Artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis

Abstract

The invention provides an artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis, and the method combines multi-scale morphological feature analysis including surface texture and internal pore structure, and utilizes a deep learning algorithm to accurately predict concrete interface strength. The method comprises the following steps: simulating a pore filling rate and microcrack propagation resistance of an interface area by combining multi-scale concrete morphology information to obtain performance parameters of an interface transition area; calculating the deviation between the predicted strength and the actual strength to obtain a prediction precision index, and if the prediction precision index does not reach a preset threshold value, updating the strength prediction model parameters to obtain an updated strength prediction model; and calculating a strength prediction result according to the quantitative morphology feature vectors, the microcrack propagation path parameters and the cement paste hydration degree of various concrete material systems to obtain strength prediction precision. The method not only improves the accuracy of traditional strength evaluation, but also provides a scientific basis for optimizing the concrete proportion design.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis. Background Art

[0002] Concrete strength assessment is a crucial research area in civil engineering, directly impacting the safety and durability of building structures. With the introduction of artificial intelligence (AI) technology, data-driven strength prediction models have become a research hotspot, offering a new path for efficient and accurate material performance assessment. However, existing AI-based assessment methods have significant limitations when dealing with complex raw material properties. Traditional models typically rely on macroscopic parameters such as water-cement ratio and aggregate particle size, while neglecting the influence of microscopic features such as sand particle morphology. This results in significant fluctuations in prediction accuracy across different material systems, particularly in the production of sand-cement concrete. This limitation makes it difficult for models to adapt to the diverse raw material properties, limiting their practical application. A key challenge in predicting the strength of sand-cement concrete lies in accurately quantifying the impact of sand particle morphology on strength. Sand particle angularity, surface roughness, and porosity distribution are key factors in determining interfacial bond strength, microcrack development, and the performance of the aggregate-to-interface transition zone. However, existing methods often fail to effectively integrate these microscopic parameters, for example, failing to accurately capture microcrack development patterns and inaccurately quantifying the strength contribution of the interfacial transition zone. These technical difficulties make it difficult for the strength assessment system to adaptively adjust according to the morphological characteristics of the raw materials, which in turn affects the reliability and engineering applicability of the prediction results. Summary of the Invention

[0003] The present invention provides an artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis, which mainly includes:

[0004] Acquire the morphological data of concrete sand particles, extract the sharpness of sand particles’ edges and corners, surface texture complexity and pore space connectivity, and obtain the sand particle morphological feature vector;

[0005] The interfacial contact behavior between sand and cement slurry is simulated. Based on the sharpness of the edges and corners and the complexity of the surface texture in the sand morphology feature vector, combined with the cement slurry adhesion and interfacial stress gradient, the stress distribution in the interfacial area is calculated to obtain the interfacial bond strength parameter.

[0006] If the interface bonding strength parameter is lower than the preset threshold, the initial starting point of the microcrack is analyzed. Based on the pore space connectivity and the initial width of the microcrack, the crack distribution density and the initial stress concentration at the crack starting point are calculated to obtain the initial characteristic parameters of the microcrack.

[0007] The fusion interface bonding strength parameters, microcrack initial characteristic parameters, interface transition zone porosity and pore filling material distribution are simulated to simulate the pore filling rate and microcrack propagation resistance of the interface area and obtain the interface transition zone performance parameters;

[0008] Calculate the stress distribution when microcracks propagate at the mortar interface, analyze the microcrack propagation behavior, and combine the crack propagation path curvature and microcrack propagation resistance to calculate the geometric characteristics of the crack propagation path and obtain the microcrack propagation path parameters;

[0009] By integrating the interface transition zone performance parameters and microcrack propagation path parameters, and adaptively training the sand surface chemical activity and cement slurry hydration degree, a strength prediction model is obtained, which outputs the predicted value of the mortar interface strength.

[0010] By calculating the deviation between the predicted intensity and the actual intensity, a prediction accuracy index is obtained. If the prediction accuracy index does not reach a preset threshold, the intensity prediction model parameters are updated to obtain an updated intensity prediction model.

[0011] The updated strength prediction model was verified, and the strength prediction results were calculated for the quantitative morphological feature vectors, microcrack propagation path parameters and cement paste hydration degree of various concrete material systems to obtain the strength prediction accuracy.

[0012] Furthermore, the morphological data of concrete sand grains is obtained, and the sand grain angular sharpness, surface texture complexity, and pore space connectivity are extracted to obtain a sand grain morphological feature vector. This includes: obtaining a grayscale image of the sand grain surface using a two-dimensional sand grain image acquisition device, extracting sand grain contour segments using a Sobel operator, calculating the curvature of the contour segments to obtain a first curvature value array, and segmenting the contour segments according to a preset curvature threshold to obtain initial sand grain angular region data. A three-dimensional point cloud data acquisition device is used to scan the sand grains to obtain raw point cloud data, and a Gaussian filter is used to remove noise points to obtain filtered point cloud data. The filtered point cloud data is then registered with the initial sand grain angular region data to obtain a sand grain angular region point cloud. A Fourier transform is performed on the sand grain angular region point cloud to obtain a frequency domain energy distribution matrix. A frequency domain energy threshold is used to segment the surface texture region point cloud to obtain a set of surface texture region point clouds. A discrete wavelet transform is performed on the surface texture region point cloud set to obtain a multi-scale texture feature vector. The point cloud of the sand grain's angular regions was voxelized, and a region growing algorithm was used to obtain the boundary point set of the pore region. The Poisson reconstruction algorithm was then used to triangulate the boundary point set to create a pore space mesh model. A connected domain labeling matrix was calculated based on the pore space mesh model. A breadth-first search algorithm was used to count the number of connected channels to obtain a connectivity index. Morphological analysis of the connected channels was performed to obtain channel diameter distribution data. Local curvature radius and angle data were extracted from the point cloud of the sand grain's angular regions. A density clustering algorithm was used to calculate the angular sharpness eigenvalue. The sand grain morphology feature vector was constructed by combining the multi-scale texture feature vector and the connectivity index.

[0013] Furthermore, the interfacial contact behavior between sand and cement slurry is simulated. Based on the angular sharpness and surface texture complexity in the sand morphology feature vector, combined with the cement slurry adhesion and interfacial stress gradient, the stress distribution in the interface region is calculated to obtain the interfacial bond strength parameter. This involves constructing a first sand surface mesh based on the angular sharpness data in the sand morphology feature vector, subdividing the sand and cement slurry contact region into a second subdivided mesh using a tetrahedral mesher, and calculating the stress distribution matrix at the interface mesh nodes using a contact stress calculator based on the principle of virtual work. Wetting contact angle values are calculated based on the sand surface texture complexity parameter. The adhesion matrix of the cement slurry to the sand surface is obtained using a surface energy-based interfacial tension calculator. Combined with the stress distribution matrix at the interface mesh nodes, the first interfacial force distribution matrix is obtained. Based on the first interfacial force distribution matrix, the stress field distribution in the interface region is calculated using a Newton-Raphson iterative solver. Finite strain analysis is performed on the stress field to obtain the interface deformation matrix, and the first interfacial bond coefficient is calculated from the interface deformation matrix. A stress-strain constitutive relationship was established for the first interface cohesion coefficient. An explicit stress transfer calculator was used to calculate the internal stress state matrix of the interface region, obtaining the interface shear stress and compressive stress components. The principal stress values were calculated based on the interface shear stress and compressive stress components. The interface stress transfer rate matrix was obtained using the Mohr-Coulomb criterion, and the interface bond strength parameters were obtained by combining the first interface cohesion coefficient.

