Rock mineral micromechanical evolution characterization method based on convolutional neural network

Abstract

The invention provides a rock mineral micromechanical evolution characterization method based on a convolutional neural network. The rock mineral micromechanical evolution characterization method comprises the following steps: step 1, obtaining rock sample micromechanical properties through a nanoindentation experiment; 2, performing deconvolution statistical analysis by using a Gaussian mixture model, and obtaining mechanical and volume fractions of each phase through a nanoindentation experiment; step 3, utilizing a Mori-Tanaka scale upgrading method to understand a convolution result, and performing rock macromechanical property prediction in combination with a convolutional neural network; and step 4, performing precision checking calculation on an output result of the convolutional neural network model until the prediction precision meets a preset requirement. According to the method, the microscopic mechanical parameters of the rock sample are obtained through the grid nanoindentation, statistical analysis is carried out on the nanoindentation test result by utilizing a deconvolution method, the mechanical property of the macroscale of the rock is obtained in combination with machine learning and a scale upgrading model, and a bridge for connecting the macroscale and the microscale is established; the method can effectively solve the problem of inaccurate prediction of rock micromechanical parameters on the macroscale and the problem of mechanical property characterization caused by scale difference.

Description

Technical Field

[0001] This application relates to the field of rock mechanics, specifically a method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks. It is an improvement on traditional rock mechanics analysis methods and is particularly suitable for predicting the mechanical properties of rocks. Background Technology

[0002] As the core constituent material of the Earth's surface and deep spheres, rocks are key research subjects in fields such as geological evolution, resource development, engineering construction, and disaster prevention. From mineral extraction, oil and gas reservoir development, and geothermal utilization to tunnel engineering, slope stability, nuclear waste disposal, and carbon sequestration, the mechanical parameters of rocks, such as strength, elasticity, and fracture toughness, directly determine the safety threshold, construction efficiency, and long-term stability of engineering projects. Traditional macroscopic mechanical experiments are a classic means of characterizing the mechanical properties of rocks, but these methods generally face many limitations: on the one hand, the experiments have strict requirements for the dimensional regularity and integrity of the samples, while most natural rocks have primary fractures and pores, and some soft or brittle rocks (such as coal, shale, and limestone) also face the problem of difficulty in preparing standard samples and easy breakage; on the other hand, macroscopic experiments are destructive, and once a sample fails, it cannot be reused, making it difficult to achieve multi-dimensional and dynamic characterization of the mechanical properties of the same rock mass.

[0003] The emergence of nanoindentation technology has provided a new approach for studying the micromechanical properties of rocks. This technique requires only micrometer- or millimeter-scale experimental areas, causes minimal damage to samples, allows for repeatable testing, and can directly obtain parameters such as hardness, elastic modulus, and fracture toughness of specific micro-components in rocks (e.g., vitrinite or inertinite in coal). Studies have shown that the micromechanical properties of rocks exhibit significant heterogeneity; the elastic modulus and hardness of different components typically differ, and the mechanical parameters show a regular evolution with variations in the content of quartz and other elements in the rock. Furthermore, external environmental factors such as supercritical carbon dioxide can alter the pore structure by dissolving minerals, leading to a decline in the micromechanical properties of rocks, further highlighting the importance of microscopic research in explaining macroscopic phenomena.

[0004] While nanoindentation technology can accurately characterize microscopic mechanical properties, a significant gap exists between its measurement scale and macroscopic engineering applications. Directly extrapolating nanoscale data to the macroscopic scale ignores the interactions, porosity effects, and size effects of multiphase components in rocks. Therefore, scale-up research becomes a crucial bridge connecting microscopic mechanical parameters and macroscopic mechanical behavior. By introducing homogenization methods and training convolutional neural networks based on deep learning datasets, a cross-scale predictive model of the quantitative relationship between sample microscopic behavior and macroscopic response is established. Using the mechanical properties and volume fractions of each phase obtained from nanoindentation as input parameters, the equivalent macroscopic elastic modulus of the material can be theoretically predicted, thereby revealing the influence mechanisms of mineral content, pore structure, etc., on macroscopic mechanical behavior. Summary of the Invention

[0005] This application proposes a method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks, in order to accurately establish the relationship between the micromechanical behavior and macroscopic response of rocks and realize cross-scale prediction of rock mechanical properties.

[0006] The technical solution of this application is as follows: One aspect of this application provides a method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks, including: Step 1: Obtain the microscopic and macroscopic mechanical properties of the rock sample, including hardness and elastic modulus, through nanoindentation and uniaxial compression experiments. Obtain the mineral composition of the sample through XRD and obtain the planar rock composition using the fully automated mineral analysis system TIMA.

