Reef limestone strength prediction method based on pore feature machine learning

Abstract

The invention discloses a reef limestone strength prediction method based on pore feature machine learning, and belongs to the technical field of crossing of geotechnical engineering and artificial intelligence, and the method comprises the steps: obtaining a reef limestone three-dimensional digital core, and extracting pore features after preprocessing; establishing a multi-source pore-physical coupling feature data fusion framework based on the pore features; integrating a random forest and a gradient boosting tree, dynamically weighting parameters in the pore features, designing a composite loss function, performing hierarchical training, and embedding physical constraints to construct a reef limestone strength prediction machine learning model; layered K-fold cross validation is adopted, the robustness of the reef limestone strength prediction machine learning model is tested, and interpretability is verified. According to the method, through data processing, the reef limestone strength prediction machine learning model is constructed and optimized, the relationship between the pore characteristics and the reef limestone strength is deeply excavated, the model prediction accuracy, reliability and interpretability are improved, and an effective solution is provided for reef limestone strength prediction.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of geotechnical engineering and artificial intelligence, and specifically relates to a method for predicting the strength of reef limestone based on pore characteristics and machine learning. Background Art

[0002] In the field of geotechnical engineering and the strength assessment of geological materials, the traditional detection of the mechanical properties of reef limestone mainly relies on physical and mechanical experimental methods, such as uniaxial / triaxial compression tests, point load tests, etc. Although these methods can directly obtain the rock strength parameters, their inherent defects significantly restrict the engineering application efficiency: First, the experimental process requires destructive loading of the samples, resulting in the inability to reuse precious geological samples; Second, it usually takes 2 - 4 weeks from sample collection, preparation to the completion of strength tests, which is difficult to meet the rapid decision - making needs of large - scale projects; Third, the cost of a single experiment is as high as several thousand yuan, which poses a heavy economic burden on the reef stability assessment projects that require a large number of tests.

[0003] Although existing theoretical models attempt to predict the strength of reef limestone through a single parameter such as porosity, their prediction accuracy (usually the coefficient of determination R² < 0.7) far fails to meet the engineering design requirements. The root cause is that the pore structure of reef limestone is controlled by the coral biological skeleton, showing a biogenic spatial distribution law that is completely different from sedimentary rocks - the pore network forms a gradient density distribution along the coral growth direction in three - dimensional space, and there are significant topological connectivity differences between pore clusters. Such complex spatial characteristics have a strong non - linear relationship with mechanical properties, while traditional empirical formulas only rely on the macroscopic statistical parameter of porosity and completely ignore key microscopic structure information such as pore geometry, spatial arrangement directionality, and connectivity, resulting in a deviation of 30% - 50% between the model prediction results and the measured values.

[0004] In view of the above - mentioned technical bottlenecks, there is an urgent need to develop a non - destructive, high - efficiency, and high - precision method for predicting the strength of reef limestone. Summary of the Invention

[0005] Aiming at the deficiencies in the background art, the purpose of the present invention is to propose a method for predicting the strength of reef limestone based on pore characteristics and machine learning. By integrating three - dimensional digital core technology and machine learning algorithms, it breaks through to achieve the quantitative characterization of pore space distribution characteristics and the accurate mapping of non - linear mechanical responses, providing an innovative solution for the strength assessment of complex biogenic rocks.

[0006] The technical solution adopted by the present invention is as follows: A method for predicting the strength of reef limestone based on pore characteristics and machine learning, comprising the following steps: Step S1) Obtain three - dimensional digital core data and pre - process it, and extract pore characteristics at multiple scales; Step S2) Establish a multi-source pore-physical coupling feature data fusion framework based on pore characteristics; Step S3) Integrate random forest and gradient boosting tree, dynamically weight the parameters in pore characteristics, design a composite loss function, train in layers and embed physical constraints to construct a machine learning model for predicting the strength of reef limestone; Step S4) Based on hierarchical K-fold cross-validation, optimize physical constraints, jointly optimize hyperparameters, test the robustness of the machine learning model for predicting the strength of reef limestone and verify its interpretability.