[0014] Furthermore, if the interface bonding strength parameter is lower than a preset threshold, the initial microcrack starting point is analyzed. Based on the pore space connectivity and the initial microcrack width, the crack distribution density and initial crack stress concentration at the crack starting point are calculated to obtain the initial microcrack characteristic parameters. This includes: numerical comparison of the interface bonding strength parameter with the preset threshold. If the interface bonding strength parameter is lower than the preset threshold, a molecular dynamics calculator is used to calculate the interatomic potential energy distribution at the interface. A stress singular value calculator is used to locate high stress regions at the interface to obtain a stress singular distribution matrix. A three-dimensional pore network is constructed based on the pore space connectivity data. A fracture stress calculator is used to identify stress peak points in the stress singular distribution matrix. A mesh refinement processor is used to optimize the mesh in the stress peak region to obtain the initial crack stress field distribution. The crack initiation location is determined based on the initial crack stress field distribution. A stress gradient calculator is used to analyze the stress gradient at the crack initiation location. The initial crack width value is obtained using a crack boundary identifier to obtain the crack geometry matrix. Morphological features are extracted from the crack geometry matrix. The local crack distribution density is calculated using a fracture morphology analyzer. The crack density distribution function is obtained using a statistical analyzer to obtain the first crack density characteristic data. A crack propagation criterion is established based on this first crack density characteristic data. The crack propagation direction is determined using an energy release rate calculator. The stress concentration distribution at the crack tip is obtained using a stress concentration calculator. Combined with the initial crack stress field distribution, the initial characteristic parameters of the microcrack are obtained.

[0015] Furthermore, the interface bond strength parameters, microcrack initial characteristic parameters, interface transition zone porosity, and pore filling material distribution are integrated to simulate the pore filling rate and microcrack propagation resistance of the interface region and obtain the performance parameters of the interface transition zone. This includes: obtaining a grayscale image of the interface transition zone using a scanning electron microscope image acquisition system, optimizing the contrast of the grayscale image using a histogram equalization enhancer, and extracting the pore region boundary contours using a maximum inter-class variance segmenter to obtain a first pore feature matrix. A three-dimensional reconstructed image is constructed from CT tomography data of the interface transition zone. The reconstructed image is density-corrected using an attenuation coefficient compensator, and the filling material region is extracted using a region marker to obtain a first filling material distribution matrix. Feature fusion is performed between the first pore feature matrix and the first filling material distribution matrix, and the fused data is spatially aligned using a least squares aligner to obtain a structural distribution map of the interface transition zone. A stress field is constructed based on the structural distribution map of the interface transition zone. The stress distribution of the interface region is obtained using a finite element stress calculator, and the stress gradient field matrix is obtained using a stress gradient calculator. The crack propagation resistance is calculated based on the stress gradient field matrix and the initial characteristic parameters of the microcracks. The interfacial stress transfer characteristics are obtained through an interlaminar stress integrator, resulting in a second interfacial stress matrix. Mechanical feature extraction is performed on this second interfacial stress matrix and the interfacial bond strength parameters. A deep convolutional network is used to extract local structural features, which are then classified using a random forest classifier to obtain the performance parameters of the interface transition zone.

[0016] Furthermore, the stress distribution of microcracks as they propagate at the mortar interface is calculated, the microcrack propagation behavior is analyzed, and the geometric characteristics of the crack propagation path are calculated by combining the crack propagation path curvature and the microcrack propagation resistance to obtain the microcrack propagation path parameters. This includes: constructing a first crack distribution map based on the crack distribution density in the initial characteristic parameters of the microcracks, obtaining the crack tip stress intensity factor using a fracture critical value calculator, and calculating the crack tip stress distribution using a stress analyzer to obtain a first stress distribution matrix. An elastic-plastic mechanics solver is used to calculate the interface deformation field for the first stress distribution matrix, a grid curvature calculator is used to perform curvature analysis on the deformation field, and a minimum energy path planner is used to calculate the crack propagation direction to obtain a first path curvature matrix. A crack propagation criterion is established based on the first path curvature matrix and the microcrack propagation resistance, the atomic bond energy distribution is obtained using a molecular dynamics calculator, and the crack propagation driving force is calculated using a bond energy gradient analyzer to obtain a first propagation driving force matrix. Fracture evolution calculations are performed based on the first propagation driving force matrix. A fracture path tracker is used to obtain crack propagation trajectories. A morphological processor extracts crack propagation morphological features to generate a second path feature matrix. A geometric feature space is established based on this second path feature matrix. A deep convolutional network is used to extract path geometric features. A feature fuser combines microcrack initial stress concentration data with path geometric features to obtain microcrack propagation path parameters.

[0017] Furthermore, feature fusion is performed on the interface transition zone performance parameters and microcrack propagation path parameters, and adaptive training is performed based on the sand surface chemical activity and cement slurry hydration degree. This results in a strength prediction model that outputs the predicted value of the mortar interface strength. This model includes: constructing a first feature matrix based on the interface transition zone performance parameters and microcrack propagation path parameters, performing interval mapping on the matrix data using a maximum-minimum normalizer, and extracting characteristic principal components using a principal component analyzer to obtain first characteristic principal component data. The surface functional group distribution of the sand grains is obtained using a Raman spectrometer, the cement hydration degree data is obtained using a nuclear magnetic resonance calculator, and the interface reaction ion distribution is obtained using an ion concentration analyzer to obtain first reaction characteristic data. Data alignment is performed based on the first characteristic principal component data and the first reaction characteristic data, a multidimensional feature space is constructed using a feature combiner, and training and validation data are partitioned using a cross-validator to obtain a first training data matrix. A residual neural network structure is established for the first training data matrix, network weights are iteratively updated using a backpropagation optimizer, and prediction errors are calculated using a cross-entropy loss function calculator to obtain the first network parameter matrix. A strength prediction function is constructed based on the first network parameter matrix. The training process is dynamically adjusted using a learning rate adaptor, and the prediction error is minimized using a gradient descent optimizer to obtain the strength predictor. A validation dataset is input to the strength predictor, and the model evaluator calculates the prediction accuracy index. The parameter fine-tuner is used to locally optimize the predictor to obtain the strength prediction model, which outputs the mortar interface strength value.

