A reef limestone mechanical property prediction method, device, equipment and storage medium

Abstract

This invention belongs to the interdisciplinary field of rock mechanics, engineering geological exploration, and artificial intelligence. Specifically, it relates to a method for predicting the mechanical properties of reef limestone. The method includes: acquiring and preprocessing a Micro-CT scan image sequence of reef limestone to obtain a binary three-dimensional digital core model of the reef limestone; then constructing a skeleton network model using the maximum sphere algorithm to extract node features and edge attribute features; next, inputting the node features and edge attribute features into a preset physical information graph neural network, and performing feature embedding and high-dimensional mapping through its mechanical message passing layer to obtain bias features that highly conform to the laws of rock mechanics; finally, aggregating the bias features of each node to obtain a full-image feature vector representing the macroscopic state of the reef limestone, and inputting it into two independent multilayer perceptron regression branches to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of rock mechanics, engineering geological exploration and artificial intelligence, and specifically relates to a method, device and equipment for predicting the mechanical properties of reef limestone and a storage medium. Background Technology

[0002] Reef limestone, as a core foundation and building material for major marine infrastructure projects such as marine island and reef engineering, cross-sea bridges, and subsea tunnels, has its mechanical properties so accurately assessed that the load-bearing safety, stability, and long-term service life of the engineering structure are directly determined. This is a crucial prerequisite for ensuring high-quality construction and operation of marine engineering projects. Compared to conventional terrigenous sedimentary rocks, reef limestone has a unique formation environment, formed from the long-term compaction, cementation, and recrystallization of marine biological remains such as corals and algae. It inherently possesses significant characteristics such as high porosity, strong heterogeneity, a fragile biological skeleton structure, and a complex interconnected pore network. These inherent properties lead to extremely strong spatial variability, nonlinearity, and anisotropy in its mechanical response, posing a significant challenge to engineering mechanical assessment. Repeated engineering practice has verified that the mechanical strength of reef limestone is not solely determined by total porosity, but is also closely related to microscopic features such as the topological structure of the micro-skeleton, the type and strength of the cement, and the distribution and transmission efficiency of the internal force chain network. Even subtle differences in the microstructure can cause significant fluctuations in mechanical properties.

[0003] Currently, the assessment methods for the mechanical properties of reef limestone, both domestically and internationally, still have significant shortcomings, making it difficult to meet the actual needs of major marine engineering projects for efficient, accurate, and reliable assessments. They mainly face three core bottlenecks, as follows: (1) Limited indoor testing methods and insufficient data representativeness. Reef limestone is brittle and fragile, with a well-developed pore structure and complex connectivity. It is prone to disintegration and fragmentation during field drilling and core sampling and indoor sample preparation, resulting in a very low success rate in preparing standard test specimens. It is often difficult to obtain sufficient and qualified specimens for mechanical testing. At the same time, indoor destructive mechanical tests (such as uniaxial compressive strength and shear tests) are not only costly and complicated, but also have large data dispersion. The test results of a single sample are difficult to reflect the heterogeneous characteristics of the reef limestone in the entire island and reef site, and cannot provide comprehensive and reliable mechanical parameter support for engineering design.

[0004] (2) Traditional numerical simulation faces a serious computational disaster and low efficiency. At present, the three-dimensional microstructure reconstruction technology of reef limestone based on CT scan images provides a visual basis for the simulation of mechanical properties. However, in order to accurately restore the micropores and skeleton structure, micron-level fine meshes are required for discretization. For centimeter-level reef limestone samples, mesh division alone can generate hundreds of millions of computational units. The single numerical simulation operation takes several days to several weeks, and the computational efficiency is extremely low. It cannot meet the urgent need for rapid assessment of mechanical properties in the marine engineering survey and design stage, and it is difficult to adapt to the timeliness requirements of the engineering site.

[0005] (3) Pure data-driven AI models lack physical foundation and have weak generalization ability. Existing reef limestone mechanical property prediction models based on deep learning methods such as convolutional neural networks (CNN) rely solely on the pixel texture statistical features of CT images for modeling, completely ignoring the core physical essence of reef limestone such as the connection characteristics of the micro-skeleton and stress transfer mechanism. This results in a lack of physical interpretability of the model and may even lead to prediction results that violate the basic common sense of rock mechanics. At the same time, such pure data-driven models have a huge demand for labeled data, while the collection of island and reef limestone samples is limited by marine environment and drilling conditions. The number of samples is scarce and the acquisition cost is high. In small sample scenarios, the model is prone to overfitting, and the prediction accuracy for unseen samples drops significantly. The generalization ability is difficult to meet the needs of complex scenarios in actual engineering.

[0006] In summary, current technologies for assessing the mechanical properties of reef limestone all suffer from insurmountable limitations: laboratory experiments are constrained by the difficulty of sample acquisition, traditional numerical simulations are limited by computational bottlenecks, and pure AI models are limited by the lack of physical foundations and data dependence. Therefore, overcoming the computational limitations of traditional numerical simulations, compensating for the physical blind spots and data dependence of AI models, and constructing an intelligent prediction method for the mechanical properties of reef limestone that is efficient, accurate, and physically interpretable, to achieve rapid and accurate assessment of reef limestone mechanical parameters, has become a critical technical challenge urgently needing to be solved in the fields of rock mechanics and marine engineering. This has significant theoretical and engineering application value for promoting the safe construction of major marine engineering projects, reducing engineering risks, and extending the service life of engineering projects. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a method, device, equipment, and storage medium for predicting the mechanical properties of reef limestone. It effectively overcomes the defect of data-driven approaches being disconnected from physical laws in existing technologies, solves the technical problems of low prediction accuracy and poor robustness of mechanical parameters of complex porous media under small sample conditions, and realizes efficient, accurate, and non-destructive evaluation of macroscopic mechanical properties based on microstructure.