[0007] Preferably, in the step S1), the specific process of obtaining and preprocessing three-dimensional digital core data is as follows: First, use a micron-level high-resolution CT scanning system to obtain three-dimensional digital core data of the reef limestone sample; eliminate noise through the Otsu adaptive threshold segmentation algorithm combined with morphological closing operation, and accurately segment the rock matrix and pore structure; In the step S1), the specific process of multi-scale extraction of pore characteristics is as follows: Use a three-dimensional image analysis framework to extract pore characteristics hierarchically and multi-scale: ① Quantify geometric pore characteristics Based on the pore skeletonization algorithm, calculate the equivalent pore size distribution histogram, use the triangular mesh surface fitting technology to quantify the pore curvature, and calculate the shortest connected path and tortuosity in the pore network through the Dijkstra algorithm to generate the pore connectivity index; ② Analyze the spatial distribution of pore characteristics Establish a three-dimensional space coordinate system along the original growth direction of the coral, use the Voronoi diagram to divide the reef limestone sample into 5×5×5 mm³ sub-units, count the proportion of pore volume in each unit and calculate the density gradient vector; at the same time, extract the main extension direction of the pore cluster through principal component analysis, and calculate the cosine value of the angle between it and the coral growth axis as the direction correlation coefficient; ③ Model topological pore characteristics Based on the three-dimensional pore network model, count the number of adjacent pores connected to each pore node to generate a coordination number distribution matrix, and use the breadth-first search algorithm to identify the largest connected pore cluster and calculate its proportion in the total pore volume; All pore characteristic parameters are stored in the structured feature database after Z-score standardization to provide an input vector.

[0008] Preferably, in the step S2), the specific steps of establishing the pore-physical coupling feature data fusion framework are as follows: Sb1) Calibrate physical properties Measure the geometric dimensions of the reef limestone sample to calculate the volume V, obtain the mass M, and determine the dry density ρ_d; Sb2) Align feature dimensions Use the Box-Cox transformation to eliminate the order-of-magnitude differences between pore features and physical properties, perform dimensional unification on pore features and physical properties, and then align the feature dimensions; Sb3) Establish a spatio-temporal correlation model Through the spatial registration algorithm, bind the spatially distributed pore features and physical properties to three-dimensional spatial coordinates to ensure that the spatial positions of each dimension in the feature matrix X = [x 1 ,..., x 12 , V, M, ρ_d] can be traced, thereby establishing a spatio-temporal correlation model; Sb4) Enhance the interaction of pore-physical coupling features Based on the spatio-temporal correlation model, introduce pore-physical coupling feature terms, use the mutual information method to screen the feature subset with a correlation with strength > 0.3, and fuse the pore-physical coupling feature matrix to enhance the interaction of pore-physical coupling features; Sb5) Optimize the data robustness of pore-physical coupling features Fill in the missing values and correct the outliers in the fused pore-physical coupling feature matrix. Finally, generate a standardized feature matrix X ∈ R^(n×15), and its condition number κ(X^TX) < 10^3 to ensure the stability of the feature matrix input.

[0009] Preferably, in the step S3), the specific steps for constructing the reef limestone strength prediction machine learning model are as follows: Sc1) Integrate heterogeneous models Construct a dual-channel learning framework of random forest and gradient boosting tree to integrate heterogeneous models, where RF sets 200 decision trees to capture the discretized correlation law of pore topology features, and GBDT configures 500 rounds of iteration to model the continuous gradient effect of spatially distributed pore features; Sc2) Dynamic feature weighting Based on the SHAP value analysis, quantify the contribution of each pore feature to strength prediction, apply a weight coefficient of 1.2 - 2.0 times to the parameters in the pore features, and update the weight matrix every 10 rounds of training through the sliding window mechanism to dynamically adjust the weight matrix during the training process; Sc3) Design a composite loss function Define a hybrid loss function of L = 0.7MAE + 0.3Huber_loss, where MAE is used to constrain the overall prediction deviation, and Huber_loss (δ = 1.5MPa) is used to suppress the interference of outliers. The hyperparameters α = 0.7 and β = 0.3 are determined by Bayesian optimization; Sc4) Hierarchical training strategy First, freeze the GBDT layer and train the RF branch alone until the MAE on the validation set is < 2.5 MPa, that is, until it reaches the convergence state, and then unfreeze the entire network for joint fine-tuning; use the AdamW optimizer with Nesterov momentum acceleration, and the initial learning rate of 3e-4 is decreased to 1e-5 by cosine annealing; Sc5) Embed physical constraints Introduce a porosity-strength monotonicity penalty term in the output layer. When the predicted value increases as the porosity increases, apply a gradient correction with an L1 regularization intensity coefficient λ = 0.1; Construct a machine learning model for predicting the strength of reef limestone by integrating heterogeneous models, dynamic feature weighting, designing a composite loss function, a hierarchical training strategy, and embedding physical constraints.

