Intelligent prediction method for rock strength parameters based on mineral data and model fusion
By constructing a three-dimensional digital core model and a machine learning regression model, combined with mineral data and microscopic numerical simulation, the problem of insufficient samples in rock strength prediction was solved, and high-precision and highly adaptable rock strength prediction was achieved.
Patent Information
- Application Number
- CN202611122632.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies rely on destructive testing in rock strength prediction and struggle to effectively integrate mineral microstructure and mechanical information, resulting in insufficient reliability and low accuracy in predictions with limited sample sizes.
By acquiring multi-source mineral parameters, constructing a three-dimensional digital core model and performing mesh generation, and combining microscopic numerical simulation and machine learning regression model, rock strength is predicted using mineral phase type and content data.
It enables high-precision rock strength prediction with limited measured data, reduces reliance on destructive testing, improves the convenience and real-time nature of prediction, and enhances adaptability to changes in mineral assemblage.
Smart Images

Figure CN123072541A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent rock strength prediction technology, specifically to an intelligent prediction method for rock strength parameters based on the fusion of mineral data and models. Background Technology
[0002] In rock mechanics parameter assessment, obtaining rock strength data relies on indoor uniaxial compression tests or triaxial tests. These methods have strict requirements for sample preparation and long testing cycles, making them unsuitable for large-scale rapid testing of irregular rock fragments or precious rock samples. Predicting strength using empirical formulas based on mineral composition and simple regression can reduce the workload to some extent, but mineral content data alone cannot reflect the control of microstructural factors such as mineral particle morphology, contact relationships, and cementation characteristics on macroscopic mechanical behavior, resulting in significant prediction bias. Mineral-strength prediction models based on machine learning require a large amount of labeled strength data as training samples, but the measured destructive strength data available in actual engineering environments or scientific research scenarios is often very limited. Insufficient samples can lead to model overfitting or weak generalization ability. While discrete element method (DEM) and other microscopic numerical simulation techniques can reproduce the influence of mineral particle heterogeneity on the rock fracture process, accurate numerical models require detailed information on mineral types, structures, single-mineral mechanical parameters, and interparticle contact parameters. These conditions are difficult to simultaneously meet in the field, and the simulation results deviate from the actual rock mass strength, making direct use of these simulations for calibration or substitution of measured values uncertain. Existing technologies fail to fully integrate multi-source mineralogical information to simultaneously drive micromechanical simulation and data-driven modeling. This makes it difficult to obtain results with both physical consistency and statistical accuracy for intelligent prediction of rock strength under limited measured sample conditions. It is necessary to address how to generate simulated strength data with micromechanical implications as supplementary labels when parameters such as mineral type, content, structure, elastic modulus, and hardness are known. Furthermore, it is necessary to solve the problem of how to organically integrate a small amount of measured strength with simulated strength to establish a high-precision strength mapping model with readily available mineral composition characteristics as input. Summary of the Invention
[0003] The purpose of this invention is to provide an intelligent prediction method for rock strength parameters based on the fusion of mineral data and models, in order to solve the problems of existing methods relying on a large number of destructive tests, insufficient prediction reliability under small sample conditions, and difficulty in effectively integrating mineral microstructure and mechanical information for strength prediction.
[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an intelligent prediction method for rock strength parameters based on mineral data and model fusion, comprising the following steps: acquiring multi-source mineral parameters of the rock to be predicted, wherein the multi-source mineral parameters include at least mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, hardness data of each mineral, and measured macroscopic strength data of the rock; constructing a three-dimensional digital core model based on the multi-source mineral parameters, and performing meshing of the three-dimensional digital core model based on the Voronoi algorithm to establish a mesoscopic numerical simulation model; applying uniaxial compression boundary conditions to the mesoscopic numerical simulation model, performing mesoscopic numerical simulation calculations, and extracting simulated stress-strain curves and simulated macroscopic strength data; using the mineral phase type data and the mineral phase content data as input features, and using the measured macroscopic strength data of the rock and the simulated macroscopic strength data as label values, training a machine learning regression model to obtain a mapping relationship model; inputting the field-measured mineral phase type data and mineral phase content data of the rock to be predicted into the mapping relationship model, and outputting the predicted rock strength value.
[0005] As a technical solution of the present invention, in the step of obtaining multi-source mineral parameters of the rock to be predicted, a portable X-ray diffractometer is used to obtain mineral phase type data and mineral phase content data of the rock; a TIMA automated mineral analysis system combined with a polarizing microscope is used to obtain mineral type data and mineral structure data of the rock; a nanoindenter is used to obtain elastic modulus data and hardness data of each mineral in the rock; and a uniaxial testing machine is used to obtain measured data of the rock's macroscopic strength. The mineral parameters obtained through multi-source testing methods cover cross-scale information such as composition, structure, and mechanical properties, which can provide a more comprehensive data foundation for digital core modeling and machine learning.
[0006] Preferably, the step of constructing a three-dimensional digital core model based on the multi-source mineral parameters specifically includes: cross-scale data fusion of the mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, hardness data of each mineral, and measured macroscopic strength data of the rock; determining typical microstructural parameters based on the fused data; and constructing a three-dimensional digital core model based on the typical microstructural parameters. The typical microstructural parameters include mineral grain size distribution data, mineral contact relationship data, porosity data, and cementation strength data. The three-dimensional digital core model constructed in this way can realistically reflect the mineral composition and microstructural characteristics inside the rock, improving the fidelity of subsequent numerical simulations.
[0007] Furthermore, the steps of meshing the three-dimensional digital core model and establishing a mesoscopic numerical simulation model specifically include: meshing the three-dimensional digital core model using the discrete element method; generating a Voronoi mesh numerical model based on the Voronoi mesh algorithm; setting the mesoscopic mechanical parameters of the Voronoi mesh numerical model, including mineral particle density, mineral particle contact stiffness, mineral particle friction coefficient, and mineral particle bonding strength; and using the Voronoi mesh numerical model with the set mesoscopic mechanical parameters as the mesoscopic numerical simulation model. This model fully considers the geometric morphology and contact mechanical properties of mineral particles, enabling the mesoscopic simulation to effectively reflect the interaction between mineral particles and their influence on macroscopic mechanical behavior.
[0008] In one embodiment of the present invention, the steps of applying uniaxial compressive boundary conditions to the microscopic numerical simulation model, performing microscopic numerical simulation calculations, and extracting simulated stress-strain curves and simulated macroscopic strength data specifically include: applying a downward displacement load to the upper surface of the microscopic numerical simulation model, with the lateral surface of the microscopic numerical simulation model being a free surface, and applying fixed constraints to the lower surface of the microscopic numerical simulation model as uniaxial compressive boundary conditions; performing microscopic failure simulation calculations under uniaxial compression conditions using the discrete element iterative calculation method; recording stress and strain data at each calculation step during the simulation calculation process to generate simulated stress-strain curves; and extracting peak stress from the simulated stress-strain curves as simulated macroscopic strength data. The simulated strength data obtained through microscopic numerical simulation can serve as a valuable supplement to measured strength data, compensating for the limited number of field test samples.