[0007] Step 2: Statistical analysis of the nanoindentation data was performed using the Gaussian mixture model deconvolution method to identify and quantify different rock phases. The phase of each sample was determined based on XRD and TIMA results, and the mechanical parameters and volume fraction of each phase were obtained through nanoindentation experiments. The deconvolution method was then validated and optimized. Step 3: Process the deconvolution results using the Mori-Tanaka scale upgrade method, and combine them with a fully automated convolutional neural network mineral analysis system to predict the macroscopic mechanical properties of rocks, upgrading to the macroscopic scale level; establish a deep learning dataset with nanoindentation data as input and macroscopic mechanical properties of rocks as output.

[0008] Step 4: Verify the accuracy of the output results of the convolutional neural network model until the prediction accuracy meets the preset requirements, thus accurately realizing the cross-scale prediction model of the quantitative relationship between the microscopic behavior and macroscopic response of the sample.

[0009] Preferably, in step 2, the Gaussian mixture model deconvolution method is used to perform statistical analysis on the nanoindentation data to identify and quantify different rock phases, including: a) Model construction and probability density function: The micromechanical properties of each phase are regarded as a two-dimensional Gaussian distribution with respect to elastic modulus and hardness using the Gaussian mixture model, wherein the probability density function follows a maximum likelihood estimation algorithm; b) Parameter estimation and EM algorithm: The micromechanical property parameters of each phase are obtained by iteratively optimizing the EM algorithm by maximizing the log-likelihood function.

[0010] Preferably, the parameter estimation of the Gaussian mixture model is performed in the following manner: for each data point, the probability of it belonging to each Gaussian cluster is calculated based on the current weight coefficients, mean, and covariance matrix; the mean and covariance matrix of each cluster are iteratively updated to maximize the joint probability density function of all data points until the convergence criterion of the EM algorithm is met.

[0011] Preferably, in step 2, the GMM deconvolution method is verified and optimized through the following operations to ensure the reliability of the phase mechanical parameters and volume fraction: a) The quantity and type of phases are determined and verified using X-ray diffraction and transmission infrared spectroscopy analysis results; b) The mineral distribution and volume fraction obtained by XRD and TIMA are compared with the nanoindentation experimental results, and the phase division is optimized by adjusting the deconvolution parameters until the phase mechanical parameters and volume fraction match the actual situation.

[0012] Preferably, in step 3, the deconvolution results are processed using the Mori-Tanaka scaling method, and the macroscopic mechanical properties of rocks are predicted by combining a convolutional neural network. Specifically, this includes: a) applying the Mori-Tanaka method to calculate the homogeneous bulk modulus and shear modulus of a material with n phases, where clay minerals are used as the reference medium; b) using the calculated homogeneous Young's modulus and Poisson's ratio, combined with the training of the convolutional neural network, to establish a deep learning dataset.

[0013] Preferably, the calculation steps of the Mori-Tanaka method are as follows: i) Based on the micromechanical property parameters and volume fraction of each phase, the homogenized bulk modulus and shear modulus of each phase are calculated using the Mori-Tanaka formula; ii) Using the modulus value of clay minerals as a reference medium, the calculation results are corrected to obtain the macroscopic homogenized bulk modulus and shear modulus of the rock material.

[0014] Preferably, the construction of the deep learning dataset and the training process of the convolutional neural network prediction model specifically include: a) collecting a large amount of nanoindentation data and corresponding measured values ​​of rock macroscopic mechanical properties to form a training dataset; b) using a convolutional neural network model, with microscopic mechanical parameters (X1, X2, …, Xn) as network input and rock macroscopic mechanical properties as output labels, optimizing network parameters through iterative training, and finally establishing a prediction model Y=F(X1,X2,⋯,Xn)+C, where C is the error term, to achieve prediction of mechanical behavior from microscopic to macroscopic scales.

[0015] Preferably, the accuracy verification process of the convolutional neural network model output in step 4 is as follows: a) Compare the macroscopic mechanical properties of the rock predicted by the model with the results obtained through actual macroscopic experiments, and evaluate the prediction accuracy using root mean square error or mean absolute error index; b) If the accuracy meets the preset threshold, verify the effectiveness of the model and output it directly as a core tool for research and application; otherwise, perform iterative optimization.