[0010] Preferably, in the step S4), the specific steps for testing the robustness of the machine learning model for predicting the strength of reef limestone and verifying its interpretability are as follows: Sd1) Hierarchical K-fold cross-validation Divide the machine learning model for predicting the strength of reef limestone by porosity intervals, perform 5-fold cross-validation, and synchronously monitor the MAE curves of the training set and the validation set during each fold of training. When the MAE of the validation set rises continuously for 3 rounds by more than 0.5 MPa, trigger the early stopping mechanism to stop the training of the current fold to prevent overfitting; Sd2) Optimize the embedding of physical constraints Introduce a porosity-strength monotonicity correction module in the output layer of the machine learning model for predicting the strength of reef limestone, which is the physical constraint. Real-time detect the trend of the predicted value changing with the porosity. If the porosity increases while the predicted strength does not decrease, then force the predicted value to be adjusted to the theoretical monotonic decreasing curve by linear interpolation, so that the corrected predicted value 100% satisfies the physical law; Sd3) Joint hyperparameter optimization Adopt a hybrid strategy of grid search and Bayesian optimization to perform two-stage parameter tuning within the preset space: first use grid search for rough tuning to determine the optimal region, and then use Bayesian optimization for fine search. Use the weighted sum of the RMSE and MAE of the validation set as the objective function to find the global optimal hyperparameter combination; Sd4) Test the robustness of the machine learning model for predicting the strength of reef limestone Construct a noisy dataset and a feature missing dataset, respectively test the performance decay rate of the machine learning model for predicting the strength of reef limestone, and use adversarial sample training to enhance the stability of the decision boundary; Sd5) Interpretability verification Globally analyze the contribution degree of each feature by SHAP values to ensure that the SHAP values of the parameters in the pore features rank among the top 5, and the consistency with rock mechanics theory is > 90%.

[0011] Compared with the prior art, the present invention proposes a method for predicting the strength of reef limestone based on pore feature machine learning. The advantages of this method are as follows: (1) Non-destructive and efficient detection, breaking through the limitations of traditional methods. Through CT scanning and digital core analysis technology, the present invention completely replaces traditional destructive mechanical experiments, realizes non-destructive evaluation of reef limestone strength, and the samples can be reused 100%, solving the problem of scarcity of geological samples; the detection cycle is shortened from 2-4 weeks of traditional methods to 2-4 hours (including scanning and model prediction), the single cost is reduced, and the efficiency is improved; the prediction accuracy is high, and the error is significantly reduced compared with the traditional porosity model, meeting the requirements of high-precision engineering design; (2) Multi-dimensional feature fusion, revealing the mechanical mechanism of biological pores, breaking through the limitation of a single porosity parameter. For the first time, 12-dimensional features of three major categories including pore geometry (equivalent pore size, curvature), space (density gradient, direction correlation), and topology (coordination number, proportion of pore clusters) are fused, and a 15-dimensional composite feature space is constructed in combination with physical properties to comprehensively characterize the characteristics of biogenic pores; the non-linear correlation law between the pore network and strength is revealed through SHAP, filling the theoretical gap in the mechanics mechanism of biological rock masses; (3) Machine learning guided by physical constraints to ensure engineering reliability. A monotonicity correction module of porosity-strength is embedded in the model to force the prediction results to conform to the laws of rock mechanics, and the prediction error in the extreme porosity region is reduced; through adversarial training such as noise injection and feature masking, the performance decay rate of the model under complex working conditions is <15%, which is significantly better than traditional models; (4) Significant environmental and cost benefits. Tens of thousands of tons of destructive consumption of reef limestone samples can be avoided every year, reducing the impact of ecological exploration; the system has low comprehensive energy consumption, is more energy-efficient than traditional experimental equipment, and saves project costs; the technology can be extended to projects such as island development and subsea tunnels. Description of the Drawings

[0012] Figure 1 It is a flowchart of the method for predicting the strength of reef limestone based on pore feature machine learning of the present invention. Detailed Embodiments

[0013] Next, the technical solutions in the embodiments of the present application will be further clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. It should be noted that the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present application.