[0009] As another technical solution of the present invention, the steps for training the machine learning regression model specifically include: encoding the mineral phase type data into a category feature vector, encoding the mineral phase content data into a numerical feature vector, and concatenating the category feature vector and the numerical feature vector into an input feature vector; weightedly fusing the measured data of the rock macroscopic strength and the simulated macroscopic strength data to obtain a fused strength label value; using the input feature vector as the model input and the fused strength label value as the model output target, training an initial regression model using a random forest regression algorithm; and performing cross-validation and hyperparameter optimization on the initial regression model to obtain a trained mapping relationship model. This solution organically integrates measured data and simulated data, and the label value retains the authenticity of the field measurements while incorporating the physical laws revealed by numerical simulation, which is beneficial to improving the generalization ability and prediction accuracy of the mapping relationship model.
[0010] When training the initial regression model, preferably, the input feature vector is normalized to eliminate dimensional differences; hyperparameters of the random forest regression algorithm are set, including the number of decision trees N, the maximum depth of the decision tree D, and the minimum number of samples M for node splitting; N subsets of samples with replacement are drawn from the training set using the Bootstrap sampling method; a decision tree is constructed based on each subset, for a total of N decision trees, each trained based on the input feature vector and the fusion strength label value. The process of constructing a decision tree based on each subset specifically includes: when splitting at each node of the current decision tree, a feature subset is randomly selected from the input feature vector; the optimal splitting feature and optimal splitting threshold are selected from the feature subset based on the mean squared error criterion; the sample set of the current node is split into a left child node sample set and a right child node sample set according to the optimal splitting feature and the optimal splitting threshold; the node splitting process is recursively executed until the depth of the current decision tree reaches the maximum depth D or the number of node samples is less than the minimum number of samples M for node splitting. The optimal splitting feature and threshold are selected based on the mean squared error criterion as follows: For each candidate feature in the feature subset, all values of the candidate feature are traversed as candidate splitting thresholds; the rock strength variances of the left and right child node sample sets under each candidate splitting threshold are calculated; the weighted sum of the rock strength variances of the left and right child node sample sets is calculated as the mean squared error corresponding to the candidate splitting threshold; the candidate feature and candidate splitting threshold with the smallest mean squared error are selected as the optimal splitting feature and optimal splitting threshold. The random forest regression model constructed in this way can effectively handle the complex nonlinear mapping relationship between multi-dimensional input features such as mineral phase type and content and rock strength, and by integrating multiple decision trees, the risk of overfitting can be significantly reduced.
[0011] In the hyperparameter optimization process, a five-fold cross-validation method is used to divide the training set into five subsets. Four subsets are used alternately as the training subset, and one subset as the validation subset, repeating the training and validation process five times. During each validation process, the coefficient of determination (R²) and root mean square error (RMSE) between the predicted intensity value and the fused intensity label value are calculated. The number of decision trees (N), the maximum depth of the decision trees (D), and the minimum number of samples for node splits (M) are adjusted, and the five-fold cross-validation process is repeated until the coefficient of determination (R²) is maximized and the RMSE is minimized. The hyperparameter combination corresponding to maximizing the coefficient of determination (R²) and minimizing the RMSE is taken as the optimized hyperparameter combination. The optimized mapping model can maintain prediction accuracy while exhibiting good stability.
[0012] This invention enables intelligent prediction of rock strength parameters based on the fusion of mineral data and models. It deeply integrates multi-source mineral parameter testing, microscopic numerical simulation and machine learning regression modeling. It can achieve high-precision prediction of rock strength using only mineral phase type and content data that can be quickly obtained on site, which greatly improves the convenience and real-time performance of on-site prediction and reduces the reliance on a large number of indoor rock mechanics tests.
[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows:
[0014] By acquiring data on mineral phase types, content, types, structures, elastic moduli, and hardness of each mineral, a three-dimensional digital core model is constructed based on these multi-source mineral parameters. This model is then meshed using the Voronoi grid algorithm and assigned micromechanical parameters to establish a micro-scale numerical simulation model. Uniaxial compressive boundary conditions are applied to the model to perform simulation calculations, extracting simulated stress-strain curves and simulated macroscopic strength data. This approach directly embeds the actual mechanical properties of mineral particles into the numerical model, enabling the simulation process to reproduce the contribution of micromechanical mechanisms such as mineral interface slip, particle fragmentation, and cementation failure to macroscopic strength. The generated simulated macroscopic strength data carries the heterogeneous distribution and contact mechanical characteristics of minerals, maintaining physical consistency with measured strength. This simulated data can serve as an effective extension of measured labels, enriching the training sample space when measured destructive data is scarce, reducing reliance on numerous laboratory experiments, and enhancing sample diversity by batch generating supplementary labels covering various lithological scenarios through changes in mineral assemblage parameters.
[0015] Using mineral phase type and content data as input features, the measured and simulated macroscopic strength data of rocks are weighted and fused to form a label value, which is then used to train a machine learning regression model to form a mapping relationship model. In the prediction stage, only the mineral phase types and contents of the rock to be tested in the field need to be obtained using portable X-ray diffraction and other methods and input into the model to directly output the predicted strength value. The method of constructing a model that integrates measured and simulated labels allows the model to learn the implicit nonlinear relationship between mineral microstructure features and strength from simulated labels even with only a small number of measured strength samples, avoiding overfitting and performance degradation due to insufficient samples. The weighted fusion strategy allows adjusting the contribution ratio of measured and simulated labels according to data reliability. Without compromising the priority of measured data, simulated labels are used to fill the gaps in sample distribution. The resulting mapping relationship, which is more adaptable to changes in mineral phase combination, significantly improves the prediction stability and accuracy under complex mineral composition conditions, while meeting the engineering need for rapid acquisition of rock strength parameters in the field. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0017] Figure 1 This is a flowchart of an intelligent prediction method for rock strength parameters based on the fusion of mineral data and models;
[0018] Figure 2 This is a flowchart of the process for establishing a mesoscopic numerical simulation model based on the Voronoi mesh;
[0019] Figure 3 This is a flowchart of a uniaxial compressional mesoscopic damage simulation based on the Voronoi discrete element method.
[0020] Figure 4 It is the measured stress-strain curve of the macroscopic strength of the rock;
[0021] Figure 5 It is the particle size distribution function curve of mineral particles. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.
[0023] See Figure 1This invention provides an intelligent prediction method for rock strength parameters based on mineral data and model fusion, comprising the following steps: acquiring multi-source mineral parameters of the rock to be predicted, including at least mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, hardness data of each mineral, and measured macroscopic strength data of the rock; constructing a three-dimensional digital core model based on the multi-source mineral parameters, and performing meshing on the three-dimensional digital core model based on the Voronoi algorithm to establish a mesoscopic numerical simulation model; applying uniaxial compression boundary conditions to the mesoscopic numerical simulation model, performing mesoscopic numerical simulation calculations, and extracting simulated stress-strain curves and simulated macroscopic strength data; training a machine learning regression model using mineral phase type data and mineral phase content data as input features, and using measured macroscopic strength data and simulated macroscopic strength data of the rock as label values to obtain a mapping relationship model; inputting the field-measured mineral phase type data and mineral phase content data of the rock to be predicted into the mapping relationship model, and outputting the predicted rock strength value.