[0018] Furthermore, by calculating the deviation between the predicted strength and the actual strength, a prediction accuracy index is obtained. If the prediction accuracy index does not reach a preset threshold, the parameters of the strength prediction model are updated to obtain an updated strength prediction model, including: constructing a first validation data set based on the quantized morphological feature vector and the microcrack propagation path parameters, grouping the data set through a K-fold cross-validator, and performing prediction calculations on each group of data using a strength predictor to obtain first predicted strength data. An error is calculated between the first predicted strength data and the actual strength value, a predicted deviation value is obtained through a root mean square error calculator, and the deviation value is normalized using a standardization processor to obtain a first prediction accuracy index. If the first prediction accuracy index is less than a preset threshold, a first gradient matrix is constructed based on the interface stress gradient data, the neural network parameters are adjusted through a weight optimizer, and the network weights are updated using a backpropagation calculator to obtain a first optimized weight matrix. The parameters of the predictor are updated according to the first optimized weight matrix, the updated parameters are trained through a batch gradient descent, and the training termination conditions are determined using an early stopping rule to obtain a second prediction model. The second prediction model was used to calculate the surface energy of sand particles. The surface energy distribution data was obtained using an energy distribution calculator. The energy deviation analyzer was used to calculate the quantified deviation value to obtain the second accuracy index. The prediction model was fine-tuned based on the second accuracy index. The network parameters were adjusted using an adaptive learning rate optimizer. The model performance was verified using a cross-validator to obtain the updated strength prediction model.

[0019] Furthermore, the updated strength prediction model was validated by calculating strength prediction results based on the quantitative morphological feature vectors, microcrack propagation path parameters, and cement slurry hydration levels of various concrete material systems to obtain strength prediction accuracy. This process involved constructing a first test dataset based on the quantitative morphological feature vectors of the various concrete material systems, normalizing the feature data using a maximum-minimum normalizer, and extracting the characteristic principal components using a principal component analyzer to obtain a first characteristic principal component matrix. Materials were classified based on the first characteristic principal component matrix, grouping different material systems using a cluster analyzer, and extracting the microcrack propagation path parameters for each data set using a feature mapper to obtain first path feature data. Feature combinations were established based on the first path feature data and cement slurry hydration level data. The features were combined using a multidimensional data fuser and spatiotemporally aligned using a data aligner to obtain a first training data matrix. Strength prediction was performed on the first training data matrix, with predicted strength values calculated using a deep residual network and statistically analyzed using a prediction result analyzer to obtain a first prediction result matrix. The first prediction matrix is compared with the measured strength values, and the prediction deviation is obtained using an error calculator. The prediction reliability is evaluated using a random forest classifier to obtain a first reliability index. The accuracy of the first reliability index is evaluated, and the prediction accuracy distribution is calculated using a statistical analyzer. The final evaluation result is generated using a comprehensive evaluator to obtain the final strength prediction accuracy.

[0020] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0021] This paper discloses an artificial intelligence-based concrete strength prediction method based on multi-scale morphological feature analysis. By acquiring sand grain morphology data, extracting features such as angular sharpness, surface texture complexity, and pore connectivity, the method simulates the interface contact behavior between sand and cement slurry, analyzes the initiation and propagation characteristics of microcracks, integrates performance parameters of the interface transition zone, and constructs a deep neural network model to predict interface strength. This method integrates multiple factors, including sand grain morphology, interfacial stress distribution, and microcrack behavior. Through adaptive training and cross-validation, the model is continuously optimized, achieving accurate predictions of the interface strength of various concrete material systems. This method provides an important basis for concrete material design and performance evaluation, and helps improve the durability and reliability of concrete structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of an artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis of the present invention. DETAILED DESCRIPTION

[0023] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] In this application's examples, concrete strength assessment involves digitally extracting sand grain morphology, simulating interfacial behavior, and building a deep learning model. Sand grain morphology features include sharpness, surface texture complexity, and pore connectivity, which directly influence interfacial bond strength and microcrack propagation behavior. The assessment process utilizes multiscale modeling and molecular dynamics analysis, combined with deep neural networks, to achieve feature fusion and strength prediction.

[0025] Figure 1 The implementation process of the artificial intelligence-based concrete strength assessment method provided in the embodiment of the present application is shown and is detailed as follows:

[0026] S101 obtains two-dimensional and three-dimensional morphological data of concrete sand particles, digitally extracts the sharpness of edges and corners, surface texture complexity, and pore connectivity, and generates quantitative morphological feature vectors.

[0027] In an embodiment of the present application, the strength assessment function can be integrated into the concrete analysis system or run as a standalone application. The specific startup method can be triggered by the user through an interface operation, such as clicking the assessment icon, or activated by a remote command. After startup, a two-dimensional grayscale image and three-dimensional point cloud data of the sand particles are first obtained for subsequent morphological feature extraction. After obtaining the two-dimensional and three-dimensional data of the sand particles, an edge detection algorithm (such as the Sobel operator) is used to extract contour segments from the grayscale image, and the curvature value is calculated to determine the angular area. For the three-dimensional point cloud data, filtering and registration techniques are used to generate an accurate angular area point cloud. Subsequently, the eigenvalues of the surface texture complexity and pore connectivity are extracted through methods such as Fourier transform and discrete wavelet transform. These eigenvalues together constitute the sand particle morphology feature vector. Although surface features such as angular sharpness and surface texture complexity significantly affect the interfacial bond strength, they cannot fully represent the overall strength of concrete. Therefore, the designed concrete strength prediction model does not rely solely on surface features, but takes into account multiple factors such as internal porosity and microcrack distribution to achieve a more comprehensive and accurate strength prediction.

[0028] S1011 obtains a two-dimensional grayscale image of sand particles, extracts contour segments using an edge detection algorithm, calculates curvature values, and divides angular regions according to a preset threshold to generate initial angular region data.

[0029] A scanning electron microscope (SEM) imager acquires a high-resolution grayscale image of the concrete surface for analyzing surface texture complexity. Edge detection is performed using the Sobel operator, which calculates horizontal and vertical gradients to extract contour segments. For example, when the grayscale value changes dramatically from high to low, the gradient increases significantly, indicating the presence of an edge feature. The curvature of the contour segment is calculated using a three-point circle fitting method. If the curvature exceeds a preset threshold, such as 0.5, it is identified as an angular region, and initial angular region data is generated.

[0030] S1012 performs filtering processing on the three-dimensional point cloud data, uses the initial angular area data for registration, generates precise angular area point clouds, and extracts surface texture and pore features.