[0008] The technical means adopted by this invention to solve its technical problem is: a method for predicting the mechanical properties of reef limestone, the method comprising: S1. Obtain the Micro-CT scan image sequence of the reef limestone and preprocess it to obtain the binary three-dimensional digital core model of the reef limestone. S2. Based on the binarized three-dimensional digital core model, the maximum sphere algorithm is used to construct the skeleton network model of the reef limestone core, and the node features and edge attribute features of each node in the skeleton network model are extracted. S3. Input the node features and edge attribute features into a preset physical information graph neural network, and perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly conform to the laws of rock mechanics. S4. Aggregate the bias features of each node to obtain a full-map feature vector representing the macroscopic state of the reef limestone. Input the full-map feature vector into two independent multilayer perceptron regression branches in a preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

[0009] Preferably, S1 includes: The reef limestone was non-destructively scanned using an industrial Micro-CT scanner to obtain an image sequence. After performing median filtering for noise reduction and three-dimensional threshold segmentation on the image sequence, a binarized three-dimensional digital core model of the reef limestone is obtained.

[0010] Preferably, S2 includes: The maximum sphere algorithm is used to find several inscribed spheres in the binary three-dimensional digital core model. Based on the centers of the inscribed spheres and their connection relationships, a skeleton network model is constructed. The centroid of the skeleton particles in the skeleton network model is defined as node V in graph theory, and the feature definition formula of node V is as follows: (1); Among them, X v R represents the feature vector of node V; v V represents the equivalent inscribed sphere radius of the skeletal particle; v H represents the actual volume of the skeletal particles; m This represents the local Vickers hardness value of the mineral composition obtained through nanoindentation testing or mineral composition inversion; C n This indicates the coordination number of node V; T represents the calculation of the transpose of the matrix. The physical contact surface between adjacent skeletal particles is detected using a three-dimensional morphological dilation algorithm. This physical contact surface is defined as an edge attribute E in graph theory, and the characteristic definition formula for edge attribute E is as follows: (2); Among them, E uvA represents the feature vector of the edge connecting nodes U and V; uv D represents the effective contact area between two skeletal particles; uv This represents the Euclidean distance between the centroids of two nodes; ξ represents the cosine of the angle between the normal direction of the contact surface and the macroscopic principal stress loading axis; rough It represents the geometric curvature of the contact surface, used to characterize the roughness of the contact interface and the interlocking friction effect.

[0011] Preferably, S3 includes: The node features and edge attribute features are input into a multilayer perceptron for normalization processing. The normalized features are input into a graph attention network embedded with a specific physical bias function, and the bias features of node V that are highly consistent with the laws of rock mechanics are calculated based on the attention coefficients of the graph attention network. The formula for calculating the attention coefficient is as follows: (3); Where, α uv h represents the influence weight of node U on node V. u ,h v Let U be the high-dimensional latent feature vectors of nodes U and V; W is the learnable linear transformation weight matrix; || denotes the vector concatenation operation; a T is the learnable parameter vector for the attention mechanism; N(u) represents the set of all first-order neighbor nodes of node U; LeakyReLU is a modified linear unit with leakage, which is an improved version of the classic ReLU function; Wherein, the physical bias function Ψ(E) uv The mathematical expression for ) is: (4); Wherein, β1 and β2 are the physical scaling factors that the physical information graph neural network adaptively learns during the training process; According to the attention coefficient α uv The offset characteristic of node V height conforming to the laws of rock mechanics was calculated, and the calculation formula is as follows: (5); Where σ is a nonlinear activation function, h v ” The offset characteristic of node V is that it conforms to the laws of rock mechanics.

[0012] Preferably, S4 includes: The bias features of several nodes are aggregated to obtain the full-map feature vector representing the entire reef limestone mass. The calculation formula is as follows: (6); Among them, h u Gate(h) represents the final hidden state of node U after deep network computation; u ) represents a learnable gating network; ⨀ represents the Hadamard product; h Graph This is the final generated full-image feature vector; The full-image feature vector is input into two independent multilayer perceptron regression branches, which output the elastic modulus prediction result and the uniaxial compressive strength prediction result, respectively. The calculation formula is as follows: (7); (8); Among them, E pred and UCS pred W represents the predicted macroscopic elastic modulus and uniaxial compressive strength, respectively. E and W UCS b is the mapping weight matrix; E and b UCS This is the bias term; the ReLU activation function is used to ensure that the output mechanical parameters are always non-negative.

[0013] Preferably, the preset physical information graph neural network is obtained through pre-training, and the training process of the physical information graph neural network includes: The skeleton network model of the input physical information graph neural network is subjected to random rotation in three-dimensional space and perturbation of node coordinates; Using the Adam optimizer, with the data error loss term L disabled. MSE Under the premise of only using the physical residual loss term L Phy Unsupervised warm-up training is performed; the data error loss term L MSE The physical residual loss term L is used to measure the mean square error between model predictions and actual laboratory measurements. Phy The calculation formula is based on the simplified nodal static equilibrium equations, as follows: (9); (10); Where, k eff This is the equivalent stiffness coefficient derived from mineral hardness; It characterizes the geometric stiffness between the skeleton nodes; Δu rel (uv) This refers to the relative displacement vectors between adjacent nodes inferred from the hidden layer of the model. After the preheating training, a small number of Micro-CT scan image sequences of reef limestone with real labels obtained from field boreholes were introduced, combined with the total loss function L. totalSupervised fine-tuning is performed; the total loss function L total Data error loss term L MSE and physical residual loss term L Phy Together they form the following components, and the calculation formula is as follows: (11); Where λ is a hyperparameter balancing the weights of data fitting and physical constraints.

[0014] Preferably, the method further includes: The predicted results of the macroscopic mechanical properties of the reef limestone are compared with the preset mechanical property classification thresholds of the reef limestone, and the engineering grade classification label of the reef limestone is output. The engineering grade classification label includes at least one of the following: extremely low strength, low strength, medium strength, and high strength.

