble learning. Its core principle is to achieve "high-precision formation classification prediction" through algorithm optimization, providing a reliable basis for virtual borehole generation. Stacking infrastructure (two-layer modeling): The first layer (base learner layer): The multi-belief integrated data is divided into training set and validation set. Seven classic machine learning models are selected as candidate base learners (XGB, LGB, KNN, LR, RF, DT, SVM). After each base learner is trained independently, it generates predicted values ​​for the input samples to form a new feature matrix.

[0023] The second layer (meta-learner layer): The prediction results of the base learners are used as input to train the meta-learner. By minimizing the cross-entropy loss function, the optimal combination relationship of the base learners is learned, and the final stratum classification result is output.

[0024] Genetic Algorithm Optimization Strategy: Encoding method: Encode "whether to select a certain base learner" into a binary vector (e.g., "1" means to select, "0" means not to select), forming an individual chromosome.

[0025] Fitness function: The fitness metric is the weighted F1 score of the model on the validation set, which measures the generalization performance of the combination of base learners.

[0026] Evolutionary operations: Set the population size to twice the number of base learners, 15 generations, mutation rate of 0.15, and elite retention ratio of 10%-20%; iterate to find the optimal combination of base learners through selection (retaining individuals with high fitness), crossover (recombining individual genes), and mutation (randomly changing gene positions).

[0027] Model training and parameter optimization Model training was implemented using Python 3.8 and the Scikit-Learn library, and overfitting was reduced through cross-validation. The hyperparameters of each base learner were determined by random search (e.g., alpha=0.2 and gamma=0.1 for XGB, and subsample=0.8 for LGB).

[0028] The initial population consists of a random combination of 3-6 different base learners to ensure population diversity. After 15 generations of evolution, the optimal combination is obtained by convergence: LGB, XGB, and DT are used as base learners, and LGB is used as a meta-learner.

[0029] Virtual borehole generation, such as Figure 3 As shown: The optimized GA-Stacking model was applied to the study area. Based on multi-belief geological data, stratigraphic information in sparsely populated areas was predicted, generating virtual boreholes. Each virtual borehole contains three-dimensional coordinates (x, y, z), stratigraphic category labels, and corresponding probability distributions, ensuring that its stratigraphic sequence conforms to regional sedimentary patterns (consistent with the category logic of actual strata). In the example, 25 virtual boreholes were generated at 50m intervals to fill the information gaps between the original boreholes (14) and CPT boreholes (12). When generating virtual boreholes, the generated boreholes conform to stratigraphic patterns vertically and adapt to tectonic trends horizontally. A stratigraphic statistical variogram was introduced as a spatial correlation constraint to ensure that the generated virtual points maintain reasonable spatial autocorrelation with known points. Furthermore, the generated results can be filtered and corrected using expert knowledge rules.

[0030] S3, Radial Basis Function Implicit Modeling - Achieving Fine Reconstruction of Formation Interfaces and Entities, such as Figure 2 As shown; Stratigraphic data preprocessing and control point extraction A systematic analysis was conducted on the original exploration data (14 boreholes + 12 CPT holes) and 25 virtual borehole data: the stratigraphic data of each exploration point were vertically sequenced according to depth, the upper and lower interface elevations of continuous stratigraphic units were identified, and a complete stratigraphic interface control point dataset (containing the spatial location boundary information of each stratum) was constructed.

[0031] Computational Boundary and Mesh Generation Determine the computational boundary of the study area and set up a buffer to eliminate boundary effects (the buffer size is adaptively adjusted according to the area range). An equidistant grid partitioning strategy was adopted. Based on the study area and the characteristics of the stratigraphic thickness, the grid resolution was set to 10m×10m×0.1m (x×y×z) to provide dense computational support points for surface interpolation and ensure the accuracy of interface reconstruction.