[0016] Preferably, iterative optimization includes: a) expanding the scale of the deep learning dataset, adding nanoindentation data and measured values ​​of macroscopic mechanical properties covering different working conditions and parameter combinations, and improving the representativeness and completeness of the dataset; b) retraining the convolutional neural network model based on the expanded dataset, adjusting the input parameter weights, optimizing the function mapping structure, and correcting the model constant term C to ensure that the model prediction accuracy meets the preset requirements; c) after optimization, performing the accuracy verification process of comparing the model prediction with the measured data again until the error between the predicted values ​​of rock macroscopic mechanical properties output by the model and the experimental true values ​​is controlled within the preset acceptable range, ensuring the reliability of the output cross-scale prediction model.

[0017] Preferably, in step 1, the workflow of TIMA (fully automated mineral analysis system) includes: acquiring backscattered electron images and X-ray energy dispersive spectroscopy data of the rock surface; generating a mineral planar distribution map and statistically analyzing the relative mineral content and other microstructural parameters through automatic classification and manual correction of misjudgments; acquiring mineral characteristics of the rock surface; using scanning electron microscopy and energy dispersive spectroscopy analysis to identify the mineral types and their distribution characteristics in the indentation area, thereby improving the reliability of the phase mechanical parameters.

[0018] The method proposed in this application obtains the micromechanical parameters of rock samples through grid nanoindentation, uses deconvolution method to statistically analyze the nanoindentation test results, and combines machine learning and scale-up models to obtain the macroscopic mechanical properties of rocks. It effectively integrates micromechanical experiments and macroscopic mechanical property prediction. By using Gaussian mixture model and convolutional neural network technology, it achieves accurate conversion from micromechanical properties of rocks and minerals to macroscopic mechanical behavior, breaking through the problem of scale conversion in traditional rock mechanics research.

[0019] The method proposed in this application processes the deconvolution results using the Mori-Tanaka scale-up method and combines it with deep learning technology, which not only improves the accuracy of predictions but also significantly enhances efficiency, enabling rapid and accurate analysis of the mechanical properties of complex rock samples.

[0020] This application establishes a bridge between macro and micro scales, which can effectively solve the problems of inaccurate prediction of rock micromechanical parameters at the macro scale and the difficulties in characterizing mechanical properties due to scale differences, providing strong technical support for geotechnical engineering, geological exploration, mineral resource development and other fields.

[0021] The application of the method proposed in this application can significantly reduce experimental costs, shorten the research cycle, and promote the development of rock mechanics theory and application technology. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The method flowchart of this application; Figures 2-5 Schematic diagram of the principle of rock nanoindentation; Figure 6 TIMA Fully Automated Mineral Analysis System Operating Principle Diagram; Figure 7 A schematic diagram of a fully automated mineral analysis system extracting mineral features from the surface of rocks. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the description and claims of this application contain a series of steps or units, and the process, method, system, product or device is not limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or that are inherent to these processes, methods, products or devices.

[0025] like Figure 1 The diagram shown is a flowchart of the method of this application. The method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks includes the following steps: Step 1: Obtain the micromechanical properties of the rock sample, including hardness and elastic modulus, through nanoindentation and uniaxial compression experiments, and obtain the mineral composition of the sample through XRD. The planar rock composition is obtained using the fully automated mineral analysis system TIMA.

[0026] The specific steps are as follows: 1-1) Sample Preparation In this embodiment, in accordance with the relevant standard recommendations of the International Society for Rock Mechanics (ISRM), a cylindrical standard specimen with a size of 25 mm × 50 mm was drilled and processed from the experimental rock core to ensure that the end face of the specimen was flat and the axis was perpendicular to meet the specifications of the uniaxial compression test.

[0027] 1-2) Uniaxial compression test Using standard specimens that have been processed, uniaxial compression tests are conducted on a mechanical testing machine. By recording the stress-strain curves throughout the entire process, the basic mechanical parameters of the specimens, such as Young's modulus and Poisson's ratio, are analyzed to obtain the parameters on a macroscopic scale.

[0028] 1-3) Preparation of microscopic samples Considering the load-bearing capacity of laboratory mechanical testing equipment and the special requirements of nanoindentation experiments, this embodiment uses a small square block with a size of approximately 1 mm × 1 mm × 0.5 mm cut from the same experimental rock core as a sample for micromechanical testing, ensuring that the sample surface is smooth and flat to meet the requirements of nanoindentation testing.

[0029] 1-4) See nanoindentation experiments Figures 2-3 ,include: a. Sample preparation and equipment debugging Sample fixation: Attach the sample to be tested (such as composite materials containing particles / interfaces) to the sample stage, ensuring that the sample surface is flat and free of impurities.

[0030] Equipment calibration: Locate the indentation area using an optical microscope, calibrate the initial position of the indenter, and set the thermal drift threshold (to avoid the influence of temperature on displacement measurement).