[0015] The technical means adopted by this invention to solve its technical problem is: a device for predicting the mechanical properties of reef limestone, the device comprising: The data acquisition module is used to acquire and preprocess the Micro-CT scan image sequence of the reef limestone to obtain a binary three-dimensional digital core model of the reef limestone. The feature extraction module is used to construct the skeleton network model of the reef limestone core using the maximum sphere algorithm based on the binarized three-dimensional digital core model, and to extract the node features and edge attribute features of each node in the skeleton network model. The feature optimization module is used to input the node features and edge attribute features into a preset physical information graph neural network, and perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly fit the laws of rock mechanics. The result prediction module is used to aggregate the bias features of each node to obtain a full-map feature vector representing the macroscopic state of the reef limestone. The full-map feature vector is then input into two independent multilayer perceptron regression branches in a preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

[0016] The technical means adopted by the present invention to solve its technical problem is: an electronic device, comprising: at least one processor and at least one memory, wherein the memory stores program instructions or code; the program instructions or code are loaded and executed by the processor, thereby enabling the electronic device to implement the reef limestone mechanical property prediction method as described in any of the preceding claims.

[0017] The technical means adopted by the present invention to solve its technical problem is: a storage medium storing program instructions or code thereon, characterized in that the program instructions or code are loaded and executed by a processor to realize the method for predicting the mechanical properties of reef limestone as described in any of the preceding claims.

[0018] The beneficial effects of this invention are: Micro-CT scans were performed on cored reef limestone samples to obtain image sequences. After preprocessing, binary 3D digital cores were obtained. The maximum sphere algorithm was used to identify skeletal particles and construct nodes and edges, extracting edge attribute features. These features were fed into a physical information graph neural network (MLP), where normalization and physical bias attention calculations yielded mechanically perceptual bias features. Gated aggregation was then used to obtain full-image features, and a two-branch MLP was used to output the elastic modulus and uniaxial compressive strength, completing the entire prediction process from microstructure to macroscopic mechanical parameters. This process combines microstructure characterization with physical constraint deep learning, maintaining high prediction accuracy under small sample conditions while significantly reducing computation time. It achieves rapid, non-destructive, and high-precision prediction of the mechanical properties of reef limestone, effectively solving the problems of difficult sample preparation and high cost in traditional experiments, and the lack of physical interpretability in conventional AI. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for predicting the mechanical properties of reef limestone according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the construction and physical property mapping of the microstructure network of reef limestone in an embodiment of the present invention. Figure 3 This is a schematic diagram of the physical information graph neural network model design shown in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating multi-stage model training and macroscopic mechanical parameter prediction in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the construction of the loss function in an embodiment of the present invention; Figure 6 This is a flowchart illustrating a method for predicting the mechanical properties of reef limestone, as shown in an embodiment of the present invention. Figure 7 This is a structural block diagram of a reef limestone mechanical property prediction device according to an embodiment of the present invention; Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0022] like Figure 1 As shown, this invention provides a method for predicting the mechanical properties of reef limestone, the method comprising: S1. Obtain the Micro-CT scan image sequence of the reef limestone and preprocess it to obtain a binary three-dimensional digital core model of the reef limestone.

[0023] The purpose of this step is to transform real reef limestone physical samples into a computer-processable digital three-dimensional representation. Micro CT scanning is a non-destructive imaging technique that can acquire detailed images of the internal microstructure of rocks. By preprocessing the image sequence, the solid skeleton and pore space in the grayscale image are separated to form a binary model, where the solid skeleton region is represented by the value 1 and the pore region is represented by the value 0.

[0024] S2. Based on the binarized three-dimensional digital core model, the maximum sphere algorithm is used to construct the skeleton network model of the reef limestone core, and the node features and edge attribute features of each node in the skeleton network model are extracted.

[0025] The maximum sphere algorithm aims to find a series of maximum inscribed spheres within the solid skeleton space, where the center of each inscribed sphere represents the centroid of a skeleton particle. The connections between the inscribed spheres constitute the topology of the skeleton network. Node features are used to quantify the local geometric and physical properties of each skeleton particle; edge features are used to quantify the contact mechanics between adjacent skeleton particles. This step transforms the pixelated voxel model into a structured graph model, thereby reducing the complexity of subsequent calculations while preserving crucial mechanical transfer information.

[0026] S3. Input the node features and edge attribute features into the preset physical information graph neural network, and perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly fit the laws of rock mechanics.

[0027] Conventional Graph Attention Networks (GATs) rely solely on the mathematical similarity of node features to assign weights. In contrast, the mechanical message passing layer designed in this invention can simulate the "stress wave transmission and force chain evolution mechanism" inside real rocks, making the information flow law of the network highly consistent with the laws of rock mechanics.

[0028] S4. Aggregate the bias features of each node to obtain the full-map feature vector representing the macroscopic state of the reef limestone. Input the full-map feature vector into two independent multilayer perceptron regression branches in the preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

[0029] The purpose of information aggregation is to integrate microscopic node information into a macroscopic overall representation; the full-map feature vector condenses key information such as porosity, connectivity, and force chain distribution of the rock sample; two independent multilayer perceptron regression branches are used to predict the elastic modulus and uniaxial compressive strength, respectively. Independent regression branches allow each mechanical parameter to have independent mapping weights, thus adapting to different mapping relationships between different mechanical parameters and microstructures.

[0030] The technical advantage of this invention lies in establishing a direct and interpretable mapping relationship between the microstructure of rocks and their macroscopic mechanical response. The maximum ball algorithm transforms massive CT images into lightweight graph structures, significantly reducing computational dimensionality. The physical definition of node features and edge attributes allows the subsequent neural network to fully utilize prior mechanical knowledge rather than purely relying on statistical correlations. This effectively overcomes the shortcomings of existing technologies where data-driven approaches are disconnected from physical laws, and solves the technical problems of low prediction accuracy and poor robustness of mechanical parameters in complex porous media under small sample conditions.