[0014] In order to make the invention purpose, technical solution and advantages of this application clearer, the following further details the embodiments of this application with reference to the accompanying drawings of the specification: In order to better understand the above purposes, features and advantages of the present invention, the following will further illustrate the advantages of the present invention through a comparison of embodiments in combination with the drawings and specific implementation manners.

[0015] The present invention proposes a method for predicting the strength of reef limestone based on pore characteristics and machine learning. The steps of this method are described in detail as follows: Step S1) Obtain three-dimensional digital core data and preprocess it, and extract pore characteristics at multiple scales; Specifically, in the step S1), the specific process of obtaining three-dimensional digital core data and preprocessing is as follows: First, use a micron-level high-resolution CT scanning system (scanning accuracy ≤ 5μm) to obtain three-dimensional digital core data of reef limestone samples; since there are noise points in the original three-dimensional digital core data, the noise points are eliminated by using the Otsu adaptive threshold segmentation algorithm combined with morphological closing operation. Through the operation of dilation first and then erosion, small cavities in the pore structure area are filled, adjacent pore structures are connected, and the boundaries of the pore structures are made smoother and more accurate, so as to accurately segment the rock matrix and pore structures.

[0016] Specifically, in the step S1), the specific process of extracting pore characteristics at multiple scales is as follows: Use a three-dimensional image analysis framework to extract pore characteristics hierarchically and at multiple scales: ① Quantify geometric pore characteristics Based on the pore skeletonization algorithm, calculate the equivalent pore size distribution histogram (statistically divided into 10 levels in the diameter range of 0.1 - 2.0 mm). Through the equivalent pore size distribution histogram, the distribution ratio of pore structures of different sizes in the data can be intuitively understood; use the triangular mesh surface fitting technology to discretize the pore structure surface into multiple triangular meshes, and calculate the curvature of each triangular mesh, and then quantify the pore curvature (curvature radius 0.05 - 1.5 mm). Among them, the pore curvature reflects the degree of bending of the pore structure surface, which is of great significance for studying properties such as the flow resistance of fluids in the pore structure; and calculate the shortest connected path in the pore network through the Dijkstra algorithm, and at the same time obtain the tortuosity (that is, the ratio of the actual path length to the shortest connected path). The greater the tortuosity, the more complex the flow path of the fluid in the pore structure. Further generate a pore connectivity index (taking a continuous value from 0 to 1). The closer the pore connectivity index is to 1, the better the connectivity of the pore structure, and vice versa, the worse the connectivity; ② Analyze the spatial distribution pore characteristics A three-dimensional space coordinate system is established along the original growth direction of the coral. The sample is divided into sub-units of 5×5×5 mm³ using the Voronoi diagram. The volume proportion of the pore structure in each sub-unit (i.e., the ratio of the pore structure volume in the sub-unit to the total volume of the sub-unit) is statistically analyzed, and the density gradient vector (gradient modulus length 0 - 0.8 mm⁻¹) is calculated. At the same time, the main extension direction of the pore cluster (a set of multiple pores that are interconnected and aggregated together) is extracted through principal component analysis (PCA), and the cosine value of the angle between it and the coral growth axis is calculated as the direction correlation coefficient (in the range of 0 - 1). The closer the direction correlation coefficient is to 1, the stronger the correlation between the main extension direction of the pore cluster and the coral growth axis; conversely, the weaker the correlation. ③ Modeling topological pore characteristics Based on the three-dimensional pore network model, the number of adjacent pores connected to each pore node is statistically analyzed to generate a coordination number distribution matrix (with a peak between 2 - 4). The coordination number distribution matrix records the number distribution of pore nodes with different coordination numbers in the entire pore network. Through analysis, the complexity and uniformity of the connections in the pore network can be understood. The breadth-first search (BFS) algorithm is used to identify the largest connected pore cluster (a set with the largest number of pores and the largest volume), and the proportion of its volume in the total pore volume is calculated (dynamic range 10% - 95%). All pore characteristics are standardized by Z-score and stored in a structured feature database. After the standardization process, all pore characteristics are on the same measurement scale, providing an input vector.