[0024] Example 1:
[0025] In practical implementation, when obtaining multi-source mineral parameters of the rock to be predicted, a portable X-ray diffractometer is used to acquire data on the types and contents of mineral phases. The portable X-ray diffractometer performs in-situ detection on the surface of the rock sample, emitting X-rays onto the rock sample and collecting the diffraction patterns of the mineral crystals in the rock sample. Based on the positions and intensities of the diffraction peaks in the diffraction patterns, the types of mineral phases contained in the rock sample are identified by matching them with a standard mineral phase diffraction pattern database. Based on the quantitative relationship between diffraction peak intensity and mineral phase content, the mass percentage of each mineral phase in the rock sample is calculated, thus obtaining the mineral phase content data.
[0026] The TIMA automated mineral analysis system, coupled with a polarizing microscope, was used to acquire mineral type and structure data of rocks. Rock samples were prepared into thin sections and placed on the sample stage of the TIMA system. The TIMA system performed electron beam scanning on the rock thin sections, acquiring backscattered electron images and characteristic X-ray signals. Backscattered electron images were analyzed to determine the geometric morphology and distribution characteristics of mineral grains, while energy dispersive spectroscopy (EDS) analysis of the characteristic X-ray signals determined the chemical elemental composition of the mineral grains. The geometric morphology and chemical elemental composition of the mineral grains were compared with a mineral database to determine the mineral type data for each mineral grain in the rock sample. The polarizing microscope was used to observe the rock thin sections under polarized light, acquiring extinction characteristics, interference color characteristics, and contact relationship characteristics between mineral grains under crossed polarized light. The extinction and interference color characteristics were used to verify and supplement the mineral type data. Based on the contact relationship characteristics, directional arrangement characteristics, and grain size distribution characteristics of the mineral grains, the structural data of the minerals were determined. The structural data included grain size distribution data, contact relationship data, cementation type data, and directional arrangement data.
[0027] The elastic modulus and hardness data of various minerals in rocks were obtained using a nanoindenter. The surface of the rock samples was polished to meet the requirements of the nanoindenter, and the processed rock samples were fixed on the sample stage of the nanoindenter. The indenter applied continuously increasing normal loads to each mineral particle region on the rock sample surface, and the relationship between the depth of the indenter into the mineral particle and the normal load was recorded in real time. Based on the Oliver-Pharr method, the reduced modulus of each mineral particle was calculated using the load-depth curve of the unloading segment, and then converted into the elastic modulus data of each mineral according to Poisson's ratio. Simultaneously, the hardness data of each mineral was calculated based on the ratio of the maximum loaded load to the projected area of the residual indentation produced by the indenter on the mineral particle surface.
[0028] The macroscopic strength data of the rock were obtained using a uniaxial testing machine. The rock to be tested was processed into a standard cylindrical rock specimen, with the flatness of both end faces meeting the requirements of the uniaxial compression test. The rock specimen was placed between the two loading plates of the uniaxial testing machine. The upper plate of the uniaxial testing machine applied axial pressure to the rock specimen at a constant loading rate. The loading process continued until the rock specimen underwent macroscopic failure. The axial force and axial deformation of the rock specimen during the loading process were continuously recorded by force sensors and displacement sensors. The macroscopic strength data of the rock was determined based on the ratio of the recorded maximum axial force to the initial cross-sectional area of the rock specimen.
[0029] See Figure 4In the figure, the horizontal axis represents axial strain in percentage (%), ranging from 0 to 2.5%; the vertical axis represents axial stress in megapascals (MPa). The curve is plotted as a solid blue line, and the legend is labeled "Measured data of macroscopic rock strength," indicating that the curve reflects the measured data of macroscopic rock strength obtained by the uniaxial testing machine in Example 1.
[0030] As can be seen from the curve trend, when the axial strain increases from 0, the axial stress exhibits a near-linear upward trend, indicating that the rock is subjected to axial pressure and undergoes elastic deformation in the elastic stage. When the strain reaches approximately 1.8%, the axial stress reaches its peak value, approximately 63 MPa, representing the rock's ultimate bearing capacity or peak strength. Subsequently, as the axial strain continues to increase to approximately 1.9%, the axial stress drops sharply to near zero, indicating that the rock undergoes macroscopic failure, rapidly loses its bearing capacity, and enters the failure stage.
[0031] Example 2:
[0032] In practice, when constructing a three-dimensional digital core model based on multi-source mineral parameters, the mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, hardness data of each mineral, and measured data of macroscopic rock strength are first fused across scales. When performing cross-scale data fusion, the mineral phase type data obtained by the portable X-ray diffractometer is used as the classification basis for mineral phases in the rock sample; the mineral phase content data obtained by the portable X-ray diffractometer is used as a quantitative constraint on the volume ratio of each mineral phase in the overall rock sample; the mineral type data obtained by the TIMA automated mineral analysis system combined with a polarizing microscope is used as the basis for identifying the mineral phase to which each independent mineral particle in the rock sample belongs; the mineral structure data obtained by the TIMA automated mineral analysis system combined with a polarizing microscope is used as the data source for determining the spatial distribution characteristics of mineral particles and the contact relationship between mineral particles; the elastic modulus data of each mineral obtained by the nanoindenter is used as the mechanical property parameter of each independent mineral particle undergoing elastic deformation under external force; the hardness data of each mineral obtained by the nanoindenter is used as the mechanical property parameter of each independent mineral particle resisting local surface indentation deformation; and the measured data of macroscopic rock strength obtained by the uniaxial testing machine is used as the target value to be fitted to the macroscopic mechanical response of the three-dimensional digital core model. In the process of cross-scale data fusion, a data assimilation method is adopted. Using the geometric position coordinates of mineral particles as a spatial index, mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, and hardness data of each mineral are uniformly mapped to the same spatial coordinate system, and a multi-source fusion dataset containing the spatial location of mineral particles, mineral particle morphology, contact relationship between mineral particles, and mechanical properties of mineral particles is established.
[0033] After the multi-source fusion dataset is established, the following operations are performed: Mineral particles in the multi-source fusion dataset are segmented using image morphological analysis methods, and the particle size data of each mineral particle is extracted. Statistical analysis is then performed on the extracted particle size data to determine the particle size distribution data. The particle size distribution data is represented by a particle size distribution function, which is:
[0034]
[0035] in, Indicates the first The particle size value of each mineral grain. The particle size value is The probability density of mineral grains appearing in a rock sample. This represents the average of the natural logarithms of the mineral particle size, obtained by statistically analyzing the natural logarithms of the particle sizes of all mineral particles in a multi-source fusion dataset and calculating the arithmetic mean. The standard deviation of the natural logarithm of the mineral particle size is obtained by statistically analyzing the natural logarithm of the particle size of all mineral particles in a multi-source fusion dataset and calculating the standard deviation. This section represents the index of mineral particles. Mineral particle size distribution data is used as a component of typical microstructural parameters. Mineral contact relationship data is determined by analyzing the spatial relationships and interface characteristics between adjacent mineral particles in the multi-source fusion dataset. This data includes the contact type, contact surface geometry, and contact surface area. Porosity data is determined by statistically analyzing the ratio of the void volume between mineral particles to the total volume of the rock sample in the multi-source fusion dataset. The specific value of porosity is equal to the sum of the void volumes between all mineral particles in the rock sample divided by the overall outer contour volume of the rock sample. The cementation strength data is determined by analyzing the bonding mechanical properties of the mineral particle contact surface area in the multi-source fusion dataset and inverting and calibrating it using measured macroscopic strength data of the rock. This cementation strength data includes the tensile strength and shear strength of the mineral particle contact surface area.