[0031] b. The indentation process (loading-holding-unloading) corresponds to Figure 4 The load-displacement curve.

[0032] During the loading phase, the indenter is controlled to apply a load to the sample surface at a set rate, and the indenter is gradually pressed into the sample. Figure 3 From the initial surface to (hm), the material undergoes elastic and plastic deformation, see... Figure 4 The Loading section.

[0033] Load holding phase: After reaching the maximum load Pmax, the load is maintained for a period of time. Figure 4 (See the Holding section in the text) to observe the creep behavior of the material. Figure 3 The deformation from hc to ht in the middle.

[0034] Unloading phase: The load is removed at a set rate, and the elastic deformation of the material partially recovers. Figure 3 The middle surface springs back to hr, and the load-displacement curve during the unloading process is recorded. See Figure 4 The Unloading section.

[0035] c. Data Acquisition and Analysis Core curve: To obtain the complete load-displacement curve, see... Figure 4 Parameters such as maximum depth hm, contact depth hc, and unloading stiffness S are extracted.

[0036] Mechanical parameter calculation: The hardness and elastic modulus of the material are calculated using the Oliver-Pharr method combined with the area function of the indenter.

[0037] Multi-region testing: testing different regions of the sample (e.g.) Figure 2 The distribution law of statistical mechanical properties of repeated indentations (matrix, particles, and interfaces) in the matrix is ​​analyzed, and the corresponding... Figure 5 The probability density curve in the image.

[0038] 1-5) Microstructure observation and composition analysis The morphology of the nanoindentation experimental area was observed using scanning electron microscopy (SEM), and energy dispersive spectroscopy (EDS) was used to identify the types of minerals contained in the indentation area and their distribution characteristics, thus establishing the correspondence between micromechanical properties and mineral composition.

[0039] 1-6) X-ray diffraction mineral composition analysis Phase analysis of core samples was performed using X-ray diffraction (XRD). By identifying characteristic peaks in the diffraction patterns, the composition and relative content of various minerals in the core were qualitatively and semi-quantitatively determined, providing mineralogical basis for scale-up analysis and analysis of macroscopic-microscopic mechanical behavior.

[0040] 1-7) The TIMA fully automated mineral analysis system obtains planar rock composition. For TIMA operating principles, please refer to Figure 6 : a. Sample pretreatment The rock sample is cut, ground, and polished until the surface is flat (ensuring clear exposure of mineral particles), and a conductive layer (such as a carbon film) is sprayed onto the surface to avoid charge accumulation during subsequent electron microscopy observation.

[0041] b. Scan Area and Parameter Settings The sample is fixed in the TIMA system, and the planar area of ​​the rock to be analyzed (usually a regular grid area, as shown in the figure) is selected by the software. The scanning parameters are set, such as electron beam accelerating voltage, beam current, scanning step size, and signal acquisition threshold.

[0042] c. Point-by-point acquisition of mineral signals (EDS energy dispersive spectroscopy) TIMA's electron beam scans each "test point" in the selected area according to a preset grid, while simultaneously acquiring EDS energy spectrum signals (the probe in the figure): the EDS signal of each test point corresponds to the elemental composition at that location (such as the characteristic spectral peaks of elements like Si, Al, and Fe).

[0043] d. Defining minerals using BSE / EDS signals BSE signal-assisted boundary delineation: The gray level of the backscattered electron (BSE) signal is positively correlated with the atomic number of the mineral, which can preliminarily distinguish the boundaries of different minerals; Accurate EDS signal identification: By combining the EDS element composition of each test point and matching the standard element ratio of the mineral, the boundary of each mineral in the rock plane is finally determined (the colored area in the figure represents different minerals).

[0044] e. Statistical analysis of rock composition The software summarizes the mineral identification results of all test points, calculates the area ratio of each mineral in the rock plane, and finally outputs the quantitative results of the mineral composition of the rock (as shown in the bar chart, which shows the proportion of different minerals).

[0045] See rock surface mineral characteristics extracted by a fully automated mineral analysis system. Figure 4 .

[0046] Step 2: Statistical analysis of nanoindentation data is performed using the Gaussian mixture model deconvolution method to identify and quantify different rock phases; the phase of each sample is determined based on XRD and TIMA results, and the mechanical parameters and volume fraction of each phase are obtained through nanoindentation experiments to verify and optimize the deconvolution method.

[0047] In one embodiment, it specifically includes: 2-1) The composition phase of the nanoindentation data was determined by deconvolution using a Gaussian mixture model (GMM).