[0031] In a preferred embodiment, such as Figure 2 As shown, S1 includes: Image sequences were obtained by non-destructive scanning of reef limestone using industrial Micro-CT equipment; After median filtering for noise reduction and three-dimensional thresholding of the image sequence, a binarized three-dimensional digital core model of the reef limestone was obtained.

[0032] Among them, the industrial Micro CT equipment has micron-level resolution, enabling clear differentiation of biopores and microfracture structures within reef limestone. Non-destructive scanning means that the rock sample does not need to be cut or damaged, thus protecting valuable island and reef core samples. During the scanning process, the rock sample is placed on a rotating stage, and X-rays emitted by the X-ray source penetrate the sample and are received by the detector. A three-dimensional grayscale image sequence is reconstructed using projection data from different angles. The grayscale value of each voxel in this sequence represents the degree of X-ray absorption at that location, reflecting the density of the material.

[0033] Median filtering is a nonlinear filtering method that uses the median gray value in the neighborhood of each voxel as the new value for that voxel. This method can effectively remove salt-and-pepper noise while preserving edge details. 3D thresholding requires determining a gray-level threshold; voxels above this threshold are labeled as solid skeletons, and those below are labeled as pores. The threshold can be automatically determined using the Otsu's method or manually set based on the gray-level distribution characteristics of solids and pores in the rock. The result of the binarization model is a 3D matrix containing only 0s and 1s, facilitating subsequent skeleton extraction and numerical calculations.

[0034] This embodiment preserves the integrity of the precious rock core through non-destructive scanning technology, allowing the same core to be used for both CT scans and subsequent verification experiments. The combination of median filtering and threshold segmentation is computationally efficient and suitable for processing large-scale CT image sequences; the binarization model simplifies complex grayscale information into binary distinctions of the skeleton and pores, providing a clear foundation for subsequent skeleton network extraction. These steps break away from traditional pixel-grid-based image representation methods, transforming unstructured reef limestone 3D CT scan images into structured mathematical graph models that accurately describe the internal mechanical transmission paths.

[0035] In a preferred embodiment, see further. Figure 2 S2 includes: S201. The maximum sphere algorithm is used to find several inscribed spheres in the binary three-dimensional digital core model. Based on the centers of the inscribed spheres and their connection relationships, a skeleton network model is constructed.

[0036] The basic principle of the maximum sphere algorithm is to attempt to expand outwards within the solid skeleton space, using each skeleton voxel as its center, until it reaches the pore boundary. Each expanded sphere is called an inscribed sphere. Among all inscribed spheres, the sphere with the largest radius is selected as the principal sphere for that local region, and the center of the principal sphere represents the centroid of the skeleton particle. When two principal spheres overlap, it indicates that the corresponding skeleton particles are spatially adjacent, thus establishing a connection. This connection constitutes the topology of the skeleton network.

[0037] S202. Define the centroid of the skeleton particles in the skeleton network model as a node V in graph theory. The feature definition formula of node V is as follows: (1); To fully reflect the local mechanical load-bearing capacity of the skeleton network model, the feature vectors of the nodes not only contain geometric information, but also incorporate the microscopic physical properties of the material.

[0038] Among them, X v R represents the feature vector of node V; vV represents the equivalent inscribed sphere radius of the skeletal particle, which visually reflects the size of the skeletal particle; v H represents the actual volume of the skeletal particles, reflecting their geometric size; m This represents the local Vickers hardness value of the mineral composition obtained through nanoindentation testing or mineral composition inversion, and represents the intrinsic stiffness of the material; C n The coordinate number of node V is represented by , and the number of neighboring nodes connected to this node is represented by , reflecting the local topological stability; T represents the transpose of the matrix.

[0039] S203. The physical contact surface between adjacent skeletal particles is detected using a three-dimensional morphological dilation algorithm. The physical contact surface is defined as the edge attribute E in graph theory. The edge attribute directly determines the stress transfer efficiency between particles. The characteristic definition formula of the edge attribute E is as follows: (2); Among them, E uv A represents the feature vector of the edge connecting nodes U and V; uv This represents the effective contact area between two skeletal particles, which determines the magnitude of force transmission flux; the larger the contact area, the higher the force transmission flux. uv It represents the Euclidean distance between the centroids of two nodes, reflecting the interparticle spacing; ξ represents the cosine of the angle between the contact surface normal direction and the macroscopic principal stress loading axis (usually set as the Z-axis), characterizing the degree of matching between the contact surface direction and the principal stress direction; rough It represents the geometric curvature of the contact surface, used to characterize the roughness of the contact interface and the interlocking friction effect.

[0040] This embodiment transforms the complex porous structure into a concise graphical model, greatly reducing the data dimensionality and completing the initial mapping from "pure geometric image features" to "physical and mechanical properties." The maximum ball algorithm can effectively handle highly heterogeneous porous materials such as reef limestone. The feature definitions of nodes and edges closely align with the basic variables of rock mechanics, enabling the subsequent graph neural network to directly process input features with clear physical meaning.

[0041] In a preferred embodiment, such as Figure 3 As shown, S3 includes: S301. Input the node features and edge attribute features into the multilayer perceptron for normalization processing.

[0042] A multilayer perceptron (MLP) is a feedforward neural network architecture consisting of an input layer, hidden layers, and an output layer. Normalization aims to eliminate differences in the dimensions and numerical ranges of different physical quantities, thereby accelerating network convergence and improving training stability. For example, the numerical range of an equivalent radius is in the micrometer range, while the numerical range of Vickers hardness is in the megapascal range; the two differ significantly, and direct input would lead to unbalanced gradient updates. The MLP learns a set of trainable scaling and translation factors to map each feature dimension to a standard normal distribution space with a mean of 0 and a variance of 1.

[0043] S302. Input the normalized features into a graph attention network that embeds a specific physical bias function, and calculate the bias features of node V that are highly consistent with the laws of rock mechanics based on the attention coefficients of the graph attention network.