[0036] When constructing a 3D digital core model based on typical microstructural parameters and the geometric boundaries of a standard cylindrical rock sample, a particle packing algorithm is used to generate the 3D digital core model. The particle packing algorithm determines the number of mineral particles required for each mineral phase based on mineral phase type and content data, determines the particle size of each mineral particle based on particle size distribution data, determines the spatial arrangement rules between mineral particles based on mineral contact relationship data, and determines the density constraints of the mineral particle packing based on porosity data. Starting from the center of the spatial region of the 3D digital core model, the particle packing algorithm places mineral particles sequentially in descending order of particle size. For each mineral particle placed, its spatial coordinates are randomly generated. The algorithm checks whether the overlap between the newly generated mineral particle and the already placed mineral particles meets the contact tolerance range specified by the mineral contact relationship data. If the overlap exceeds the contact tolerance range, the spatial coordinates are regenerated until all mineral particles have been placed. After the ore particles are placed, the mechanical parameters of the cementing material layer are set on the contact surface between the mineral particles based on the cementing strength data. The mechanical parameters of the cementing material layer include the normal stiffness and tangential stiffness of the cementing material. The mineral particle aggregate and the cementing material layer together constitute a three-dimensional digital core model. Each mineral particle in the three-dimensional digital core model is assigned corresponding mineral type data, elastic modulus data, and hardness data.
[0037] See Figure 5 The figure shows the probability density curves of particle size distribution for three minerals: quartz, calcite, and feldspar. The horizontal axis represents the particle size in micrometers (μm) using a logarithmic scale, and the vertical axis represents the probability density P(d) for the corresponding particle size. Quartz particle size distribution is represented by a solid red line, calcite particle size distribution by a dashed green line, and feldspar particle size distribution by a dashed blue line.
[0038] As shown in the figure, calcite exhibits the highest peak value in its grain size distribution curve, with a probability density peak of approximately 0.02. This peak corresponds to a grain size range of approximately 30 μm to 50 μm, indicating that the calcite grains are primarily concentrated within this range and have a relatively concentrated grain size distribution. Quartz shows a lower peak value than calcite, at approximately 0.011, corresponding to a grain size range of approximately 70 μm to 100 μm, indicating that quartz grains are slightly larger and have a wider distribution than calcite grains. Feldspar has the lowest peak value, at approximately 0.005, corresponding to a grain size range of approximately 100 μm to 200 μm, indicating that feldspar grains are even larger and have a wider grain size distribution range.
[0039] Overall, the particle size distribution of the three minerals shows an increasing trend from small to large. The particle size distribution curves for quartz and feldspar are relatively flat, indicating a more dispersed particle size distribution; while the particle size distribution curve for calcite is sharper, indicating a more concentrated particle size distribution. The logarithmic horizontal axis of the particle size distribution function clearly displays the distribution characteristics of mineral particles over a wide range from 1 μm to approximately 1000 μm, which helps in the precise characterization of fine-grained structural parameters.
[0040] Example 3:
[0041] In specific implementation, please refer to Figure 2 To establish a mesoscopic numerical simulation model, the Discrete Element Method (DEM) was used to mesh the 3D digital core model. The DEM treats each mineral grain in the 3D digital core model as an independent rigid discrete element, the cementing layer between mineral grains as the connecting bond between discrete elements, and the outer contour boundary of the 3D digital core model as a rigid boundary restricting the displacement of the discrete elements. Based on the geometric morphology and spatial arrangement of each mineral grain in the 3D digital core model, the DEM discretizes each mineral grain into an independent computational node. Each computational node stores the spatial coordinates, grain size, mineral type, elastic modulus, and hardness of the corresponding mineral grain. Mechanical interactions are established between adjacent computational nodes through connecting bonds. The two ends of each connecting bond connect to the computational nodes corresponding to two adjacent mineral grains. The connecting bonds store the normal stiffness, tangential stiffness, and strength parameters of the cementing layer.
[0042] After processing using the Discrete Element Method (DEM), a Voronoi grid numerical model is generated based on the Voronoi grid algorithm. The Voronoi grid algorithm uses the spatial coordinates of the computational node corresponding to each mineral grain in the 3D digital core model as Voronoi seed points. It calculates the Voronoi cell corresponding to each Voronoi seed point in 3D Euclidean space. A Voronoi cell is defined as the set of all spatial points in 3D Euclidean space that are closer to the corresponding Voronoi seed point than to any other Voronoi seed point. The Voronoi grid algorithm determines the spatial interfaces between Voronoi cells by solving for the perpendicular bisectors of the lines connecting adjacent Voronoi seed points. All the spatial interfaces of the Voronoi cells together constitute the grid skeleton of the Voronoi grid numerical model. Each Voronoi cell in the Voronoi grid numerical model corresponds to a mineral grain in the 3D digital core model. The geometric volume of the Voronoi cell is equal to the geometric volume of the corresponding mineral grain, and the contact area between Voronoi cells is equal to the contact area between the corresponding mineral grains. The Voronoi grid numerical model preserves the particle size distribution characteristics, spatial arrangement characteristics, and contact relationship characteristics of mineral particles in the three-dimensional digital core model.
[0043] After generating the Voronoi mesh numerical model, the mesomechanical parameters of the Voronoi mesh numerical model are set. These parameters include mineral particle density, mineral particle contact stiffness, mineral particle friction coefficient, and mineral particle bond strength. Mineral particle contact stiffness includes normal contact stiffness and tangential contact stiffness. Normal contact stiffness represents the ability of the contact surface between two adjacent Voronoi cells to resist normal deformation, while tangential contact stiffness represents the ability of the contact surface between two adjacent Voronoi cells to resist tangential slip deformation. The specific value of normal contact stiffness is calculated based on the elastic modulus data of the mineral particles corresponding to the two adjacent Voronoi cells and the contact surface area between the mineral particles. The calculation formula is as follows:
[0044]
[0045] in, Indicates adjacent Voronoi cells With Voronoi cell Normal contact stiffness on the contact surface between them This represents the index of the first Voronoi cell in a pair of adjacent Voronoi cells. This represents the index of the second Voronoi cell in a given pair of adjacent Voronoi cells. Voronoi cell The elastic modulus data corresponding to the mineral particles, Voronoi cell The elastic modulus data corresponding to the mineral particles, Voronoi cell With Voronoi cell The area of the contact surfaces between them. Voronoi cell The perpendicular distance from the geometric center to the contact surface. Voronoi cell The perpendicular distance from the geometric center to the contact surface. The specific value of the tangential contact stiffness is set as a proportion of the normal contact stiffness, and the proportion is determined based on the properties of the cementing material in the mineral particle contact area. The mineral particle friction coefficient represents the frictional resistance characteristics when relative slippage occurs between the contact surfaces of Voronoi cells. The specific value of the mineral particle friction coefficient is determined by looking up tables using mineral hardness and mineral type data, and the value ranges from 0.3 to 0.9. The mineral particle bond strength includes the normal bond strength and the tangential bond strength of the contact surface. The normal bond strength represents the maximum normal stress resisting tensile failure of the bonds between Voronoi cells, and the tangential bond strength represents the maximum tangential stress resisting shear failure of the bonds between Voronoi cells. The specific value of the mineral particle bond strength is directly assigned based on the tensile strength and shear strength of the mineral particle contact area in the cementing strength data.