[0048] a. Model construction and probability density function: The micromechanical properties of each phase are treated as a two-dimensional Gaussian distribution with respect to the elastic modulus and hardness using the GMM, where the probability density function follows the maximum likelihood estimation (MLE) algorithm.

[0049] Assuming that the micromechanical properties of each phase of the sample are considered in relation to the elastic modulus and hardness The two-dimensional Gaussian distribution follows a Gaussian mixture model based on the maximum likelihood estimation (MLE) algorithm. Its probability density function is: The probability density of the single-component two-dimensional Gaussian distribution is: in, It is a column vector of two-dimensional nanoindentation data points. It is the first The weighting coefficients of each component satisfy the following conditions: , and They are the first The mean and covariance matrix of each component (k=1,2…n).

[0050] b. Parameter Estimation and EM Algorithm: The EM algorithm is used for iterative optimization by maximizing the log-likelihood function to obtain the micromechanical property parameters of each phase. This includes calculating the probability of each data point belonging to each Gaussian cluster based on the current weight coefficients, mean, and covariance matrix; iteratively updating the mean and covariance matrix of each cluster to maximize the joint probability density function of all data points until the convergence criterion of the EM algorithm is met.

[0051] In one embodiment, it specifically includes: First, for each data point and each cluster Calculate the probability of belonging to this cluster. This is based on the current weighting coefficients. and Gaussian distribution parameter mean and covariance matrix of: After obtaining the probability of belonging to each cluster, the mean parameter of each cluster is iteratively updated. and covariance matrix To maximize the joint probability of all data points.

[0052] Where, N k is the number of valid data points belonging to the k-th component, and N is the total number of data points. Through iterative convergence of the EM algorithm, the deconvolution result of the multi-phase micromechanical property distribution of the sample is obtained, which is used for subsequent quantitative determination of the elastic modulus and hardness of each phase.

[0053] In this embodiment, a Gaussian Mixture Model (GMM) deconvolution method was employed to perform refined statistical analysis on nanoindentation data, aiming to identify and quantify the micromechanical characteristics of different rock phases. Specifically, regarding model construction and probability density functions, the micromechanical properties of each phase were treated as a two-dimensional Gaussian distribution of elastic modulus and hardness, and a probability density function based on the Maximum Likelihood Estimation (MLE) algorithm was constructed, laying the mathematical foundation for subsequent parameter estimation. In the parameter estimation and EM algorithm stage, by maximizing the log-likelihood function and applying the iterative optimization strategy of the EM algorithm, the key parameters of the micromechanical properties of each phase, including the average value and distribution characteristics of elastic modulus and hardness, were accurately solved. This series of statistical analyses and iterative processes not only ensured the accuracy of phase identification but also quantified the volume fraction of each phase, providing solid data support for subsequent scale-up and cross-scale mechanical behavior prediction. Through the scientific application of the GMM deconvolution method, the complex variations and aliasing problems existing in the nanoindentation experimental data were effectively addressed, enabling in-depth analysis of the micromechanical properties of rocks.

[0054] 2-2) Based on the results of the nanoindentation experiment, the sample will be divided into different phases, and the mechanical parameters and volume fraction of each phase will be obtained. The accuracy of the phase division and deconvolution method will be verified based on the XRD and TIMA results.

[0055] When classifying phases using nanoindentation, mechanical parameters are extracted from the load-displacement curves. Indentation points with similar characteristics are clustered into phase groups. After outlier removal, the volume fractions are statistically analyzed, and mechanical parameters are calculated. To verify accuracy, XRD and TIMA are combined: XRD identifies mineral phase types and verifies the number of phases through characteristic peaks; TIMA provides the distribution and volume fraction of minerals on the rock surface, allowing for comparison with the indentation results. Deconvolution parameters are then adjusted and optimized to ensure the reliability of phase mechanical parameters and volume fractions.

[0056] In this embodiment, the GMM deconvolution method is verified and optimized by analyzing the TIMA results using X-ray diffraction (XRD) and transmission infrared spectroscopy (TEM) to ensure the reliability of phase mechanical parameters and volume fractions. Specifically, XRD analysis is combined with TEM to determine the quantity and type of phases, and the results are compared with TIMA to provide mineralogical basis for phase identification in nanoindentation experiments. The mineral distribution and volume fraction data obtained from XRD and TIMA are compared with the nanoindentation experimental results, and the phase classification is optimized by adjusting the deconvolution parameters until the phase mechanical parameters and volume fractions match the actual situation. This process not only verifies the accuracy of the GMM deconvolution method but also ensures that the deconvolution results can truly reflect the microscopic mechanical properties of the rock sample, providing a reliable data foundation for subsequent scale upgrades and the establishment of deep learning models. Through this series of verification and optimization steps, the calculation results of phase mechanical parameters and volume fractions are more accurate, thereby achieving accurate prediction of the macroscopic mechanical behavior of rocks. Furthermore, the optimized GMM deconvolution method improves the efficiency and accuracy of nanoindentation data processing, reduces errors in the data processing process, and lays a solid foundation for establishing a predictive model of the mechanical properties of rocks from microscopic to macroscopic.