[0044] When rock is under pressure, stress is not uniformly distributed, but preferentially transmitted along the skeletal path with high stiffness, large contact area, and parallel to the principal stress direction, forming a so-called "strong force chain". To simulate this physical phenomenon, this invention embeds a specific physical bias function into the graph attention mechanism.

[0045] First, the formula for calculating the attention coefficient is as follows: (3); Where, α uv h represents the influence weight of node U on node V. u ,h v Let U be the high-dimensional latent feature vectors of nodes U and V; W is the learnable linear transformation weight matrix; || denotes the vector concatenation operation; a T is the learnable parameter vector for the attention mechanism; N(u) represents the set of all first-order neighbor nodes of node U; LeakyReLU is a modified linear unit with leakage, which is an improved version of the classic ReLU function; Among them, the physical bias function Ψ(E) uv The mathematical expression for ) is: (4); Where β1 and β2 are the physical scaling factors that the physical information graph neural network adaptively learns during the training process.

[0046] The bias term Ψ(E) uv The physical significance of this is that it explicitly increases the weight of edges with "large contact area (large Auv)," "short distance (small Duv)," and "aligned along the principal stress direction (|θz| approaches 1)." Through this forced intervention, the neural network automatically favors those mechanically "strong connections" when aggregating neighbor node information, perfectly simulating the real transmission law of stress flow between rock skeletons.

[0047] Based on the attention coefficient α uv The offset characteristic of node V height conforming to the laws of rock mechanics was calculated, and the calculation formula is as follows: (5); Where σ is a nonlinear activation function, h v ” The offset characteristic of node V is that it conforms to the laws of rock mechanics.

[0048] This embodiment sets a physical bias function to force the neural network to focus on mechanically important connections. This mechanism enables the network to correctly identify strong force chain networks within rocks even when training data is scarce. The learnable parameters of the physical bias function allow the network to automatically adapt to the mechanical response characteristics of different types of reef limestone; the attention mechanism also provides interpretability for the model, visually displaying stress concentration areas and potential failure paths.

[0049] In a preferred embodiment, see further. Figure 3 S4 includes: S401. Aggregate the information from the bias features of several nodes to obtain the full-map feature vector representing the entire reef limestone. The calculation formula is as follows: (6); Where Gate() represents a learnable gated network; ⨀ represents the Hadamard product; h Graph σ represents the final generated full-image feature vector; σ is the non-linear activation function.

[0050] The information aggregation function uses a gated summation form, and the bias feature h of each node V... v ” First, a gating value between 0 and 1 is calculated using a gating network Gate(h). This gating value characterizes the contribution of node V to the macroscopic mechanical response; the gating value is close to 1 for strong chain nodes bearing the main load, and close to 0 for suspended or almost unloaded weak skeleton nodes. The gating network is a small multilayer perceptron whose output is compressed to between 0 and 1 using a sigmoid function. The full-image feature vector h... Graph It is obtained by multiplying the hidden state of each node with its corresponding gating value element by element (i.e., the Hadamard product) and summing the results over all nodes. It highly condenses all the key microstructural information of the entire rock sample, such as porosity, connectivity, anisotropy, and force chain distribution.

[0051] S402. Input the full-image feature vector into two independent multilayer perceptron regression branches, and output the elastic modulus prediction result and the uniaxial compressive strength prediction result respectively. The calculation formula is as follows: (7); (8); Among them, E pred and UCS pred W represents the predicted macroscopic elastic modulus and uniaxial compressive strength, respectively. E and W UCS b is the mapping weight matrix; E and b UCS This is the bias term; the ReLU activation function is used to ensure that the output mechanical parameters are always non-negative.

[0052] This embodiment intelligently identifies and focuses on the force chain network within the rock through a gated aggregation mechanism, while ignoring suspended skeletons that contribute little to the macroscopic mechanical response. This selective aggregation significantly improves the signal-to-noise ratio and accuracy of the prediction. The bi-branch independent regression design allows the elastic modulus and compressive strength to have their own independent feature-to-parameter mapping relationships, fully considering the different dependency patterns between different mechanical parameters and microstructure.

[0053] In a preferred embodiment, the preset physical information graph neural network is obtained through pre-training, such as... Figure 4 As shown, the training process of the physical information graph neural network includes: The skeleton network model of the input physical information graph neural network is subjected to random rotation in three-dimensional space and perturbation of node coordinates.

[0054] The purpose of random rotation is to enhance the rotation invariance of the model, ensuring that the model provides the same prediction results even if the rock samples are placed in different orientations during scanning. The rotation operation randomly generates rotation angles around the three coordinate axes and then applies the rotation matrix to the coordinates of all nodes. Uniformly distributed sampling of rotation angles can be used, where each angle is randomly selected with equal probability between 0 and 360°; alternatively, incremental sampling of specific angles can be used, rotating only in integer multiples of 90° to maintain the alignment properties in voxel space.

[0055] The perturbation of node coordinates involves adding a small random noise vector to the coordinates of each node. The purpose of the perturbation is to simulate the natural positional fluctuations of skeleton particles in real rock cores and the unavoidable boundary errors during CT image segmentation. The magnitude of the perturbation is typically set to within half the voxel size to ensure that the topology of the skeleton network is not altered.

[0056] Using the Adam optimizer, with the data error loss term L disabled. MSE Under the premise of only using the physical residual loss term L Phy Conduct unsupervised warm-up training.

[0057] The Adam optimizer is an adaptive learning rate gradient descent algorithm that automatically adjusts the learning rate of each parameter based on its gradient history. Disabling the data error term during the warm-up training phase means the network doesn't need to fit any real label data temporarily, allowing the network parameters to quickly converge to an initial state that satisfies the laws of mechanical equilibrium.