[0017] Step S2) Establish a multi-source pore-physical coupling feature data fusion framework based on pore characteristics. Specifically, in the said step S2), the specific process of establishing the pore-physical coupling feature data fusion framework is as follows: Sb1) Calibrate physical properties A laser rangefinder is used to measure the geometric dimensions of the reef limestone sample to calculate the volume V. Through the relevant dimension data, the volume V of the reef limestone sample is calculated according to the corresponding geometric volume calculation formula. An electronic balance is used to obtain the mass M. The dry density ρ_d is measured by the oven drying method, that is, the reef limestone sample is placed in an oven, and under certain temperature and time conditions, the moisture in the reef limestone sample is completely evaporated. Then, based on the mass of the dried reef limestone sample and the volume of the original reef limestone sample, the dry density ρ_d is calculated. The volume, mass, and dry density are physical properties. Sb2) Align feature dimensions Pore characteristics (such as pore curvature, tortuosity, etc.) and the obtained physical properties (volume, mass, dry density) may have different dimensions. The Box-Cox transformation is used to eliminate the differences between pore characteristics and physical properties, so as to achieve the dimensional unification of pore characteristics and physical properties. After dimensional unification and difference elimination, each feature has the same weight and comparability in subsequent data analysis and modeling, and then aligns the feature dimensions; Sb3) Establish a spatio-temporal correlation model Through the spatial registration algorithm, the spatially distributed pore characteristics and physical properties are bound to the three-dimensional space coordinates. The spatially distributed pore characteristics and physical properties are bound in the three-dimensional space. In the constructed feature matrix X = [x 1 ,..., x 12 , V, M, ρ_d] (where x 1 to x 12 represent the parameters in the pore characteristics), the features of each dimension correspond to clear three-dimensional space positions, so that the spatial positions of each dimension in the feature matrix X = [x 1 ,..., x 12 , V, M, ρ_d] can be traced, thus establishing a spatio-temporal correlation model; by binding the physical properties with the spatially distributed pore characteristics, the physical properties can be related to the spatial structure; Sb4) Enhance the interaction of pore-physical coupling characteristics Based on the spatio-temporal correlation model, pore-physical coupling characteristic terms are introduced, including derivative parameters such as pore volume ratio (V_pores / V), pore curvature per unit mass (∑ curvature / M), etc. The mutual information method is used to screen the feature subset with a correlation with strength > 0.3, and a pore-physical coupling feature matrix is fused to enhance the interaction of pore-physical coupling characteristics; Sb5) Optimize the data robustness of pore-physical coupling characteristics Missing value filling and outlier correction are performed on the fused 15-dimensional pore-physical coupling feature matrix. Through missing value filling and outlier correction, the quality and integrity of the pore-physical coupling feature matrix are improved. After processing, a standardized feature matrix X ∈ R^(n×15) is finally generated, and its condition number κ(X^TX) < 10^3 ensures the stability of the input of the standardized feature matrix.