[0046] The Voronoi mesh numerical model with its micromechanical parameters set is used as the micromechanical simulation model. The micromechanical simulation model stores the geometric information of all Voronoi cells, the mechanical property parameters of the mineral grains corresponding to all Voronoi cells, the micromechanical parameters of the contact surfaces between all adjacent Voronoi cells, and the boundary constraint definitions on the outer surface of the Voronoi mesh numerical model. The micromechanical simulation model is used to perform subsequent uniaxial compression numerical simulation calculations.
[0047] Example 4:
[0048] In specific implementation, please refer to Figure 3When applying uniaxial compression boundary conditions to the microscopic numerical simulation model, all surface nodes of the Voronoi cells on the upper surface of the microscopic numerical simulation model are treated as the objects of load application. A vertically downward displacement load, pointing inwards towards the interior of the microscopic numerical simulation model, is applied to each surface node on the upper surface. The application rate of the displacement load is controlled by the time step of the discrete element iteration calculation. The increment of the displacement load within each time step remains constant, and the increment is set to maintain the axial strain rate of the microscopic numerical simulation model within the quasi-static compression range. The cumulative value of the displacement load gradually increases from zero. When applying fixed constraints to the lower surface of the microscopic numerical simulation model, all surface nodes of the Voronoi cells on the lower surface of the microscopic numerical simulation model are treated as the objects of constraint application. Each surface node on the lower surface of the microscopic numerical simulation model is constrained in its translational degrees of freedom in three orthogonal spatial directions, ensuring that the displacement of each surface node on the lower surface of the microscopic numerical simulation model in the vertical direction and the two horizontal directions is always zero. The downward displacement load applied to the upper surface of the mesoscopic numerical simulation model and the fixed constraint applied to the lower surface of the mesoscopic numerical simulation model together constitute the uniaxial compression boundary condition. The uniaxial compression boundary condition causes the mesoscopic numerical simulation model to undergo compressive deformation in the vertical direction and free expansion in the horizontal direction.
[0049] When performing mesoscopic failure simulation calculations under uniaxial compression conditions using the discrete element method (DEM), the DEM employs the explicit central difference method to perform time integration on the motion equations of each Voronoi cell in the mesoscopic numerical simulation model. Within each DEM iteration time step, during the contact detection step, all adjacent Voronoi cell pairs in the mesoscopic numerical simulation model are traversed, and the relative displacement increment between two adjacent Voronoi cells is calculated. This relative displacement increment includes both normal and tangential displacement increments. The normal displacement increment is obtained by projecting the change in the spatial coordinates of the geometric centers of two adjacent Voronoi cells onto the normal vector of the contact surface, and the tangential displacement increment is obtained by projecting the change in the spatial coordinates of the geometric centers of two adjacent Voronoi cells onto the tangential plane of the contact surface. When calculating the contact force on the contact surface between two adjacent Voronoi cells, the normal contact force increment is calculated using the force-displacement relationship based on the normal displacement increment and the normal contact stiffness in the mineral grain contact stiffness of the mesomechanical parameters of the Voronoi mesh numerical model. The tangential contact force increment is calculated using the force-displacement relationship based on the tangential displacement increment and the tangential contact stiffness in the mineral grain contact stiffness of the mesomechanical parameters of the Voronoi mesh numerical model. The normal contact force increment and the tangential contact force increment are then accumulated with the normal contact force and the tangential contact force of the contact surface in the previous time step to obtain the normal contact force and the tangential contact force of the contact surface in the current time step. The normal contact force at the contact surface is limited by the contact surface normal bond strength, which is a component of the mineral particle bond strength in the micromechanical parameters of the Voronoi mesh numerical model. When the contact surface normal contact force exceeds the contact surface normal bond strength, the connecting bonds on the contact surface undergo tensile failure, the contact surface normal contact force becomes zero, and the connecting bonds are marked as fractured. The tangential contact force at the contact surface is limited by the contact surface tangential bond strength, which is a component of the mineral particle bond strength in the micromechanical parameters of the Voronoi mesh numerical model. When the contact surface tangential contact force exceeds the contact surface tangential bond strength, the connecting bonds on the contact surface undergo shear failure, the contact surface tangential contact force remains at the residual friction force value, and the connecting bonds are marked as fractured. The residual friction force value is calculated from the contact surface normal contact force and the mineral particle friction coefficient in the micromechanical parameters of the Voronoi mesh numerical model.When solving the equations of motion, the contact forces and volume forces acting on all contact surfaces of each Voronoi cell are vector-summed to obtain the resultant force and resultant torque acting on the Voronoi cell. The translational acceleration of the Voronoi cell is calculated using Newton's second law based on the resultant force and the mass of the Voronoi cell. The rotational angular acceleration of the Voronoi cell is calculated using Euler's equations of motion based on the resultant torque and the moment of inertia of the Voronoi cell. The translational acceleration and rotational angular acceleration are integrated over time using the central difference method to update the translational velocity and rotational angular velocity of the Voronoi cell. The spatial position coordinates and spatial attitude angles of the Voronoi cell are then updated based on the translational velocity and rotational angular velocity. The above discrete element iterative calculation time steps are repeated for the mesoscopic failure simulation calculation until the axial strain of the mesoscopic numerical simulation model reaches a preset termination strain threshold, which is set to three percent of the axial strain value.
[0050] During the simulation, stress and strain data are recorded for each calculation step. When generating the simulated stress-strain curve, after each discrete element iteration time step is completed, the vertical reaction force components of the surface nodes of all Voronoi cells on the upper surface of the mesoscopic numerical simulation model are extracted. The total reaction force of the upper surface is obtained by summing the vertical reaction force components of all surface nodes on the upper surface of the mesoscopic numerical simulation model. Similarly, the vertical reaction force components of the surface nodes of all Voronoi cells on the lower surface of the mesoscopic numerical simulation model are extracted. The total reaction force of the lower surface is obtained by summing the vertical reaction force components of all surface nodes on the lower surface of the mesoscopic numerical simulation model. The average of the total reaction force of the upper surface and the total reaction force of the lower surface is taken as the axial force borne by the mesoscopic numerical simulation model in the current calculation step. The axial stress is calculated by dividing the axial force by the initial cross-sectional area of the mesoscopic numerical simulation model, which is the cross-sectional area of the mesoscopic numerical simulation model perpendicular to the vertical direction in its undeformed state. The vertical displacements of all Voronoi cells on the upper surface of the mesoscopic numerical simulation model are extracted, and their average value is calculated as the upper surface average displacement. Similarly, the vertical displacements of all Voronoi cells on the lower surface are extracted, and their average value is calculated as the lower surface average displacement. Axial strain is calculated by dividing the absolute value of the difference between the upper and lower surface average displacements by the initial height of the mesoscopic numerical simulation model, which is the vertical distance between the upper and lower surfaces in the undeformed state. The axial stress and strain values obtained at each discrete element iteration time step are treated as data points. All data points are connected in chronological order to form a simulated stress-strain curve. The horizontal axis of the simulated stress-strain curve represents the axial strain value, and the vertical axis represents the axial stress value.