[0058] like Figure 5 As shown, the data error loss term L MSE The physical residual loss term L is used to measure the mean square error between model predictions and actual laboratory measurements. Phy The calculation formula is based on the simplified nodal static equilibrium equations, as follows: (9); (10); Where, k eff This is the equivalent stiffness coefficient derived from mineral hardness; It characterizes the geometric stiffness between the skeleton nodes; Δu rel (uv) This is the relative displacement vector between adjacent nodes inferred from the hidden layer of the model.

[0059] Physical residual loss term L Phy The calculation formula is based on the summation of the squares of the residual forces of all nodes by the simplified nodal static equilibrium equation. The core idea of ​​this equation is that for each skeleton node, the resultant force exerted on it by all its surrounding neighboring nodes should be close to zero.

[0060] The introduction of this physical residual loss term is equivalent to constructing a virtual "spring-mass" mechanical system inside the neural network, forcing the feature distribution learned by the neural network to conform to the most basic mechanical equilibrium law, thereby transforming the pure mathematical black-box fitting into gray-box inference constrained by physical laws.

[0061] After the preheating training, a small number of Micro-CT scan image sequences of reef limestone with real labels obtained from field boreholes were introduced, combined with the total loss function L. total Conduct supervised fine-tuning.

[0062] Total loss function L total Data error loss term L MSE and physical residual loss term L Phy Together they form the following components, and the calculation formula is as follows: (11); Where λ is a hyperparameter balancing the weights of data fitting and physical constraints.

[0063] When λ is large, the weight of physical constraints is greater, and the network tends to output predictions that better conform to the laws of mechanics; when λ is small, the weight of data fitting is greater, and the network tends to fit the training samples more accurately. The fine-tuning stage uses a small amount of real-label data to adjust the network parameters so that it not only satisfies physical laws but also has the ability to accurately predict specific lithologies.

[0064] This embodiment ensures that the network possesses basic mechanical knowledge before encountering real labels, thus avoiding getting trapped in physically unreasonable local optima under small sample conditions. The physical residual loss term L... phy As a soft constraint, it acts continuously throughout the training process, effectively suppressing overfitting through regularization. The two-stage training strategy fully utilizes the value of a small number of true labels, enabling the model to both conform to the laws of mechanics and possess accurate predictive capabilities for specific reef limestone. The adjustability of the λ hyperparameter allows engineers to flexibly adjust the weights of the two losses based on data quality and the reliability of physical priors.

[0065] In one specific embodiment: During the pile foundation design phase of a large breakwater on a certain island reef, the survey team drilled a typical high-porosity reef limestone core with a diameter of 50mm and a height of 100mm at a depth of 40 meters underwater. Due to the extreme fragility of the core, a standard uniaxial compression test could not be performed directly. The engineering team used a portable industrial Micro-CT device on-site to perform a non-destructive scan of the core, setting the scan resolution to 5μm, and obtained a high-precision three-dimensional grayscale image sequence containing approximately 2000×2000×4000 voxels. The CT images were imported into a workstation equipped with the algorithm of this invention. The system automatically ran the maximum sphere algorithm, extracting approximately 3.2 million skeleton nodes and 8.5 million contact edges from the massive number of voxels in just 5 minutes, successfully constructing the microscopic skeleton map of the core, and calculating the equivalent radius, contact area, and other physical properties of each node. The constructed map data was then input into a PI-GNN model that had already completed "pre-training" in the cloud. The model's internal mechanical message passing layer, based on a preset physical bias, quickly identifies the strong force chain network within the rock core along the vertical direction (simulating the axial force of the pile foundation) and uses a gated network to filter out approximately 15% of the suspended invalid pore skeleton. After forward inference, the model outputs the predicted result within 1.2 seconds: the macroscopic elastic modulus E of the reef limestone sample. pred =15.2GPa, uniaxial compressive strength UCS pred =24.8 MPa.

[0066] In a preferred embodiment, the method further includes: The predicted results of the macroscopic mechanical properties of reef limestone are compared with the preset mechanical property classification thresholds of reef limestone, and the engineering grade classification label of reef limestone is output. The engineering grade classification label includes at least one of the following: extremely low strength, low strength, medium strength, and high strength.

[0067] The preset grading thresholds are numerical ranges pre-defined based on engineering practice experience and relevant design specifications. The grading labels include four types: extremely low strength, low strength, medium strength, and high strength. Each strength grade corresponds to a range of elastic modulus and a range of uniaxial compressive strength. The comparison process involves comparing the model-predicted elastic modulus and uniaxial compressive strength values ​​with these thresholds one by one to determine the grade of the rock sample. The purpose of outputting engineering grade classification labels is to provide an intuitive reference for engineering design, facilitating designers to quickly determine the foundation bearing capacity and select appropriate construction parameters. This classification result can be directly applied to engineering practices such as calculating the rock embedment depth of pile foundations, rating the difficulty of waterway dredging, and assessing slope stability.

[0068] Specifically, the output results of the model after prediction in the above embodiment are an elastic modulus of 15.2 GPa and a uniaxial compressive strength of 24.8 MPa.

[0069] The system's internally preset tiered threshold rules are as follows: Extremely low strength corresponds to an elastic modulus less than 5 GPa and a uniaxial compressive strength less than 10 MPa; low strength corresponds to an elastic modulus between 5 and 15 GPa or a uniaxial compressive strength between 10 and 20 MPa; medium strength corresponds to an elastic modulus between 15 and 30 GPa or a uniaxial compressive strength between 20 and 40 MPa; high strength corresponds to an elastic modulus greater than 30 GPa and a uniaxial compressive strength greater than 40 MPa.

[0070] Comparing the predicted values ​​of 15.2 GPa and 24.8 MPa with the aforementioned thresholds reveals that the elastic modulus of 15.2 GPa falls just below the lower limit of medium strength, while the uniaxial compressive strength of 24.8 MPa falls within the medium strength range. Based on this comprehensive assessment, the system classifies the reef limestone sample as medium strength. Therefore, the system outputs an engineering grade classification label of medium strength and recommends that engineering designers perform pile foundation bearing capacity calculations and foundation design according to the relevant design parameters for medium-strength rock.