[0018] Step S3) Integrate the random forest and the gradient boosting tree, dynamically weight the parameters in the pore characteristics, design a composite loss function, and perform hierarchical training and embed physical constraints to construct a machine learning model for predicting the strength of reef limestone; Specifically, in the said step S3), the specific process of constructing a machine learning model for predicting the strength of reef limestone is as follows: Sc1) Integrate heterogeneous models Construct a dual-channel learning framework of Random Forest (RF) and Gradient Boosting Decision Tree (GBDT), integrating heterogeneous models. Among them, RF is set with 200 decision trees (maximum depth 15) to capture the discretized correlation law of topological pore features; GBDT is configured with 500 rounds of iteration (learning rate 0.05), and through iterative training and gradual correction of errors, the continuous gradient effect of spatially distributed pore features is modeled; by integrating heterogeneous models, the relationship between pore features and the strength of reef limestone is learned more comprehensively; Sc2) Dynamic feature weighting Based on the analysis of SHAP values (Shapley Additive exPlanation), the contribution degree of each pore feature to strength prediction is quantified, a weight coefficient of 1.2 - 2.0 times is applied to the parameters in the pore features (including pore curvature, direction correlation coefficient, etc.), and the weight matrix is updated every 10 rounds of training through a sliding window mechanism, dynamically adjusting the weight matrix during the training process; different pore features have different importance for reef limestone strength prediction. Through SHAP value analysis, the contribution degree of each pore feature can be accurately evaluated, and dynamically updating the weight matrix can adapt to data changes and the learning process of the model; Sc3) Design of composite loss function Define a hybrid loss function of L = 0.7MAE + 0.3Huber_loss, where MAE is used to constrain the overall prediction deviation, and Huber_loss (δ = 1.5MPa) is used to suppress the interference of outliers. The hyperparameters α = 0.7 and β = 0.3 are determined by Bayesian optimization; MAE can intuitively reflect the deviation degree of the overall prediction but is sensitive to outliers; Huber_loss can effectively reduce the impact of outliers on the model. By combining them and using Bayesian optimization to determine the optimal weights, the robustness can be enhanced; Sc4) Hierarchical training strategy First, freeze the GBDT layer and train the RF branch alone until the MAE on the validation set < 2.5MPa, that is, reach the convergence state, and then unfreeze the entire network for joint fine-tuning; use the AdamW optimizer with Nesterov momentum acceleration (γ = 0.9), and the initial learning rate of 3e - 4 is decreased to 1e - 5 through cosine annealing; Sc5) Embedding physical constraints Introduce a porosity-strength monotonicity penalty term in the output layer. When the predicted value rises with the increase of porosity, apply a gradient correction with an L1 regularization strength coefficient λ = 0.1; ensure that the predicted value rises monotonically with the increase of porosity through the penalty term, which conforms to the actual physical law, and suppress the excessive complexity through gradient correction to prevent overfitting; Through methods such as integrating heterogeneous models, dynamic feature weighting, designing composite loss functions, hierarchical training strategies, and embedding physical constraints, construct a machine learning model for predicting the strength of reef limestone.

[0019] Step S4) Based on hierarchical K-fold cross-validation, optimize the physical constraints, jointly optimize the hyperparameters, test the robustness of the machine learning model for predicting the strength of reef limestone, and verify the interpretability; Specifically, in the step S4), the specific process of testing the robustness of the machine learning model for predicting the strength of reef limestone and verifying the interpretability is as follows: Sd1) Hierarchical K-fold cross-validation Divide the machine learning model for predicting the strength of reef limestone into multiple sub-intervals according to the porosity interval (6% - 32%, step size 5%), and perform 5-fold cross-validation (training:validation = 8:2). During each fold of training, synchronously monitor the MAE curves of the training set and the validation set. When the MAE of the validation set rises continuously for 3 rounds by more than 0.5 MPa, trigger the early stopping mechanism to stop the training of the current fold to prevent overfitting; This method can not only effectively evaluate the performance, but also reduce unnecessary computational effort through the early stopping mechanism and improve the optimization efficiency; Sd2) Optimize the embedding of physical constraints Introduce a porosity-strength monotonicity correction module in the output layer of the machine learning model for predicting the strength of reef limestone, which is the physical constraint. Real-time detect the trend of the predicted value changing with the porosity. If the porosity increases while the predicted strength does not decrease, then use linear interpolation to force the predicted value onto the theoretical monotonic decreasing curve (the interpolation reference point is taken from the training set quantile), so that the corrected predicted value 100% satisfies the physical law, that is, the strength must decrease when the porosity increases, thereby improving the credibility and scientificity of the model; Sd3) Jointly optimize the hyperparameters Adopt a hybrid strategy of Grid Search and Bayesian Optimization to perform two-stage parameter tuning within the preset space (RF tree depth 8 - 20, GBDT learning rate 0.01 - 0.1, loss weight α = 0.5 - 0.8): first use grid search for rough tuning to determine the optimal region, and then use Bayesian optimization for fine search. Use the weighted sum of the RMSE and MAE of the validation set (weight ratio 6:4) as the objective function to find the global optimal hyperparameter combination. The combination of grid search and Bayesian optimization better balances the search efficiency and accuracy; Sd4) Test the robustness of the machine learning model for predicting the strength of reef limestone Construct a noisy dataset (adding ±5% Gaussian noise) and a dataset with missing features (randomly masking 30% of the feature values) to simulate errors and incomplete data acquisition respectively. Test the performance decay rate of the machine learning model for predicting the strength of reef limestone (a RMSE increase of less than 15% is considered qualified), and use adversarial sample training to enhance the stability of the decision boundary and improve robustness. These tests aim to evaluate the robustness of the machine learning model for predicting the strength of reef limestone in different scenarios and ensure that it can maintain a high performance level in the face of noise, missing data, or adversarial attacks. Sd5) Interpretability verification Quantify the contribution degree of each feature through global SHAP value analysis to ensure that the SHAP values of the parameters in the pore features (such as pore connectivity index, direction correlation coefficient, etc.) rank among the top 5, and the consistency with rock mechanics theory is greater than 90%. Verify the interpretability of the machine learning model for predicting the strength of reef limestone to ensure that the importance of the parameters in the pore features is consistent with the theoretical expectation. Through SHAP value analysis, it is possible to clearly understand which features have the greatest impact on the prediction results of the machine learning model for predicting the strength of reef limestone, thereby enhancing the transparency and credibility of the machine learning model for predicting the strength of reef limestone.