[0051] When extracting peak stress from the simulated stress-strain curve as simulated macroscopic strength data, all data points on the simulated stress-strain curve are traversed. The axial stress value corresponding to each data point is compared with the currently recorded maximum stress value. If the axial stress value corresponding to the current data point is greater than the currently recorded maximum stress value, the currently recorded maximum stress value is updated to the axial stress value corresponding to the current data point. After traversal, the currently recorded maximum stress value is the peak stress, and the peak stress is used as simulated macroscopic strength data.
[0052] Example 5:
[0053] In practice, mineral phase type data and mineral phase content data are used as input features, and measured data and simulated macroscopic strength data of rocks are used as label values. When training the machine learning regression model, mineral phase type data is encoded as a category feature vector, mineral phase content data is encoded as a numerical feature vector, and the category feature vector and the numerical feature vector are concatenated into an input feature vector.
[0054] Mineral phase type data is converted into category feature vectors using hot-coded unique encoding. A complete set of mineral phase types is pre-constructed based on the mineral phase types appearing in all rock samples. This set includes mineral phase types such as quartz, feldspar, calcite, dolomite, and clay minerals. Each mineral phase type in the complete set corresponds to one dimension in the category feature vector, and the total number of dimensions in the category feature vector equals the total number of mineral phase types in the complete set. For a rock sample, if the rock sample contains the [number]th [mineral phase type] in the complete set... For each type of mineral phase, the th element in the category feature vector... The value of each dimension is set to 1. If the rock sample does not contain the first dimension of the complete set of mineral phases... For each type of mineral phase, the th element in the category feature vector... Set the values of each dimension to 0. A numbered index representing the mineral phase types in the complete set of mineral phase types. The value range is from 1 to the total number of mineral phase types.
[0055] Mineral phase content data is converted into numerical feature vectors using numerical encoding. The number of dimensions in the numerical feature vectors is the same as the total number of mineral phase types in the complete set of mineral phase types. The position of each dimension in the numerical feature vector corresponds one-to-one with the order of mineral phase types in the complete set of mineral phase types. For a rock sample, the [missing information - likely a specific element or sequence] of the mineral phase types in the complete set of mineral phase types is extracted. The mineral phase content data corresponding to each mineral phase type is the mass percentage of that mineral phase type in the rock sample. This mineral phase content data is directly entered into the numerical feature vector. One dimension. If the rock sample does not contain the first dimension of the complete set of mineral phases. For each type of mineral phase, the th element in the numerical eigenvector... The values for each dimension are set to 0.
[0056] The categorical feature vector and the numerical feature vector are concatenated end-to-end according to their dimensionality. All dimensions of the categorical feature vector are placed at the beginning of the concatenated vector, and all dimensions of the numerical feature vector are placed at the end of the concatenated vector. The concatenated vector is used as the input feature vector. The total number of dimensions of the input feature vector is equal to twice the total number of mineral phase types in the complete set of mineral phase types.
[0057] The measured and simulated macroscopic strength data of rocks are weighted and fused to obtain a fused strength label value. The weighted fusion calculation method is as follows: the first weight coefficient corresponding to the measured macroscopic strength data of rocks is set to... The second weighting coefficient corresponding to the simulated macroeconomic intensity data is set as follows: First weighting coefficient With the second weighting coefficient Satisfying Relationships First weighting coefficient The specific values are determined comprehensively based on the experimental accuracy of the measured macroscopic strength data of the rock obtained by the uniaxial testing machine and the simulation accuracy of the simulated macroscopic strength data obtained by the microscopic numerical simulation calculation. The first weighting coefficient is... The setting value range is 0.4 to 0.6, the second weighting coefficient. The set value ranges from 0.4 to 0.6. In one specific implementation, the first weighting coefficient... Set to 0.5, second weighting coefficient Setting it to 0.5 ensures that the measured and simulated macroscopic strength data of the rock contribute equally to the fused strength label value. The measured macroscopic strength data of the rock is then compared with the first weighting coefficient. Multiplying yields the first accumulation, which combines the simulated macroscopic intensity data with the second weighting coefficient. Multiplying the first and second multiplications yields the second multiplication accumulation. Adding the first and second multiplication accumulations together gives the fusion strength label value.
[0058] Using the input feature vector as the model input and the fusion strength label value as the model output target, the initial regression model is trained using the random forest regression algorithm. The random forest regression algorithm consists of an ensemble of multiple regression decision trees. Each regression decision tree is constructed independently and makes independent predictions on the input feature vector. The final output of the random forest regression algorithm is the arithmetic mean of the predictions of all regression decision trees.
[0059] The specific process of training the initial regression model using the random forest regression algorithm is as follows:
[0060] The input feature vector is normalized to eliminate dimensional differences. The normalization process uses the Min-Max normalization method, which calculates the minimum value of all sample data in each dimension of the input feature vector for that dimension. and maximum value For each sample, the original value in that dimension The value is converted to a normalized value using a normalization formula. The normalization formula is:
[0061]
[0062] in, This represents the original value in one dimension of the input feature vector. This represents the minimum value of all sample data in this dimension of the input feature vector. This represents the maximum value of all sample data in this dimension of the input feature vector. This represents the value after normalization. The value of is between 0 and 1. Normalization is performed on each dimension of the input feature vector to obtain the normalized input feature vector.
[0063] Set the hyperparameters for the random forest regression algorithm, including the number of decision trees. Maximum depth of decision tree Minimum number of samples for node splitting Number of decision trees The initial value is set to 100, and the maximum depth of the decision tree is... The initial value is set to 10, and the minimum number of samples for node splitting is set. The initial value is set to 5.
[0064] The Bootstrap sampling method is used to sample data with replacement from the training set. A subset of samples. The training set consists of the normalized input feature vectors and fused intensity label values corresponding to all rock samples. The Bootstrap sampling method randomly selects one sample from the training set each time, records its contents, and then puts it back into the training set, ensuring it has the same probability of being selected again in the next selection. This sampling process is repeated until the number of samples selected reaches 100% of the total number of samples in the training set. All selected samples constitute a subset of samples. The Bootstrap sampling method is then repeatedly executed. Next, generate Each of the three independent subsample sets is used to construct a regression decision tree.
[0065] A regression decision tree is constructed based on each subset of samples, for a total of [number] trees. Each regression decision tree is trained based on an input feature vector and a fusion strength label value. When constructing a regression decision tree, the normalized input feature vector of all samples in the corresponding subset is used as the training input, and the fusion strength label value of all samples in the corresponding subset is used as the training target.
[0066] The specific process of constructing a regression decision tree based on each subset of samples is as follows:
[0067] When constructing the root node of the current regression decision tree, a subset of features is randomly selected from all feature dimensions of the input feature vector. The number of feature dimensions in the subset is set to the integer part of the total number of dimensions of the input feature vector divided by 3. The optimal splitting feature and the optimal splitting threshold are selected from the subset based on the mean squared error criterion.