[0031] Raw sand point cloud data was acquired using a 3D scanning device. A Gaussian filter was used to remove noise, with a filter radius of 0.5 mm and a standard deviation of 0.2 to ensure point cloud data accuracy. Using the initial angular region data, the point cloud was registered using an iterative closest point algorithm, with an error controlled within 0.1 mm, generating a precise angular region point cloud. A Fourier transform was performed on this point cloud to obtain a frequency domain energy distribution matrix. A preset energy threshold, such as 15% of the total energy, was used to segment the surface texture region point cloud. A discrete wavelet transform was used with the db4 wavelet basis function and a three-level decomposition scale to extract multi-scale texture feature vectors.

[0032] S1013 performs voxel processing on the point cloud of the precise angular area, constructs a pore space grid model, calculates the connected domain label matrix, counts the number of connected channels, and generates a connectivity index value.

[0033] The point cloud space was divided into a 0.2 mm cubic grid, and the number of points within the grid was counted. If the number of points was less than 3, it was determined to be a porous region. A region growing algorithm was used to merge adjacent grids using boundary points as seed points to generate a set of boundary points for the porous region. A smooth porous space grid model was constructed using the Poisson reconstruction algorithm with an octree depth of 8. Based on the 26-neighborhood connectivity criterion, the connected domain labeling matrix was calculated, and the number of connected channels was counted using a breadth-first search algorithm with an upper limit of 50 search depth and 10 mm path length. Combined with morphological analysis, the channel diameter distribution was calculated to generate a connectivity index value.

[0034] S1014 extracts the angular sharpness eigenvalue, combines the texture eigenvector and the connectivity index value, and constructs the sand grain morphology eigenvector.

[0035] For the point cloud of precise corner areas, a density clustering algorithm was used to calculate the local curvature radius and angle data, generating the corner sharpness characteristic value. Combining multi-scale texture feature vectors and connectivity index values, a 31-dimensional sand grain morphology feature vector was constructed for interface behavior analysis.

[0036] In this application, multi-dimensional processing of two-dimensional and three-dimensional sand grain morphology data enables precise extraction of angular, texture, and pore characteristics, avoiding the neglect of microscopic parameters often observed in traditional methods. The resulting morphological feature vectors provide a reliable data foundation for interfacial bond strength and microcrack analysis, thereby improving the accuracy and stability of strength predictions.

[0037] It's worth noting that while the smoothness of the formwork primarily affects the surface flatness of the poured concrete rather than its internal structural strength, in actual engineering applications, a smooth formwork can help reduce surface defects and improve the appearance of concrete. Therefore, when considering the overall performance of concrete, in addition to its internal structural properties, the importance of surface treatment should not be overlooked. For example, using a smooth formwork can ensure a more uniform and flat concrete surface, which is crucial for aesthetics and functionality in certain application scenarios. However, this surface treatment does not significantly change the internal mechanical properties of concrete, such as key indicators such as compressive strength or tensile strength.

[0038] Through the above steps, key morphological features were extracted from the 2D and 3D data of the sand particles, and their interfacial contact behavior with the cement paste was simulated. Next, based on these features and the simulation results, the stress state in the interfacial region was analyzed in depth to generate interfacial bond strength parameters. This step not only laid the foundation for subsequent microcrack analysis but also provided important input data for the overall concrete strength prediction model.

[0039] S102 calculates the stress state of the interface area and generates interface bonding strength parameters based on the angular sharpness and surface texture complexity in the quantified morphological feature vector, combined with the adhesion characteristics of the cement slurry and the interfacial stress distribution characteristics.

[0040] In the embodiment of the present application, interface performance evaluation is a key link in concrete strength prediction, which aims to quantify the mechanical behavior of sand and cement slurry in the contact area. Based on the sharpness of the edges and corners and the complexity of the surface texture in the eigenvector, the interface contact behavior of sand and cement slurry is further simulated. By calculating the stress distribution matrix at the interface grid nodes, and combining the adhesion characteristics of the cement slurry, a bonding strength parameter reflecting the interface bonding quality can be generated. This parameter is crucial for predicting the development of microcracks and the overall concrete strength. By analyzing the sharpness of the edges and corners and the complexity of the surface texture of the morphological eigenvector, combined with the adhesion characteristics of the cement slurry, the interface stress distribution can be accurately simulated, and then a bonding strength parameter reflecting the interface bonding quality can be generated. The specific implementation method can be determined by those skilled in the art according to the actual scenario, for example, stress calculation is completed by finite element analysis or molecular dynamics simulation.

[0041] S1021 constructs a geometric model of the sand surface based on edge sharpness data and uses meshing technology to generate a high-precision contact mesh for subsequent interface stress analysis.

[0042] Using the angular sharpness data in the quantified morphological feature vector, the system first constructs a three-dimensional geometric model of the sand grain surface. For areas with sharpness values exceeding 0.8, a local mesh encryption strategy is adopted to generate an initial surface mesh using the Delaunay triangulation algorithm. Next, the initial mesh is subdivided using a tetrahedral mesher to ensure that the minimum dihedral angle of the mesh unit is greater than 15 degrees and the mesh side length is controlled between 0.01 mm and 0.05 mm to improve calculation accuracy. The subdivided high-precision contact mesh can effectively capture the geometric characteristics of the angular areas, providing a reliable geometric foundation for stress analysis.

[0043] S1022 calculates the stress distribution in the interface contact area and generates interface force distribution data by combining the surface texture complexity parameters and cement slurry adhesion characteristics.

[0044] For high-precision contact meshes, a stress calculation method based on the principle of virtual work is adopted. The stress distribution matrix of the interface mesh nodes is generated by balancing the energy of the virtual displacement field and the actual displacement field. To further consider the influence of surface texture complexity, the system calculates the wetting contact angle based on the texture complexity parameter and analyzes the tension balance of the solid-liquid-gas three-phase interface based on Young's equation. When the texture complexity value is greater than 0.6, it indicates that the surface has significant concave-convex features, and the wetting contact angle is usually less than 60 degrees, showing strong wettability. Using surface energy theory, the spreading coefficient of cement slurry on the sand surface is calculated. If the spreading coefficient exceeds 0.5, the corresponding adhesion force matrix is generated. Combining the stress distribution matrix and the adhesion force matrix, the system generates the first interface force distribution data for subsequent stress field analysis.

[0045] S1023 generates the interface stress field distribution through iterative solution, calculates the interface deformation characteristics based on finite strain analysis, and generates the interface bonding index.

[0046] Based on the first interface force distribution data, the Newton-Raphson iterative solver is used for nonlinear calculation to generate the stress field distribution in the interface area. During the solution process, the iterative convergence accuracy is set to 0.001 and the maximum number of iterations is 100 to ensure the stability of the calculation results. Finite strain analysis is performed on the stress field to generate an interface deformation matrix to reflect the spatial deformation characteristics of the contact area. When the maximum principal strain exceeds 0.002, it indicates that significant deformation may have occurred in the local area. The interface bonding index is calculated by the ratio of deformation energy to surface energy. When the value is between 0.3 and 0.7, it indicates that the interface bonding degree is moderate. This index provides a key basis for subsequent bonding strength calculations.