[0020] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0021] Obviously, those skilled in the art can make various changes and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for predicting reef limestone strength based on pore feature machine learning, characterized in that: include: Step S1) acquiring and preprocessing three-dimensional digital core data, and extracting pore characteristics at multiple scales; Step S2) establishing a multi-source pore-physical coupling feature data fusion framework based on pore features; Step S3) integrating random forest and gradient boosting tree, dynamically weighting the parameters in pore characteristics, designing a composite loss function, hierarchical training and embedding physical constraints to construct a machine learning model for reef limestone strength prediction; Step S4) Based on stratified K-fold cross-validation, physical constraints are optimized, hyperparameters are jointly optimized, the robustness of the reef limestone strength prediction machine learning model is tested, and the interpretability is verified.

2. The method for predicting reef limestone strength based on pore characteristic machine learning according to claim 1 is characterized in that: In step S1), the specific process of obtaining and preprocessing the three-dimensional digital core data is as follows: A micron-level high-resolution CT scanning system is used to obtain three-dimensional digital core data of reef limestone samples. The noise of the three-dimensional digital core data is eliminated by combining the Otsu adaptive threshold segmentation algorithm with morphological closing operations to accurately segment the rock matrix and pore structure.

3. The method for predicting reef limestone strength based on pore characteristic machine learning according to claim 1 is characterized in that: In step S1), the specific process of multi-scale extraction of pore characteristics is as follows: Using the 3D image analysis framework, pore features are extracted hierarchically and at multiple scales: ①Quantitative geometric pore characteristics The equivalent pore size distribution histogram is calculated based on the pore skeletonization algorithm, the pore curvature is quantified using the triangular mesh surface fitting technology, and the shortest connected path and tortuosity in the pore network are calculated using the Dijkstra algorithm to generate the pore connectivity index. ②Analysis of spatial distribution pore characteristics A three-dimensional spatial coordinate system was established along the original growth direction of the coral, and the reef limestone samples were divided into 5×5×5mm³ subunits using the Voronoi diagram. The volume proportion of the pore structure in each unit was counted and the density gradient vector was calculated. At the same time, the main extension direction of the pore cluster was extracted through principal component analysis, and the cosine value of the angle between it and the coral growth axis was calculated as the directional correlation coefficient. ③Modeling topological pore characteristics Based on the three-dimensional pore network model, the number of adjacent pores connected to each pore node is counted to generate a coordination number distribution matrix, and the breadth-first search algorithm is used to identify the largest connected pore cluster and calculate its proportion of the total pore volume. All pore features are normalized by Z-score and stored in the structured feature database to provide input vectors.

4. The method for predicting reef limestone strength based on pore characteristic machine learning according to claim 1 is characterized in that: In step S2), the specific steps of establishing the pore-physical coupling feature data fusion framework are as follows: Sb1) Calibration of physical properties Measure the geometric dimensions of the reef limestone sample to calculate the volume V, obtain the mass M, and determine the dry density ρ_d; Sb2) Align feature dimensions The Box-Cox transformation is used to eliminate the order of magnitude difference between pore characteristics and physical properties, unify the pore characteristics and physical properties, and align the characteristic dimensions. Sb3) Establishing spatiotemporal correlation model The spatial distribution pore characteristics and physical properties are bound to the three-dimensional spatial coordinates through the spatial registration algorithm to ensure that the characteristic matrix X=[x1,...,x 12 ,V,M,ρ_d], thereby establishing a spatiotemporal correlation model; Sb4) Enhanced pore-physical coupling feature interaction Based on the spatiotemporal correlation model, the pore-physical coupling feature term is introduced, and the feature subset with intensity correlation greater than 0.3 is screened using the mutual information method to fuse the pore-physical coupling feature matrix to enhance the pore-physical coupling feature interaction. Sb5) Optimize the robustness of pore-physics coupling feature data The fused pore-physical coupling feature matrix is ​​filled with missing values ​​and corrected for outliers, and finally a standardized feature matrix X∈R^(n×15) is generated, whose condition number κ(X^TX)<10^3 ensures the stability of the standardized feature matrix input.