[0032] Radial basis function interpolation modeling, such as Figure 4 As shown: The core process of constructing the formation interface surface using the radial basis function algorithm is as follows: Interpolation function definition: For a given set of stratigraphic interface control points, the interpolation function takes the form: , Where: Φ is the kernel function, λ i Let ||(x,y)-(x) be the coefficients to be determined. i ,y i || is the Euclidean distance norm; Regularization optimization: Introducing regularization techniques to balance interpolation accuracy and surface smoothness; Model generation: Based on the interpolated stratigraphic interface, a three-dimensional stratigraphic solid model is constructed to realize the visual reconstruction of eight types of strata (miscellaneous fill, clay, silty clay, etc.), ensuring both mathematical smoothness and conformity to the stratigraphic deposition law.

[0033] Modeling accuracy verification The accuracy of the model is quantified by measuring the fit between the thinning test and the profile: Average error: The average error between the stratigraphic distribution generated by the model and the actual stratigraphic distribution is 0.23m, which is 0.24m less than that of modeling with only the original data (0.47m) and 0.40m less than that of the kriging method (0.63m). Profile fit: Defined as "profile fit = (overlapping area between model profile and actual profile / total area of ​​actual profile) × 100%", the average fit of each stratum is 84%, with the silt 1 stratum having the highest fit (88%), which is significantly better than using only the original data (79%) and the Kriging method (72%).

[0034] S4. Uncertainty Assessment and Drilling Layout Optimization - Providing Scientific Drilling Layout Guidance; Uncertainty Quantification Based on Information Entropy Information entropy is used as an uncertainty evaluation index to quantitatively describe the local reliability of the formation model: Information entropy definition: For each formation unit of a virtual borehole, the entropy value H = −∑k=1mpklogpk, where pk is the probability that the unit belongs to the k-th formation, and m is the total number of formation categories (8 categories). Entropy value meaning: The value ranges from 0 to 1. The smaller the entropy value, the closer the stratigraphic distribution is to the actual situation and the lower the uncertainty. The larger the entropy value, the higher the uncertainty and the greater the possibility of a large error in the stratigraphic distribution. Uncertainty Distribution: The three-dimensional distribution of information entropy in the study area was obtained through interpolation. It was found that uncertainty is mainly concentrated at both ends (shallow and deep layers) and in the central region. These areas pose engineering safety hazards due to the lack of original boreholes. Figure 5 As shown.

[0035] New borehole layout optimization Based on the uncertainty quantification results, a borehole layout strategy prioritizing high uncertainty is formulated: Calculate the average entropy value of the virtual boreholes in each region, and determine the region with the largest average entropy value as the location of the first new borehole; After a new borehole is drilled, the information entropy of that location is set to 0 (indicating that the stratigraphic information has been determined through exploration), the average entropy value of the remaining area is recalculated, and the location of the next new borehole is determined. Iterate through the above process until the uncertainty is reduced to an acceptable range for the engineering.

[0036] Optimization effect verification, such as Figure 6 As shown: In this example, 5 new holes were added in 5 iterations, and the optimization results are as follows: The initial average entropy of the region decreased from 0.110 to 0.073, and the total uncertainty decreased by 34.5%. All new boreholes are located in sparse areas of the original boreholes (without existing exploration data) and are in areas with concentrated entropy values ​​and high uncertainty, which meets the actual needs of the project and provides quantitative guidance for the layout of subsequent geological exploration boreholes.

Claims

1. A three-dimensional formation modeling method for borehole strategy optimization, characterized in that, Includes the following steps: S1. Integrate discrete standard penetration test data, continuous static cone penetration test data, predicted unsampled location data, and existing stratigraphic data into multi-reliability geological exploration data. S2. Based on multi-belief geological exploration data, the GA-Stacking machine learning algorithm is used to generate virtual boreholes with geological rationality. S3. Based on the original exploration data and virtual borehole data, the radial basis function implicit modeling method is used to model the formation properties and reconstruct the interface. S4. Based on the constructed formation model, perform uncertainty assessment and borehole layout optimization.

2. The three-dimensional formation modeling method for borehole strategy optimization according to claim 1, characterized in that, The unsampled location data predicted in S1 is obtained by using a machine learning algorithm with continuous static penetration test data as input features to train the model to predict the stratigraphic type and engineering property parameters of the unsampled locations in the standard penetration test.