[0068] When selecting the optimal splitting feature and optimal splitting threshold from the feature subset based on the mean squared error criterion, for each candidate feature in the feature subset, the value of that candidate feature in all samples of the current node's sample set is used as the candidate splitting threshold. Under each candidate splitting threshold, the variance of the fusion intensity label value in the left child node sample set and the variance of the fusion intensity label value in the right child node sample set are calculated after splitting the current node's sample set into a left child node sample set and a right child node sample set. The left child node sample set consists of samples in the current node's sample set whose candidate feature values are less than or equal to the candidate splitting threshold, and the right child node sample set consists of samples in the current node's sample set whose candidate feature values are greater than the candidate splitting threshold. The variance of the fusion intensity label values in the left child node sample set is obtained by dividing the sum of the squares of the differences between all fusion intensity label values in the left child node sample set and the mean of the fusion intensity label values in the left child node sample set by the number of samples in the left child node sample set. Similarly, the variance of the fusion intensity label values in the right child node sample set is obtained by dividing the sum of the squares of the differences between all fusion intensity label values in the right child node sample set and the mean of the fusion intensity label values in the right child node sample set by the number of samples in the right child node sample set. The weighted sum of the variances of the fusion intensity label values in the left and right child node sample sets is used as the mean squared error corresponding to the candidate splitting threshold. The weights of the weighted sum are the proportions of the number of samples in the left child node sample set to the number of samples in the current node sample set, and the proportions of the number of samples in the right child node sample set to the number of samples in the current node sample set. After traversing all candidate splitting thresholds for all candidate features in the feature subset, the candidate feature and candidate splitting threshold with the smallest mean squared error are selected as the optimal splitting feature and optimal splitting threshold.
[0069] Based on the optimal splitting feature and the optimal splitting threshold, the sample set of the current node is split into a left child node sample set and a right child node sample set. The left child node sample set contains all samples in the current node sample set whose optimal splitting feature value is less than or equal to the optimal splitting threshold, and the right child node sample set contains all samples in the current node sample set whose optimal splitting feature value is greater than the optimal splitting threshold.
[0070] The node splitting process is executed recursively. The sample set of the left child node is used as the new current node sample set, and the sample set of the right child node is used as the new current node sample set. The process of randomly selecting a feature subset from the input feature vector, selecting the optimal splitting feature and the optimal splitting threshold based on the mean squared error criterion, and splitting the sample set according to the optimal splitting feature and the optimal splitting threshold is repeated for the two newly generated child nodes until the depth of the current regression decision tree reaches the maximum depth. Or the number of samples in the current node is less than the minimum number of samples required for node splitting. When the depth of the regression decision tree reaches its maximum depth... When the number of samples in the current node is less than the minimum number of samples required for node splitting, further splitting of the current node is stopped, and the current node is marked as a leaf node. When the current node is not further split, it is marked as a leaf node. Leaf nodes do not undergo any splitting operations; their predicted value is the arithmetic mean of the fusion strength label values of all training samples falling into that leaf node. For all... Each regression decision tree undergoes the above construction process to obtain... A set of trained regression decision trees, The trained regression decision trees together constitute the initial regression model.
[0071] Cross-validation and hyperparameter optimization were performed on the initial regression model. A five-fold cross-validation method was used to divide the training set into five subsets. All samples in the training set were randomly sorted and then divided into five equally sized subsets, each containing one-fifth of the total number of samples in the training set. Four subsets from each subset were used alternately as the training subset, and the remaining subset as the validation subset. This training and validation process was repeated five times, ensuring that each subset was used as a validation subset at least once.
[0072] In a single training and validation cycle, a cross-validation regression model is trained using all samples in the training subset according to the Bootstrap sampling method and the regression decision tree construction process. The normalized input feature vectors of all samples in the validation subset are then input into the cross-validation regression model. Each regression decision tree in the cross-validation regression model outputs a predicted value for each sample in the validation subset. The final predicted value of the cross-validation regression model is the arithmetic mean of the predicted values output by all regression decision trees, thus obtaining the predicted strength value for each sample in the validation subset.
[0073] The coefficient of determination between the predicted intensity value and the fusion intensity label value is calculated during each verification process. and root mean square error Coefficient of determination The calculation process is as follows: calculate the average value of all fusion intensity label values in the validation subset; calculate the sum of squares of the differences between all fusion intensity label values in the validation subset and the average value as the total sum of squares; calculate the sum of squares of the differences between the predicted intensity value and the corresponding fusion intensity label value for each sample in the validation subset as the residual sum of squares; and calculate the coefficient of determination. It equals the value 1 minus the sum of squared residuals divided by the total sum of squares. Root mean square error. The calculation process is as follows: Calculate the square of the difference between the predicted intensity value and the corresponding fused intensity label value for each sample in the validation subset; calculate the arithmetic mean of all squared values; and then take the square root of the arithmetic mean. Five determination coefficients are calculated during the five training and validation processes. The arithmetic mean was used as the average coefficient of determination for cross-validation. The five root mean square errors calculated during the five training and validation processes. The arithmetic mean was used as the mean root mean square error of cross-validation. .
[0074] Adjusting the hyperparameter combination of the random forest regression algorithm, including the number of decision trees. Maximum depth of decision tree Minimum number of samples for node splitting Number of decision trees Starting with an initial value of 100, the value is adjusted successively within the interval [50, 500] with a step size of 50. The maximum depth of the decision tree is determined by this step. Based on an initial value of 10, the value is adjusted successively within the interval [4, 30] with a step size of 2 to determine the minimum number of samples required for node splitting. Starting with an initial value of 5, the hyperparameters are adjusted sequentially within the interval [2, 20] with a step size of 1. For each hyperparameter combination, the five-fold cross-validation process is repeated to calculate the average coefficient of determination for that hyperparameter combination. and root mean square error Compare the average coefficient of determination for all combinations of hyperparameters. and root mean square error Select the average coefficient of determination To maximize and average root mean square error The hyperparameter combination that achieves minimization is taken as the optimized hyperparameter combination. When there are two or more sets of hyperparameter combinations, the average coefficient of determination... When all values are the same and reach their maximum value, select the average root mean square error. The smallest hyperparameter combination is taken as the optimized hyperparameter combination.
[0075] The optimized hyperparameter combination is used as the final hyperparameters for the random forest regression algorithm. Using all training set samples, the bootstrap sampling and regression decision tree construction processes are re-executed according to the optimized hyperparameter combination to train the mapping relationship model. The mapping relationship model is determined by the number of optimized decision trees. value constructed The system consists of several regression decision trees, each with a maximum depth equal to the maximum depth of the optimized decision tree. The minimum number of samples required for node splitting in each regression decision tree is the optimized minimum number of samples required for node splitting. value.