[0071] This embodiment transforms continuous mechanical parameter predictions into discrete engineering grade classifications, enabling even non-experts to quickly understand the quality status of rocks. The classification results are highly compatible with existing engineering standards and can be directly integrated into existing design processes without additional conversion. This classification output format is particularly suitable for rapid on-site decision-making scenarios. For example, on offshore construction platforms, construction workers can directly determine whether an area requires special reinforcement measures based on color-coded labels.

[0072] like Figure 6 The flowchart shown is a flowchart of the reef limestone mechanical property prediction method provided by the present invention. By converting the microscopic image of the rock into a skeleton map, and innovatively introducing a physical information graph neural network (PI-GNN) that integrates static equilibrium and microscopic force transmission mechanism, it effectively overcomes the defect of data-driven and physical law separation in the prior art, solves the technical problems of low prediction accuracy and poor robustness of mechanical parameters of complex porous media under small sample conditions, and realizes efficient, accurate and non-destructive evaluation of macroscopic mechanical properties based on microstructure.

[0073] The following are embodiments of the apparatus of the present invention, which can be used to execute the method for predicting the mechanical properties of reef limestone involved in the present invention. For details not disclosed in the embodiments of the apparatus of the present invention, please refer to the method embodiments of the method for predicting the mechanical properties of reef limestone involved in the present invention.

[0074] Please see Figure 7 This invention provides a device 50 for predicting the mechanical properties of reef limestone, including but not limited to: a data acquisition module 510, a feature extraction module 520, a feature optimization module 530, and a result prediction module 540, wherein: The data acquisition module 510 is used to acquire and preprocess the Micro-CT scan image sequence of the reef limestone to obtain a binary three-dimensional digital core model of the reef limestone. The feature extraction module 520 is used to construct the skeleton network model of the reef limestone core based on the binarized three-dimensional digital core model using the maximum sphere algorithm, and to extract the node features and edge attribute features of each node in the skeleton network model. The feature optimization module 530 is used to input node features and edge attribute features into a preset physical information graph neural network, and to perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly fit the laws of rock mechanics. The result prediction module 540 is used to aggregate the bias features of each node to obtain a full-map feature vector representing the macroscopic state of the reef limestone. The full-map feature vector is then input into two independent multilayer perceptron regression branches in the preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

[0075] It should be noted that the above embodiments of the reef limestone mechanical property prediction device are only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the reef limestone mechanical property prediction device will be divided into different functional modules to complete all or part of the functions described above.

[0076] Furthermore, the embodiments of the reef limestone mechanical property prediction device and the reef limestone mechanical property prediction method provided in the above embodiments belong to the same concept. The specific way each module performs its operation has been described in detail in the method embodiments, and will not be repeated here.

[0077] Please see Figure 8 This invention provides an electronic device 4000, which includes at least one processor 4001 and at least one memory 4003.

[0078] The data interaction between the processor 4001 and the memory 4003 can be achieved through at least one communication bus 4002. This communication bus 4002 may include a path for transmitting data between the processor 4001 and the memory 4003. The communication bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0079] Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0080] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0081] The memory 4003 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (CoMPact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program instructions or code in the form of instructions or data structures and accessible by the electronic device 4000, but not limited thereto.

[0082] The memory 4003 stores program instructions or code, and the processor 4001 can read the program instructions or code stored in the memory 4003 through the communication bus 4002.

[0083] When the program instructions or code are executed by the processor 4001, the method for predicting the mechanical properties of reef limestone in the above embodiments is implemented.

[0084] Furthermore, this embodiment of the invention provides a storage medium storing program instructions or code, which is loaded and executed by a processor to realize the reef limestone mechanical property prediction method as described above.

[0085] This invention provides a computer program product, which includes program instructions or code stored in a storage medium. The processor of an electronic device reads the program instructions or code from the storage medium, loads and executes the program instructions or code, thereby enabling the electronic device to implement the method for predicting the mechanical properties of reef limestone as described above.

[0086] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for predicting the mechanical properties of reef limestone, characterized in that, The method includes: S1. Obtain the Micro-CT scan image sequence of the reef limestone and preprocess it to obtain the binary three-dimensional digital core model of the reef limestone. S2. Based on the binarized three-dimensional digital core model, the maximum sphere algorithm is used to construct the skeleton network model of the reef limestone core, and the node features and edge attribute features of each node in the skeleton network model are extracted. S3. Input the node features and edge attribute features into a preset physical information graph neural network, and perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly conform to the laws of rock mechanics. S4. Aggregate the bias features of each node to obtain a full-map feature vector representing the macroscopic state of the reef limestone. Input the full-map feature vector into two independent multilayer perceptron regression branches in a preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

2. The method for predicting the mechanical properties of reef limestone according to claim 1, characterized in that, S1 includes: The reef limestone was non-destructively scanned using an industrial Micro-CT scanner to obtain an image sequence. After performing median filtering for noise reduction and three-dimensional threshold segmentation on the image sequence, a binarized three-dimensional digital core model of the reef limestone is obtained.

3. The method for predicting the mechanical properties of reef limestone according to claim 1, characterized in that, S2 includes: The maximum sphere algorithm is used to find several inscribed spheres in the binary three-dimensional digital core model. Based on the centers of the inscribed spheres and their connection relationships, a skeleton network model is constructed. The centroid of the skeleton particles in the skeleton network model is defined as node V in graph theory, and the feature definition formula of node V is as follows: （1）； wherein X v represents the eigenvector of node V; R v represents the equivalent inscribed sphere radius of the skeleton particle; V v represents the actual volume of the skeleton particle; H m represents the local mineral component Vickers hardness value obtained by nanoindentation test or mineral component inversion; C n represents the coordination number of the node V; T represents the transpose calculation of the matrix; The physical contact surface between adjacent skeletal particles is detected using a three-dimensional morphological dilation algorithm. This physical contact surface is defined as an edge attribute E in graph theory, and the characteristic definition formula for edge attribute E is as follows: （2）； Among them, E uv A represents the feature vector of the edge connecting nodes U and V; uv D represents the effective contact area between two skeletal particles; uv This represents the Euclidean distance between the centroids of the two nodes; ξ represents the cosine of the angle between the normal direction of the contact surface and the macroscopic principal stress loading axis; rough It represents the geometric curvature of the contact surface, used to characterize the roughness of the contact interface and the interlocking friction effect.