[0057] Step 3: Based on the convolution results, input the scale-up model to establish a cross-scale prediction model for the quantitative relationship between the microscopic behavior and macroscopic response of the sample. This includes using the Mori-Tanaka scale-up method to process the deconvolution results, combining it with a convolutional neural network to predict the macroscopic mechanical properties of the rock, and upgrading to the macroscopic scale level; establishing a deep learning dataset with nanoindentation data as input and the macroscopic mechanical properties of the rock as output.

[0058] In one embodiment, it specifically includes: 3-1) The deconvolution results are processed using the Mori-Tanaka scale-up method. The specific calculation steps of the Mori-Tanaka method are as follows: i) Based on the micromechanical property parameters and volume fraction of each phase, the homogenized bulk modulus and shear modulus of each phase are calculated using the Mori-Tanaka formula; ii) Using the modulus value of clay minerals as a reference medium, the calculation results are corrected to obtain the macroscopic homogenized bulk modulus and shear modulus of the rock material.

[0059] Specifically, Assuming the material consists of multiple inclusions randomly distributed and embedded in the matrix material, and the far-field strain experienced by each inclusion is equal to the average strain of the matrix, the homogeneous bulk modulus and shear modulus of a material with n phases can be calculated using the following formulas: in, and They are phases The bulk modulus and shear modulus. The bulk modulus and shear modulus can be determined based on Young's modulus. Compared to Poisson Calculate, and These are the bulk modulus and shear modulus of the reference medium, respectively. In this study, clay minerals are assumed to be the reference medium due to their abundance in the samples and their more widely known modulus values.

[0060] Homogenized Young's modulus Poisson ratio It can be calculated using the following formula: 3-2) Based on the above method, a deep learning dataset is established to train a convolutional neural network model, using nanoindentation data as input and the macroscopic mechanical properties of the sample as output. The training dataset is formed by collecting a large amount of nanoindentation data and corresponding measured values ​​of the macroscopic mechanical properties of the rock. Using the above model to predict macroscopic mechanical properties Y, based on microscopic mechanical parameters (X1, X2, ..., Xn), the basic form of the prediction model is as follows: +C Where C represents the error term.

[0061] In this embodiment, step 3 utilizes the Mori-Tanaka scale-up method to process the deconvolution results and combines them with a convolutional neural network to predict the macroscopic mechanical properties of rocks. Specifically, by applying the Mori-Tanaka method, the homogeneous bulk modulus and shear modulus of a material with n phases are calculated, using clay minerals as the reference medium. This step achieves an effective conversion from microscopic mechanical parameters to macroscopic mechanical properties. Furthermore, using the calculated homogenized Young's modulus and Poisson's ratio, combined with the training of the convolutional neural network, a deep learning dataset is established. This dataset takes nanoindentation data as input and outputs the macroscopic mechanical properties of the sample, achieving quantitative prediction between the microscopic behavior and macroscopic response of rocks. This deep learning-based dataset training not only improves the accuracy and reliability of the prediction model but also provides a novel cross-scale prediction method for rock mechanics research.

[0062] Furthermore, in this embodiment, the construction of the deep learning dataset and the training process of the convolutional neural network prediction model were meticulously planned. First, a) a large amount of microscopic mechanical parameter data derived from nanoindentation experiments were integrated, along with actual measured values ​​of the macroscopic mechanical properties of rocks, forming a comprehensive training dataset covering mechanical responses from microscopic to macroscopic levels. Subsequently, b) a convolutional neural network model was deployed, using microscopic mechanical parameters (X1, X2, …, Xn) as input variables and the macroscopic mechanical properties of rocks as output targets. Through multiple rounds of iterative training, the network architecture and parameters were continuously optimized, ultimately constructing an accurate prediction model Y=F(X1,X2,⋯,Xn)+C. C in the model represents the error term, used to quantify prediction bias. This process achieves efficient conversion from microscopic mechanical data to macroscopic mechanical properties, significantly enhancing the accuracy and reliability of cross-scale predictions. During the continuous training and optimization of the model, not only was the accuracy of microscopic data analysis improved, but the model's ability to characterize the mechanical evolution of complex rock systems was also strengthened, providing a powerful analytical tool for rock mechanics research and engineering applications.