[0076] When inputting the field-measured mineral phase type and content data of the rock to be predicted into the mapping model, and outputting the predicted rock strength value, the field-measured mineral phase type data of the rock to be predicted is first encoded into a category feature vector, and the field-measured mineral phase content data of the rock to be predicted is encoded into a numerical feature vector. The category feature vector and the numerical feature vector are concatenated into an input feature vector. Normalization is then performed on the input feature vector, using the minimum value of each dimension during normalization. and maximum value The minimum value of the corresponding dimension used when training the mapping relationship model. and maximum value To maintain consistency, the normalized input feature vector is input into the mapping model. Each regression decision tree in the mapping model, based on the values of each feature in the input feature vector, starts from the root node and sequentially branches according to the optimal splitting feature and optimal splitting threshold stored at the node, until the sample falls into a leaf node. The predicted value stored in that leaf node is output. The predicted values output by all regression decision trees in the mapping model are arithmetically averaged to obtain the predicted rock strength value.
[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for intelligent prediction of rock strength parameters based on mineral data and model fusion, characterized in that, Includes the following steps: Obtain multi-source mineral parameters of the rock to be predicted. The multi-source mineral parameters include at least mineral phase type data, mineral phase content data, mineral type data, mineral structure data, elastic modulus data of each mineral, hardness data of each mineral, and measured macroscopic strength data of the rock. A three-dimensional digital core model is constructed based on the multi-source mineral parameters, and the three-dimensional digital core model is meshed based on the Voronoi algorithm to establish a microscopic numerical simulation model. Uniaxial compressive boundary conditions are applied to the microscopic numerical simulation model, and microscopic numerical simulation calculations are performed to extract the simulated stress-strain curves and simulated macroscopic strength data. Using the mineral phase type data and the mineral phase content data as input features, and the measured data of rock macroscopic strength and the simulated macroscopic strength data as label values, a machine learning regression model is trained to obtain a mapping relationship model; The data on the types and contents of mineral phases of the rock to be predicted, obtained from on-site measurements, are input into the mapping model to output the predicted rock strength value.
2. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 1, characterized in that, The steps for obtaining multi-source mineral parameters of the rock to be predicted specifically include: Portable X-ray diffractometer was used to obtain data on the types and contents of mineral phases in the rock. The TIMA automated mineral analysis system, coupled with a polarizing microscope, was used to obtain mineral type and mineral structure data of the rocks. The elastic modulus and hardness data of various minerals in the rock were obtained using a nanoindenter. The macroscopic strength data of the rock were obtained using a uniaxial testing machine.
3. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 1, characterized in that, The step of constructing a three-dimensional digital core model based on the multi-source mineral parameters specifically includes: The data on mineral phase types, mineral phase contents, mineral types, mineral structures, elastic modulus of each mineral, hardness of each mineral, and measured macroscopic strength of the rock are fused across scales. Typical microstructure parameters were determined based on the fused data; A three-dimensional digital core model was constructed based on the aforementioned typical microstructure parameters. The typical microstructural parameters include mineral particle size distribution data, mineral contact relationship data, porosity data, and cementation strength data.
4. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 1, characterized in that, The step of meshing the three-dimensional digital core model and establishing a mesoscopic numerical simulation model specifically includes: The discrete element method is used to generate a Voronoi grid numerical model based on the Voronoi grid algorithm; The micromechanical parameters of the Voronoi mesh numerical model are set, including mineral particle contact stiffness, mineral particle friction coefficient and mineral particle bonding strength. The Voronoi mesh numerical model with pre-set micromechanical parameters is used as the micromechanical numerical simulation model.
5. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 1, characterized in that, The steps of applying uniaxial compressive boundary conditions to the microscopic numerical simulation model, performing microscopic numerical simulation calculations, and extracting simulated stress-strain curves and simulated macroscopic strength data specifically include: A downward displacement load is applied to the upper surface of the microscopic numerical simulation model, and a fixed constraint is applied to the lower surface of the microscopic numerical simulation model. The lateral surface of the microscopic numerical simulation model is a free surface, which serves as a uniaxial compression boundary condition. The discrete element method iterative calculation was used to perform microscopic damage simulation calculations under uniaxial compression conditions; During the simulation, stress and strain data are recorded at each calculation step to generate simulated stress-strain curves; The peak stress is extracted from the simulated stress-strain curve as the simulated macroscopic strength data.
6. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 1, characterized in that, The step of training a machine learning regression model using the mineral phase type data and the mineral phase content data as input features, and the measured data of rock macroscopic strength and the simulated macroscopic strength data as label values, specifically includes: The mineral phase type data is encoded into a category feature vector, the mineral phase content data is encoded into a numerical feature vector, and the category feature vector and the numerical feature vector are concatenated into an input feature vector; The measured macroscopic strength data of the rock and the simulated macroscopic strength data are weighted and fused to obtain a fused strength label value; Using the input feature vector as the model input and the fusion strength label value as the model output target, the initial regression model is trained using the random forest regression algorithm. Cross-validation and hyperparameter optimization are performed on the initial regression model to obtain a trained mapping relationship model.
7. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 6, characterized in that, The steps for training the initial regression model using the random forest regression algorithm specifically include: The input feature vector is normalized to eliminate dimensional differences; Set the hyperparameters of the random forest regression algorithm, including the number of decision trees N, the maximum depth of the decision trees D, and the minimum number of samples M for node splits; The Bootstrap sampling method is used to draw N subsets of samples with replacement from the training set; A decision tree is constructed based on each subsample set, and a total of N decision trees are constructed. Each decision tree is trained based on the input feature vector and the fusion intensity label value.
8. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 7, characterized in that, The step of constructing a decision tree based on each subset of samples specifically includes: When splitting at each node of the current decision tree, a subset of features is randomly selected from the input feature vector; The optimal splitting feature and the optimal splitting threshold are selected from the feature subset based on the mean square error criterion. Based on the optimal splitting feature and the optimal splitting threshold, the sample set of the current node is split into a left child node sample set and a right child node sample set; The node splitting process is executed recursively until the current decision tree reaches the maximum depth D or the number of node samples is less than the minimum number of samples M required for node splitting.
9. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 8, characterized in that, The step of selecting the optimal splitting feature and the optimal splitting threshold from the feature subset based on the mean square error criterion specifically includes: For each candidate feature in the feature subset, all values of the candidate feature are traversed as the candidate splitting threshold. Calculate the rock strength variance of the left child node sample set and the rock strength variance of the right child node sample set under each candidate split threshold; Calculate the weighted sum of the rock strength variance of the left child node sample set and the rock strength variance of the right child node sample set, and use it as the mean square error corresponding to the candidate splitting threshold; The candidate feature and candidate splitting threshold with the smallest mean square error are selected as the optimal splitting feature and optimal splitting threshold.
10. The intelligent prediction method for rock strength parameters based on mineral data and model fusion according to claim 7, characterized in that, The steps of cross-validating and optimizing the hyperparameters of the initial regression model specifically include: The training set was divided into five subsets using the five-fold cross-validation method. Four subsets were used as the training subset and one subset was used as the validation subset in turn. The training and validation process was repeated five times. In each validation process, the coefficient of determination R² and the root mean square error RMSE between the predicted intensity value and the fused intensity label value are calculated. Adjust the number of decision trees N, the maximum depth of the decision trees D, and the minimum number of samples M for node splitting, and repeat the five-fold cross-validation process until the coefficient of determination R² is maximized and the root mean square error RMSE is minimized. The hyperparameter combination that maximizes the coefficient of determination R² and minimizes the root mean square error RMSE is taken as the optimized hyperparameter combination.