4. The method for predicting the mechanical properties of reef limestone according to claim 3, characterized in that, S3 includes: The node features and edge attribute features are input into a multilayer perceptron for normalization processing. The normalized features are input into a graph attention network embedded with a specific physical bias function, and the bias features of node V that are highly consistent with the laws of rock mechanics are calculated based on the attention coefficients of the graph attention network. The formula for calculating the attention coefficient is as follows: （3）； Where, α uv h represents the influence weight of node U on node V. u ,h v Let U be the high-dimensional latent feature vectors of nodes U and V; W is the learnable linear transformation weight matrix; || denotes the vector concatenation operation; a T is the learnable parameter vector for the attention mechanism; N(u) represents the set of all first-order neighbor nodes of node U; LeakyReLU is a modified linear unit with leakage, which is an improved version of the classic ReLU function; Wherein, the physical bias function Ψ(E) uv The mathematical expression for ) is: （4）； Wherein, β1 and β2 are the physical scaling factors that the physical information graph neural network adaptively learns during the training process; According to the attention coefficient α uv The offset characteristic of node V height conforming to the laws of rock mechanics was calculated, and the calculation formula is as follows: （5）； Where σ is a nonlinear activation function, h v ” The offset characteristic of node V is that it conforms to the laws of rock mechanics.

5. The method for predicting the mechanical properties of reef limestone according to claim 4, characterized in that, S4 includes: The bias features of several nodes are aggregated to obtain the full-map feature vector representing the entire reef limestone mass. The calculation formula is as follows: （6）； Among them, h u Gate(h) represents the final hidden state of node U after deep network computation; u ) represents a learnable gating network; ⨀ represents the Hadamard product; h Graph This is the final generated full-image feature vector; The full-image feature vector is input into two independent multilayer perceptron regression branches, which output the elastic modulus prediction result and the uniaxial compressive strength prediction result, respectively. The calculation formula is as follows: （7）； （8）； Among them, E pred and UCS pred W represents the predicted macroscopic elastic modulus and uniaxial compressive strength, respectively. E and W UCS b is the mapping weight matrix; E and b UCS This is the bias term; the ReLU activation function is used to ensure that the output mechanical parameters are always non-negative.

6. The method for predicting the mechanical properties of reef limestone according to claim 5, characterized in that, The preset physical information graph neural network is obtained through pre-training, and the training process of the physical information graph neural network includes: The skeleton network model of the input physical information graph neural network is subjected to random rotation in three-dimensional space and perturbation of node coordinates; Using the Adam optimizer, with the data error loss term L disabled. MSE Under the premise of only using the physical residual loss term L Phy Unsupervised warm-up training is performed; the data error loss term L MSE The physical residual loss term L is used to measure the mean square error between model predictions and actual laboratory measurements. Phy The calculation formula is based on the simplified nodal static equilibrium equations, as follows: （9）； （10）； Where, k eff This is the equivalent stiffness coefficient derived from mineral hardness; It characterizes the geometric stiffness between the skeleton nodes; Δu rel (uv) This refers to the relative displacement vectors between adjacent nodes inferred from the hidden layer of the model. After the preheating training, a small number of Micro-CT scan image sequences of reef limestone with real labels obtained from field boreholes were introduced, combined with the total loss function L. total Supervised fine-tuning is performed; the total loss function L total Data error loss term L MSE and physical residual loss term L Phy Together they form the following components, and the calculation formula is as follows: （11）； Where λ is a hyperparameter balancing the weights of data fitting and physical constraints.

7. The method for predicting the mechanical properties of reef limestone according to claim 1, characterized in that, The method further includes: The predicted results of the macroscopic mechanical properties of the reef limestone are compared with the preset mechanical property classification thresholds of the reef limestone, and the engineering grade classification label of the reef limestone is output. The engineering grade classification label includes at least one of the following: extremely low strength, low strength, medium strength, and high strength.

8. A device for predicting the mechanical properties of reef limestone, characterized in that, The device includes: The data acquisition module is used to acquire and preprocess the Micro-CT scan image sequence of the reef limestone to obtain a binary three-dimensional digital core model of the reef limestone. The feature extraction module is used to construct the skeleton network model of the reef limestone core using the maximum sphere algorithm based on the binarized three-dimensional digital core model, and to extract the node features and edge attribute features of each node in the skeleton network model. The feature optimization module is used to input the node features and edge attribute features into a preset physical information graph neural network, and perform feature embedding and high-dimensional mapping according to the mechanical message passing layer of the physical information graph neural network to obtain bias features that highly fit the laws of rock mechanics. The result prediction module is used to aggregate the bias features of each node to obtain a full-map feature vector representing the macroscopic state of the reef limestone. The full-map feature vector is then input into two independent multilayer perceptron regression branches in a preset physical information graph neural network to obtain the prediction results of the macroscopic mechanical properties of the reef limestone.

9. An electronic device, characterized in that, include: At least one processor and at least one memory, wherein program instructions or code are stored in the memory; the program instructions or code are loaded and executed by the processor, causing the electronic device to implement the method for predicting the mechanical properties of reef limestone as described in any one of claims 1 to 7.

10. A storage medium storing program instructions or code thereon, characterized in that, The program instructions or code are loaded and executed by the processor to implement the method for predicting the mechanical properties of reef limestone as described in any one of claims 1 to 7.

Images