[0063] Step 4: Verify the accuracy of the output results of the convolutional neural network model, and perform iterative optimization if necessary until the prediction accuracy meets the preset requirements, so as to accurately realize the cross-scale prediction model of the quantitative relationship between the microscopic behavior and macroscopic response of the sample.

[0064] In one embodiment, specifically: 4-1) First, the output of the cross-scale prediction model is systematically compared and analyzed with the measured data obtained from macroscopic experiments, focusing on verifying the degree of agreement between the model's predicted values ​​and the experimental values. If the comparison results show that the error level is controlled within the preset acceptable threshold range (the root mean square error, mean absolute error, and other accuracy indicators meet the project's set standards), then the cross-scale prediction model is directly output as a core tool for subsequent research or application.

[0065] 4-2) If the error does not meet the standard, the sample size and coverage of the deep learning prediction dataset should be expanded first (supplementing experimental or simulated data under different working conditions and parameter combinations to improve the representativeness and completeness of the dataset). Based on the new data, the prediction model Y=F(X1,X2,⋯,Xn)+C should be iteratively optimized—including adjusting the weights of the model input parameters, optimizing the function mapping structure, and correcting the constant term C. After optimization, the closed-loop process of "model prediction - macroscopic experimental data comparison - error evaluation" should be executed again until the accuracy index of the model output fully meets the preset requirements. Finally, a reliable cross-scale prediction model that has been fully validated and optimized should be output.

[0066] In this embodiment, the accuracy verification of the convolutional neural network model output includes two key steps. Step 4-1) aims to ensure that the agreement between the model's predicted values ​​and the experimental true values ​​meets a preset standard, thereby verifying the model's effectiveness. Step 4-2) proposes an iterative optimization strategy when the accuracy does not meet the standard. The model needs to be optimized by increasing the diversity and size of the training dataset, adjusting the weights of the input parameters, optimizing the function mapping structure, and correcting the constant term C, until the accuracy index of the model output fully meets the preset requirements. This rigorous accuracy verification and continuous optimization process ensures that the model can accurately achieve cross-scale prediction between the microscopic behavior and macroscopic response of rocks.

[0067] In summary, this application proposes a method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks. By integrating nanoindentation experiments, uniaxial compression experiments, XRD, TIMA techniques, Gaussian mixture model deconvolution, Mori-Tanaka scaling methods, and deep learning technology, it achieves rapid and accurate conversion from micromechanical properties to macroscopic mechanical behavior of rocks and minerals. This technical solution not only significantly improves the efficiency and accuracy of rock mechanical property analysis but also reduces experimental costs and shortens the research cycle. It provides an innovative and efficient technical means for geotechnical engineering, geological exploration, and mineral resource development, possessing significant academic value and broad application prospects.

[0068] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0069] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks, characterized in that, include: Step 1: Obtain the microscopic and macroscopic mechanical properties of the rock sample, including hardness and elastic modulus, through nanoindentation and uniaxial compression experiments, and obtain the mineral composition of the sample through XRD. Obtain the planar rock composition through the fully automated mineral analysis system TIMA. Step 2: Statistical analysis of nanoindentation data was performed using the Gaussian mixture model (GMM) deconvolution method to identify and quantify different rock phases; the phase of each sample was determined based on XRD and TIMA results, and the mechanical parameters and volume fraction of each phase were obtained through nanoindentation experiments to verify and optimize the deconvolution method. Step 3: Process the deconvolution results using the Mori-Tanaka scale-up method, and combine them with a convolutional neural network to predict the macroscopic mechanical properties of rocks, thus upgrading them to the macroscopic scale level; establish a deep learning dataset with nanoindentation data as input and macroscopic mechanical properties of rocks as output; Step 4: Verify the accuracy of the output results of the convolutional neural network model until the prediction accuracy meets the preset requirements, so as to accurately realize the cross-scale prediction model of the quantitative relationship between the microscopic behavior and macroscopic response of the sample.

2. The method for characterizing the micromechanical evolution of rocks and minerals according to claim 1, characterized in that, Step 2 uses a Gaussian mixture model (GMM) deconvolution method to perform statistical analysis on the nanoindentation data to identify and quantify different rock phases, including: a) Model construction and probability density function: The micromechanical properties of each phase are treated as a two-dimensional Gaussian distribution with respect to the elastic modulus and hardness using the GMM, where the probability density function follows the maximum likelihood estimation (MLE) algorithm. b) Parameter estimation and EM algorithm: By maximizing the log-likelihood function, the EM algorithm is used for iterative optimization to obtain the micromechanical property parameters of each phase.