[0047] S1024 establishes the interface stress-strain relationship, calculates the stress state characteristics, and generates interface bonding strength parameters.

[0048] The interface bonding index is used to establish a constitutive relationship between stress and strain to describe the mechanical response of the interface under external loads. The explicit integration algorithm is used, the time step is set to 0.0001 seconds, the stress state matrix of the interface area is calculated, and the shear stress and compressive stress components are extracted. When the ratio of shear stress to compressive stress is greater than 0.8, it indicates that the interface is prone to shear failure. The principal stress values are calculated using the eigenvalue decomposition method, and the difference between the maximum principal stress and the minimum principal stress reflects the local shear strength. Based on the Mohr-Coulomb criterion, combined with the conditions that the cohesion is greater than 2 MPa and the internal friction angle is greater than 30 degrees, a stress transfer rate matrix is generated. If the transfer rate value exceeds 0.85, it indicates that the interface bonding quality is high. The interface bonding index and the stress transfer rate are combined to generate the interface bonding strength parameter. When the value is greater than 3 MPa, it indicates that the interface meets the engineering strength requirements.

[0049] In this application's examples, accurate modeling of edge sharpness and surface texture complexity, combined with cement slurry adhesion properties and interfacial stress analysis, effectively quantifies the interfacial bonding performance between sand and cement slurry. The resulting bond strength parameters provide a reliable basis for microcrack analysis and strength prediction, enhancing the engineering applicability of the assessment results.

[0050] S103 If the interface bonding strength parameter does not reach the preset threshold, molecular dynamics simulation is used to analyze the initial formation characteristics of microcracks. Combined with the pore space connectivity and the initial crack width, the distribution density and stress concentration characteristics of the crack starting point are calculated to obtain the initial characteristic parameters of the microcracks.

[0051] In the examples of this application, when interfacial bonding strength is insufficient to ensure material stability, further analysis of the initial formation mechanism of microcracks is required to assess potential damage risks. Molecular dynamics simulations, combined with pore connectivity data and crack geometry, can accurately quantify the distribution and stress state of microcracks, providing key input data for subsequent strength predictions. The specific implementation method can be determined by those skilled in the art based on actual needs, such as performing simulations on a high-performance computing platform.

[0052] S1031 compares the interface bonding strength parameters with the preset threshold, calculates the interface atomic potential energy distribution based on the molecular dynamics method, and generates stress singular distribution data.

[0053] First, the interface bonding strength parameters are numerically compared with the preset threshold value. The threshold value is usually set to 2.5 MPa, which represents the minimum requirement for interface bonding quality. If the parameter is lower than the threshold, it indicates that there are potential defects in the interface and microscopic analysis is required. The molecular dynamics calculation method is used to simulate the interaction between atoms based on the Lennard-Jones potential energy function. The potential well depth is set to 0.5 electron volts and the equilibrium distance is 0.3 nanometers to generate the potential energy distribution matrix between interface atoms. Through stress field gradient analysis, areas where the local stress gradient exceeds 50 MPa per micron are identified and marked as stress singular points to generate stress singular distribution data. These singular points are usually located in areas with concentrated interface defects or pores, and the stress peak can reach 100 MPa.

[0054] S1032 constructs a three-dimensional pore network model, identifies the crack initiation point in the stress singularity area, and optimizes the mesh division to generate the initial crack stress field distribution.

[0055] Using pore space connectivity data, a three-dimensional pore network model was constructed to reflect the geometric distribution and connectivity characteristics of the pores. When the pore connectivity is greater than 0.6, the pores form complex connected channels, which significantly affect the stress field distribution. For the stress singularity distribution data, the maximum principal stress criterion was used to identify potential crack initiation points. Areas where the principal stress exceeds 0.8 times the material strength were marked as high-risk points. To improve the calculation accuracy, the mesh of these areas was refined, and the initial mesh size was optimized from 0.1 mm to 0.01 mm to generate high-resolution initial crack stress field distribution data. The stress field exhibits singular characteristics, and the stress intensity factor at the crack tip can reach 0.8 MPa·m^0.5.

[0056] S1033 analyzes the geometric shape of the crack starting point, calculates the initial crack width and distribution density, and generates crack geometric characteristic data.

[0057] Based on the initial crack stress field distribution, the crack initiation position is determined, and the crack boundary is identified using morphological methods. By calculating the image grayscale gradient, points with grayscale changes exceeding 150 are determined to be crack boundary points, and crack geometric morphology data is generated. The initial crack width is usually between 0.005 mm and 0.02 mm, reflecting the microscale characteristics of the crack. By counting the crack length per unit area, the local crack distribution density is calculated. When the density exceeds 5 mm per square millimeter, it indicates significant regional damage. The generated crack geometric feature data contains position, width and direction information, providing a basis for crack propagation analysis.

[0058] S1034 establishes crack extension criteria, calculates energy release rate and stress concentration characteristics, and generates initial characteristic parameters of microcracks.

[0059] Based on the crack geometric characteristic data, a statistical analysis method is used to generate a crack density distribution function to reflect the spatial distribution law of the crack. Based on the energy release rate criterion, the critical energy release rate is set to 0.1 joules per square meter to determine the direction of crack propagation. By calculating the stress concentration coefficient at the crack tip, when the value is between 2.5 and 3.5, it indicates that the crack has a tendency to expand. Combining the crack density weight of 0.4, the stress concentration weight of 0.35 and the expansion trend weight of 0.25, the initial characteristic parameters of the microcracks are comprehensively generated. These parameters quantify the distribution density, stress concentration degree and expansion potential of the crack, providing a reliable basis for interface performance evaluation.

[0060] In the examples of this application, molecular dynamics simulations and pore network analysis accurately capture the initial formation characteristics of microcracks. The generated initial characteristic parameters of microcracks effectively reflect the potential impact of interface defects on material strength, providing important technical support for optimizing concrete mix proportions and improving engineering durability.

[0061] S104 uses scanning electron microscopy and CT imaging technology to obtain porosity and filling material distribution data, and combines multi-scale modeling technology to integrate interface bonding strength, initial characteristics of microcracks and pore characteristics, simulate pore filling rate and microcrack propagation resistance, and obtain performance parameters of the interface transition zone.

[0062] In the examples of this application, the interface transition zone is a key area connecting sand and cement paste, and its performance directly affects the overall strength of concrete. By acquiring microstructural data through high-resolution imaging technology and combining it with multi-scale modeling and deep learning methods, it is possible to accurately simulate the mechanical behavior and damage characteristics of the interface transition zone, providing a reliable basis for concrete strength assessment. The specific implementation method can be optimized by those skilled in the art based on project requirements, such as adjusting imaging parameters or modeling scale.

[0063] S1041 uses a scanning electron microscope to collect grayscale images of the interface transition zone, enhance image contrast, extract pore boundary features, and generate pore distribution data.