3. The three-dimensional formation modeling method for borehole strategy optimization according to claim 1, characterized in that, S2 specifically includes: S2-1. Design the GA-Stacking model framework, including the basic Stacking structure and genetic algorithm optimization strategy; S2-2, Model training and parameter optimization for the GA-Stacking model framework; S2-3. The optimized GA-Stacking model is applied to the study area. Based on multi-belief geological data, stratigraphic information in sparse areas is predicted, and virtual boreholes are generated. Each virtual borehole contains three-dimensional coordinates, stratigraphic category labels and corresponding probability distributions to ensure that its stratigraphic sequence conforms to the regional sedimentary patterns.

4. The three-dimensional formation modeling method for borehole strategy optimization according to claim 3, characterized in that: When generating virtual boreholes in S2-3, the generated boreholes conform to the stratigraphic regularity in the vertical direction and adapt to the structural trend in the horizontal direction; the stratigraphic statistical variability function is introduced as a spatial correlation constraint to ensure that the generated virtual points maintain reasonable spatial autocorrelation with known points; and the generated results can be screened and corrected by combining expert knowledge rules.

5. The three-dimensional formation modeling method for borehole strategy optimization according to claim 1, characterized in that, S3 specifically includes: S3-1, Stratigraphic data preprocessing and control point extraction; The original exploration data and virtual borehole data were systematically analyzed. The stratigraphic data of each exploration point were vertically sequenced according to depth. The elevation of the upper and lower interfaces of continuous stratigraphic units was identified, and a complete dataset of stratigraphic interface control points was constructed. S3-2, Calculation of Boundary and Mesh Generation; The computational boundary of the study area was determined, and a buffer zone was set to eliminate boundary effects. An equal-spacing grid partitioning strategy was adopted, and the grid resolution was set in combination with the range of the study area and the characteristics of the stratigraphic thickness to provide dense computational support points for surface interpolation and ensure the accuracy of interface reconstruction. S3-3, Radial basis function interpolation modeling; The formation interface surface is constructed using the radial basis function algorithm; S3-4, Modeling accuracy verification; The accuracy of the model is quantified by measuring the fit between the thinning test and the profile.

6. The three-dimensional formation modeling method for borehole strategy optimization according to claim 5, characterized in that, S3-3 includes: Interpolation function definition; For a given set of stratigraphic interface control points, the interpolation function takes the form: ， Where: Φ is the kernel function, λ i Let ||(x,y)-(x) be the coefficients to be determined. i ,y i || is the Euclidean distance norm; Regularization optimization: Regularization techniques are introduced to balance interpolation accuracy and surface smoothness; Model generation: Based on the interpolated stratigraphic interface, a three-dimensional stratigraphic entity model is constructed to realize the visual reconstruction of multiple types of stratigraphy.

7. The three-dimensional formation modeling method for borehole strategy optimization according to claim 1, characterized in that, S4 include: S4-1. Information entropy is used as an uncertainty evaluation index to quantitatively describe the local reliability of the formation model; For each formation unit in a virtual borehole, the entropy value H = −∑k=1mpklogpk, where pk is the probability that the unit belongs to the k-th formation category, and m is the total number of formation categories. The entropy value ranges from 0 to 1. The smaller the entropy value, the closer the formation distribution is to the actual situation, and the lower the uncertainty. The larger the entropy value, the higher the uncertainty, and the formation distribution may have a large error. The three-dimensional distribution of information entropy in the study area is obtained by interpolation. S4-2. Based on the uncertainty quantification results, formulate a borehole layout strategy; Calculate the average entropy value of virtual boreholes in each region, and determine the region with the largest average entropy value as the first new borehole location. After a new borehole is added, set the information entropy of that location to 0, indicating that the stratigraphic information has been determined through exploration. Recalculate the average entropy value of the remaining regions to determine the next new borehole location. Iterate the above process until the uncertainty is reduced to an acceptable range for the project. S4-3. Verify the optimization effect.

Images