5. The method for predicting reef limestone strength based on pore characteristic machine learning according to claim 1 is characterized in that: In step S3), the specific steps of constructing a machine learning model for predicting reef limestone strength are as follows: Sc1) Integrating heterogeneous models A dual-channel learning framework of random forest and gradient boosting tree was constructed to integrate heterogeneous models, in which RF set 200 decision trees to capture the discretized association rules of topological pore characteristics, and GBDT was configured with 500 rounds of iterations to model the continuous gradient effect of spatially distributed pore characteristics. Sc2) Dynamic feature weighting The contribution of each pore feature to strength prediction is quantified based on SHAP value analysis, and a weight coefficient of 1.2-2.0 is applied to the parameters in the pore feature. The weight matrix is ​​updated every 10 rounds of training through a sliding window mechanism, and the weight matrix is ​​dynamically adjusted during the training process. Sc3) Design a composite loss function Define a mixed loss function of L=0.7MAE+0.3Huber_loss, where MAE is used to constrain the overall prediction deviation, Huber_loss (δ=1.5MPa) is used to suppress outlier interference, and hyperparameters α=0.7 and β=0.3 are determined by Bayesian optimization; Sc4) Hierarchical training strategy First, freeze the GBDT layer and train the RF branch alone until the MAE on the validation set is less than 2.5MPa, which means convergence is reached. Then unfreeze the entire network for joint fine-tuning. Use the AdamW optimizer accelerated by Nesterov momentum, and reduce the initial learning rate of 3e-4 to 1e-5 through cosine annealing. Sc5) Embedding physical constraints A porosity-intensity monotonicity penalty term is introduced in the output layer. When the predicted value increases with the increase of porosity, a gradient correction with an L1 regularization intensity coefficient λ=0.1 is applied; A machine learning model for predicting reef limestone strength is constructed by integrating heterogeneous models, dynamic feature weighting, designing composite loss functions, hierarchical training strategies, and embedding physical constraints.

6. The method for predicting reef limestone strength based on pore characteristic machine learning according to claim 1, characterized in that: In step S4), the specific steps of testing the robustness of the reef limestone strength prediction machine learning model and verifying the interpretability are as follows: Sd1) Stratified K-fold Cross Validation The reef limestone strength prediction machine learning model was divided into layers according to the porosity interval, and a 5-fold cross validation was performed. The MAE curves of the training set and the validation set were monitored synchronously during each fold training. When the MAE of the validation set increased by more than 0.5 MPa for three consecutive rounds, the early stopping mechanism was triggered to stop the training of the current fold to prevent overfitting. Sd2) Optimize physical constraint embedding A porosity-strength monotonicity correction module is introduced into the output layer of the reef limestone strength prediction machine learning model, which is a physical constraint. The trend of the predicted value changing with the porosity is detected in real time. If the porosity increases but the predicted strength does not decrease, the predicted value is forced to be adjusted to the theoretical monotonic decreasing curve through linear interpolation, so that the corrected predicted value 100% meets the physical law. Sd3) Joint optimization of hyperparameters A hybrid strategy of grid search and Bayesian optimization is used to perform two-stage parameter adjustment in the preset space: first, grid search is used to roughly determine the optimal area, and then Bayesian optimization is used for fine search, with the weighted sum of the validation set RMSE and MAE as the objective function to find the global optimal hyperparameter combination; Sd4) Testing the robustness of the machine learning model for predicting reef limestone strength Construct a noisy dataset and a feature-missing dataset to test the performance decay rate of the reef limestone strength prediction machine learning model, and use adversarial sample training to enhance the stability of the decision boundary. Sd5) Explainability Verification The contribution of each feature was quantified through global analysis of SHAP values ​​to ensure that the SHAP values ​​of the parameters in the pore characteristics ranked in the top 5 and were >90% consistent with rock mechanics theory.

Images