[0064] A scanning electron microscope was used to collect high-resolution grayscale images of the interface transition zone. To ensure image quality, a gold film about 10 nanometers thick was sprayed on the sample surface to improve conductivity. The original image often lacks contrast, making it difficult to distinguish between pores and the matrix. Histogram equalization enhancement technology was used to expand the grayscale from 128 to 256, making the grayscale value of the pore area more significantly different from that of the matrix material. Next, the maximum inter-class variance segmentation algorithm was used to automatically identify boundaries with significant grayscale differences, extract the contour features of the pore area, and generate the first pore distribution data, including pore location and area information. The porosity is usually between 15% and 25%, reflecting the degree of looseness of the microstructure of the interface transition zone.

[0065] S1042 builds a three-dimensional pore and filling material distribution model based on CT tomography, corrects image density and extracts material area features.

[0066] A CT scan using a cone-beam X-ray source captured 2048 projection images across a 360-degree range, reconstructing a three-dimensional model with a voxel resolution of 2 microns. Due to X-ray attenuation, the grayscale values in the central region of the image were approximately 20% lower. An attenuation coefficient compensation algorithm was used to correct the grayscale distribution and ensure uniform image density. A watershed algorithm was used to mark regions of different materials. When the grayscale difference between adjacent regions exceeded 50, it was determined to be filler material or matrix, generating the initial filler material distribution data. The filler material volume fraction was approximately 35%, providing a material distribution basis for subsequent stress analysis.

[0067] S1043 integrates pore and filling material characteristic data to construct a structural distribution model of the interface transition zone and perform stress field analysis.

[0068] The first pore distribution data and the first filling material distribution data were fused, and a least-squares registration algorithm was used to spatially align approximately 200 feature points, maintaining a registration error within 3 microns. This generated a structural distribution model for the interface transition zone. This model comprehensively reflects the spatial distribution characteristics of porosity and filling material. Based on this model, the stress field was constructed using the finite element method, using a hexahedral mesh with a cell size of 5 microns, which was meshed to 1 micron at the interface to capture sudden stress changes. Considering the elastic modulus ratio of the materials on both sides of the interface to be approximately 3:1, the stress gradient field was calculated to generate stress data reflecting the local stress distribution.

[0069] S1044 simulates microcrack propagation resistance and extracts mechanical characteristics, combining deep learning to generate interface transition zone performance parameters.

[0070] The crack propagation resistance was calculated using stress gradient field data and initial microcrack characteristics. When the angle between the crack propagation direction and the interface was less than 30 degrees, the resistance increased significantly, reflecting the crack resistance of the interface transition zone. Using the Gaussian integral method, 5 integration points were set along the normal direction of the interface to calculate the stress transfer efficiency, with values ranging from 70% to 85%. Next, a deep convolutional neural network was used, with a 5-layer convolution structure set up. A local image block of 64×64 pixels was input to extract a 128-dimensional feature vector. Combined with a random forest classifier, 200 decision trees were used, and 80% of the feature dimensions were randomly selected for training to classify and generate the interface transition zone performance parameters. This parameter comprehensively considers structural integrity, stress transfer efficiency, and crack propagation resistance, with weights of 0.4, 0.35, and 0.25, respectively. When the parameter value is greater than 0.75, it indicates that the interface transition zone has excellent mechanical properties.

[0071] Through high-precision imaging and multi-scale modeling, combined with deep learning techniques, the microstructure and mechanical properties of the interface transition zone are quantified. The resulting performance parameters not only reflect pore distribution and stress transfer characteristics, but also provide data support for optimizing concrete material design, improving the reliability and durability of engineering applications.

[0072] S105 calculates the stress distribution and path curvature of crack propagation through molecular dynamics simulation and mechanical analysis. It combines the crack distribution density and stress concentration characteristics in the initial characteristics of microcracks, integrates the crack propagation resistance, extracts the path geometric characteristics, and generates microcrack propagation path parameters.

[0073] In the examples of this application, analyzing microcrack propagation behavior is a key step in assessing concrete interface stability. By combining molecular dynamics simulations with elastoplastic mechanics, the stress state and geometric characteristics of crack propagation can be accurately captured, providing reliable data for predicting material damage evolution. The specific implementation method can be optimized by those skilled in the art based on computing resources and accuracy requirements, such as adjusting the simulation scale or meshing strategy.

[0074] S1051 builds a crack distribution model based on crack distribution density, calculates the stress intensity factor at the crack tip and generates stress distribution data.

[0075] Using the crack distribution density in the initial characteristics of microcracks, a two-dimensional crack distribution model is constructed to reflect the spatial distribution of cracks within a unit area. When the density reaches 0.5 per square millimeter, the interaction between cracks is significant, increasing the complexity of expansion. Using the fracture mechanics criterion, the critical state of crack expansion is determined by calculating the stress intensity factor. If the factor exceeds 1.2 MPa·m 0.5 The crack begins to propagate. Using a stress analysis algorithm, the stress distribution at the crack tip is calculated, generating the first stress distribution data. The stress at a distance of 0.1 mm from the tip reaches three times the material strength, exhibiting typical singular characteristics and providing a basis for deformation analysis.

[0076] S1052 calculates the interface deformation field and analyzes the path curvature to determine the direction of crack propagation.

[0077] For the first stress distribution data, an elastic-plastic mechanics solution method is used, taking into account the nonlinear deformation characteristics of the material, and an interface deformation field is generated through iterative calculation. The deformation field is divided into 1000×1000 1 micron×1 micron grid cells, and the curvature value of each grid point is calculated using the central difference algorithm. When the curvature value is greater than 0.5 per millimeter, it indicates the presence of significant deformation concentration. Based on the energy release rate criterion, the system uses the minimum energy path planning algorithm to select the direction with the largest energy release rate as the crack propagation path, generating the first path curvature data, reflecting the geometric trend of crack propagation.

[0078] S1053 uses molecular dynamics simulation to analyze changes in atomic bond energy and calculate the driving force for crack propagation.

[0079] Molecular dynamics methods were used to simulate crack propagation within a nanoscale region containing approximately one million atoms, using an embedded atomic potential function to describe interatomic interactions. The atomic bond energy distribution was calculated, and when the bond energy exceeded 0.8 electron volts, the atomic bonds broke, driving crack propagation. Within 50 nanometers in front of the crack tip, the bond energy gradient reached 0.02 electron volts per nanometer, reflecting the propagation driving force. A bond energy gradient analysis algorithm was used to generate the first propagation driving force data, providing a mechanical basis for path tracing.

[0080] S1054 tracks the crack propagation trajectory and extracts geometric features to generate microcrack propagation path parameters.