3. The method according to claim 2, characterized in that, The parameter estimation of the Gaussian mixture model is performed in the following way: for each data point, the probability of it belonging to each Gaussian cluster is calculated based on the current weight coefficients, mean, and covariance matrix; the mean and covariance matrix of each cluster are iteratively updated to maximize the joint probability density function of all data points until the convergence criterion of the EM algorithm is met.

4. The method according to claim 2, characterized in that, In step 2, the GMM deconvolution method is verified and optimized through the following operations to ensure the reliability of the phase mechanical parameters and volume fraction: a) Using X-ray diffraction (XRD) and transmission infrared spectroscopy (TIMA) results, determine and verify the number and types of phases; b) Compare the mineral distribution and volume fraction obtained by XRD and TIMA with the results of nanoindentation experiments. Optimize the phase division by adjusting the deconvolution parameters until the phase mechanical parameters and volume fraction match the actual situation.

5. The method according to claim 1, characterized in that, Step 3 uses the Mori-Tanaka scaling method to process the deconvolution results and combines them with a convolutional neural network to predict the macroscopic mechanical properties of rocks. Specifically, this includes: a) Calculate the homogeneous bulk modulus and shear modulus of a material with n phases using the Mori-Tanaka method, with clay minerals as the reference medium; b) Using the calculated homogenized Young's modulus and Poisson's ratio, and combining them with the training of the convolutional neural network, a deep learning dataset is established.

6. The method according to claim 5, characterized in that, The calculation steps of the Mori-Tanaka method are as follows: i) Based on the micromechanical properties and volume fraction of each phase, the homogenized bulk modulus and shear modulus of each phase are calculated using the Mori-Tanaka formula. ii) Using the modulus value of clay minerals as a reference medium, the calculation results are corrected to obtain the macroscopic homogenized bulk modulus and shear modulus of the rock material.

7. The method according to claim 5, characterized in that, The construction of the deep learning dataset and the training process of the convolutional neural network prediction model specifically include: a) Collect a large amount of nanoindentation data and corresponding measured values ​​of rock macroscopic mechanical properties to form a training dataset; b) Using a convolutional neural network model, with microscopic mechanical parameters (X1, X2, …, Xn) as network input and macroscopic mechanical properties of rock as output labels, the network parameters are optimized through iterative training, and finally a prediction model Y=F(X1, X2, …, Xn)+C is established, where C is the error term, to achieve prediction of mechanical behavior from microscopic to macroscopic scales.

8. The method for characterizing the micromechanical evolution of rocks and minerals based on convolutional neural networks according to claim 1, characterized in that, The accuracy verification process for the output of the convolutional neural network model in step 4 is as follows: a) Compare the model-predicted macroscopic mechanical properties of rocks with the results obtained through actual macroscopic experiments, and use the root mean square error (RMSE) or mean absolute error (MAE) index to evaluate the prediction accuracy. b) If the accuracy meets the preset threshold, verify the effectiveness of the model and output it directly as a core tool for research and application; otherwise, perform iterative optimization.

9. The method according to claim 8, characterized in that, The iterative optimization includes: a) Expand the size of the deep learning dataset, add nanoindentation data and measured values ​​of macroscopic mechanical properties covering different working conditions and parameter combinations, and improve the representativeness and completeness of the dataset; b) Based on the expanded dataset, retrain the convolutional neural network model, adjust the input parameter weights, optimize the function mapping structure, and correct the model's normal coefficients C to ensure that the model's prediction accuracy meets the preset requirements; c) After optimization, the accuracy verification process of comparing model predictions with measured data is executed again until the error between the predicted values ​​of rock macroscopic mechanical properties output by the model and the actual experimental values ​​is controlled within the preset acceptable range, so as to ensure the reliability of the output cross-scale prediction model.

10. The method for characterizing the micromechanical evolution of rocks and minerals according to any one of claims 1 to 9, characterized in that, In step 1, the TIMA workflow includes: acquiring backscattered electron images and X-ray energy dispersive spectroscopy data of the rock surface; generating a mineral planar distribution map and statistically analyzing the relative content of the minerals and other microstructural parameters through automatic classification and manual correction of misjudgments; obtaining the mineral characteristics of the rock surface is achieved by using scanning electron microscopy (SEM) and energy dispersive spectroscopy (EDS) to identify the mineral types and their distribution characteristics in the indentation area, thereby improving the reliability of the phase mechanical parameters.

Images