[0081] Based on the first extension driving force data, an adaptive mesh refinement technique is used to track the crack propagation trajectory, with the mesh size dynamically adjusted from 1 micron to 0.1 micron to capture path details. The curvature, bifurcation angle, and deflection angle of the crack path are extracted using a morphological processing algorithm. The mean curvature is between 0.3 and 0.8 per millimeter, and the bifurcation angles are distributed between 45 and 75 degrees. Using a 5-layer deep convolutional neural network, a 128×128 pixel path image is input to extract 128-dimensional local and global feature vectors. Combined with the feature fusion algorithm, the initial stress concentration feature weight is set to 0.4, and the path geometry feature weight is set to 0.6. Microcrack propagation path parameters are generated, including the mean path curvature, maximum bifurcation angle, cumulative deflection angle, and path fractal dimension. The fractal dimension is between 1.2 and 1.8, reflecting the complexity of the path.

[0082] In the embodiments of the present application, by quantifying the mechanical and geometric properties of microcrack propagation, the generated path parameters provide a key basis for evaluating interface damage and optimizing material design, effectively improving the accuracy of concrete durability analysis and engineering applicability.

[0083] S106 uses a deep neural network algorithm to fuse the performance parameters of the interface transition zone and the microcrack propagation path parameters, combines the surface chemical activity of sand particles and the hydration degree of cement slurry for adaptive training, constructs a strength prediction model, and outputs the predicted value of mortar interface strength.

[0084] In the examples of this application, the prediction of mortar interface strength is a core component of concrete performance evaluation. By integrating multi-source feature data using a deep neural network and optimizing it with chemical reaction characteristics, the system can generate a highly accurate strength prediction model. The specific implementation method can be adjusted by those skilled in the art based on actual scenarios, such as selecting different network structures or optimization algorithms.

[0085] S1061 constructs and preprocesses the feature matrix and extracts key features to generate principal component data.

[0086] Based on the performance parameters of the interface transition zone, such as porosity and stress transfer efficiency, and the parameters of the microcrack propagation path, such as path curvature and bifurcation angle, the system constructs an initial characteristic matrix. In order to eliminate dimensional differences, the maximum and minimum value normalization method is used to map the eigenvalues to the range of 0 to 1 to ensure data consistency. Then, using the principal component analysis algorithm, the original high-dimensional features are reduced to 10-dimensional principal components by calculating the eigenvalues and eigenvectors, retaining 95% of the variance contribution rate to generate the first principal component data. This dimensionality reduction process not only reduces the computational complexity, but also highlights the most important features for strength prediction, providing efficient input for subsequent model training.

[0087] S1062 collects chemical reaction data between the sand surface and cement slurry to generate reaction characteristic data.

[0088] Raman spectroscopy was used to analyze the chemical activity of the sand surface, collecting spectral data from 400 to 4000 wavenumbers, focusing on the peak intensity of the silicon-oxygen bond at 1100 wavenumber, which reflects the distribution characteristics of surface functional groups. Nuclear magnetic resonance technology was used to measure the degree of cement slurry hydration, and the change in the hydrogen atom T2 relaxation time from an initial 150 microseconds to 600 microseconds characterized the progress of the hydration reaction. Simultaneously, ion concentration analysis was used to monitor the change in interfacial calcium ion concentration, which decreased from 0.015 mol / L to 0.005 mol / L, reflecting chemical reaction activity. These data were integrated to generate first-line reaction characteristic data, providing input for the model from a chemical reaction perspective.

[0089] S1063 constructs a residual neural network and performs training to generate a network parameter matrix.

[0090] The first principal component data was aligned with the first reaction feature data, and a time series alignment algorithm was used to unify data at different time scales, constructing a 32-dimensional multidimensional feature space. A 5-fold cross-validation was used to partition the training and validation datasets, with 80% used for training and 20% for validation. A neural network consisting of 16 residual blocks was constructed, each consisting of two layers of convolution and skip connections to mitigate the vanishing gradient problem. Backpropagation was performed using the Adam optimizer with an initial learning rate of 0.001. The prediction error was calculated using the cross-entropy loss function, with the loss value decreasing from 0.8 to below 0.1. The resulting first network parameter matrix provided the basis for subsequent prediction functions.

[0091] S1064 optimizes the strength prediction function and generates a final model, outputting the interface strength prediction value.

[0092] Based on the first network parameter matrix, a strength prediction function was constructed. A cosine annealing strategy was used to dynamically adjust the learning rate, periodically decreasing it from 0.001 to 0.0001 to improve training stability. Mini-batch stochastic gradient descent with a batch size of 64 was used to optimize the prediction error. On a validation dataset containing 200 samples, the model achieved a root mean square error of 0.15 MPa. Hyperparameters such as the convolution kernel size and the number of hidden layer neurons were fine-tuned using grid search to generate the final strength prediction model. The model achieved a prediction accuracy of 90% on the test set, with an error within 0.3 MPa. The output reflects the mortar interface strength values at different mix ratios and ages.

[0093] By integrating interface properties, crack characteristics, and chemical reaction data, combined with the powerful fitting capabilities of residual neural networks, the system can efficiently predict mortar interface strength. The generated model provides precise guidance for the optimal design of concrete materials, helping to improve the safety and durability of engineering structures.

[0094] S107 tests the model performance through the cross-validation method. It combines the quantitative morphological feature vector, microcrack propagation path parameters and sand particle surface energy distribution to calculate the deviation between the predicted strength and the actual strength and generate an accuracy index. If the accuracy does not meet the standard, the neural network parameters are adjusted according to the interface stress gradient and surface energy deviation to generate an updated strength prediction model.

[0095] In the examples of this application, validation and optimization of the strength prediction model are key steps to ensure its reliability in engineering applications. By validating model performance with multi-dimensional feature data and adjusting parameters to address deficiencies, the system can improve prediction accuracy and adapt to the strength prediction needs of different material ratios and ages. The specific implementation method can be optimized by those skilled in the art based on the scale of the validation dataset and computing resources.

[0096] S1071 constructs a validation dataset and performs cross-validation to generate prediction strength data.

[0097] The first validation dataset, consisting of 800 sets of samples, was constructed using quantitative morphological feature vectors, including porosity, surface roughness, and crack density, as well as microcrack propagation path parameters such as path curvature and bifurcation angle. All features were normalized and mapped to the range of 0 to 1 to eliminate dimensional differences. A 5-fold cross-validation method was used to randomly divide the dataset into 5 subsets, with 4 subsets selected for training each time and 1 subset for testing. Each set of data was predicted using the strength prediction model to generate the first predicted strength data. The correlation coefficient between the predicted value and the actual strength was above 0.85, indicating that the model has good preliminary fitting ability.

[0098] S1072 calculates the prediction deviation and evaluates the model accuracy to generate a first accuracy index.

[0099] The root mean square error (RMS) method is used to calculate the prediction deviation between the first predicted intensity data and the actual intensity value. For example, if the predicted intensity is 35 MPa and the actual intensity is 32 MPa, the deviation is 3 MPa. Through normalization, the deviation value is normalized to the range of 0 to 1 to ensure that deviations of different magnitudes are comparable. The generated first prediction accuracy index reflects model performance. When the index falls below the preset threshold of 0.85, it indicates that the prediction accuracy is insufficient and further optimization is required. Deviation analysis helps identify prediction biases in specific features of the model and provides a basis for subsequent adjustments.

[0100] S1073 adjusts the neural network weights based on the interface stress gradient and updates the model parameters.

[0101] If the first prediction accuracy metric is not met, the system constructs a first gradient matrix based on the interface stress gradient data. The gradient values range from 0.5 MPa / micron to 2.5 MPa / micron, reflecting the severity of the stress change. The Adam optimizer is used to adjust the neural network weights, with an initial learning rate of 0.001 and a momentum term to accelerate convergence. During backpropagation, the weight update is proportional to the gradient, with an update coefficient of 0.1 to prevent drastic parameter fluctuations. The first optimized weight matrix is generated and trained using batch gradient descent using mini-batches of 128 samples, with each epoch containing 50 batches. Early stopping is used to monitor the validation set loss. Training is terminated when the loss has not decreased for 10 consecutive epochs, generating the updated model parameters.

[0102] S1074 fine-tunes the model based on the surface energy distribution of sand particles to generate the final strength prediction model.

[0103] For the updated model, the surface energy distribution of sand particles was calculated. Based on the surface tension test data, the surface energy value ranged from 40 millijoules per square meter to 80 millijoules per square meter, reflecting the change in interface bonding strength. The energy deviation analysis method was used to calculate the quantitative deviation and generate a second accuracy index, which comprehensively predicted the deviation weight of 0.6 and the surface energy deviation weight of 0.4. If the second accuracy index exceeds 0.9, it indicates that the model performance meets the requirements. The learning rate was further adjusted through the cosine annealing strategy, periodically reduced from 0.001 to 0.0001, and the weights of the last two fully connected layers were fine-tuned, with the adjustment range controlled within 20% of the original weights. Cross-validation was performed again, and the verification results showed that the prediction error was reduced from 15% to 10%, generating a final strength prediction model with excellent generalization ability.

[0104] In the examples presented here, multiple rounds of cross-validation and parameter optimization, combined with analysis of interfacial stress gradients and surface energy distribution, significantly improved the accuracy of the strength prediction model. The updated model can accurately predict the interfacial strength of mortars of varying mix proportions and ages, providing reliable technical support for material design and engineering applications.

[0105] S108 verifies the model through batch test data, combines the quantitative morphological feature vector, microcrack propagation path parameters and cement slurry hydration degree, calculates the predicted strength value and analyzes the deviation to generate the final strength prediction accuracy.

[0106] In the examples of this application, batch testing is a key step in verifying the generalization capabilities of the strength prediction model, aiming to ensure the model's reliability in different concrete material systems. Through multidimensional feature processing, deep learning prediction, and error analysis, high-precision strength prediction results can be generated, providing technical support for engineering applications. The specific implementation method can be optimized by those skilled in the art based on material properties and testing requirements.

[0107] S1081 preprocesses the test data and extracts the principal components to generate a feature matrix.

[0108] The system uses a maximum-minimum normalization method to map the eigenvalues of various concrete material systems, including approximately 30 parameters such as porosity, surface roughness, and crack density, to a range of 0 to 1, eliminating dimensional differences. Principal component analysis (PCA) is used to calculate eigenvalues and eigenvectors, reducing the 30-dimensional features to 8 principal components while retaining 95% of the variance contribution, generating a first-character principal component matrix. This dimensionality reduction effectively reduces computational complexity while preserving key information, providing high-quality input data for subsequent classification and prediction.

[0109] S1082 classifies the material system and extracts path features to generate path feature data.

[0110] Based on the first characteristic principal component matrix, a K-means clustering algorithm was used to classify concrete samples into four strength categories, ensuring that the Euclidean distance between cluster centers was greater than 0.5 to distinguish between different material systems. For each data set, a nonlinear mapping function was used to extract microcrack propagation path parameters, including crack propagation direction, rate, and bifurcation angle, to generate first path characteristic data. This path feature extraction ensures that the geometric characteristics of crack propagation behavior can be effectively quantified, providing dynamic damage information for model prediction.

[0111] S1083 fuses multi-dimensional features and performs spatiotemporal registration to construct a training data matrix.

[0112] Combining the first path feature data with cement slurry hydration data, the system measured the hydration process using nuclear magnetic resonance imaging. The T2 relaxation time increased from 150 to 600 microseconds, reflecting the extent of the hydration reaction. A weighted combination approach was used for multidimensional data fusion, with path feature weights set to 0.4, hydration weights to 0.3, and morphological weights to 0.3. A spatiotemporal registration algorithm was used to ensure that data collected at different time points were aligned to a consistent spatial location, generating the first training data matrix. This registration process improved data consistency and provided reliable input for the deep learning model.

[0113] S1084 uses a deep residual network to perform intensity prediction and evaluate the accuracy to generate the final prediction accuracy.

[0114] For the first training data matrix, a deep residual network (DRN) consisting of 16 residual blocks was used for strength prediction. Each residual block consists of two layers of convolution and skip connections. It takes a 32-dimensional feature vector as input and outputs a predicted strength value. The prediction results show a correlation coefficient of 0.92 between the predicted and measured strengths. For example, the predicted strength of C30 concrete was 35 MPa, while the measured value was 32 MPa, resulting in a relative error of 9.4%. Prediction deviations were calculated using an error calculation method. A random forest classifier, based on 200 decision trees, was used to assess prediction reliability. A reliability index greater than 0.85 indicates high confidence. Statistical analysis showed an average prediction accuracy of 91% with a standard deviation of 3%. The accuracy for concrete grades C20 to C50 exceeded 85%. A weighted score was used for the comprehensive evaluation, with a weight of 0.5 for prediction accuracy, 0.3 for reliability, and 0.2 for generalization performance. The resulting final strength prediction accuracy was 0.89.

[0115] In this application's examples, batch testing and multidimensional feature analysis, combined with the predictive capabilities of a deep residual network, systematically validated the model's stability and high accuracy across diverse concrete material systems. The resulting prediction accuracy provides strong data support for optimizing material mixes and improving structural durability.

[0116] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence concrete strength prediction method based on multi-scale morphological feature analysis, characterized in that: The method comprises: Acquire the morphological data of concrete sand particles, extract the sharpness of sand particles’ edges and corners, surface texture complexity and pore space connectivity, and obtain the sand particle morphological feature vector; The interfacial contact behavior between sand and cement slurry is simulated. Based on the sharpness of the edges and corners and the complexity of the surface texture in the sand morphology feature vector, combined with the cement slurry adhesion and interfacial stress gradient, the stress distribution in the interfacial area is calculated to obtain the interfacial bond strength parameter. If the interface bonding strength parameter is lower than the preset threshold, the initial starting point of the microcrack is analyzed. Based on the pore space connectivity and the initial width of the microcrack, the crack distribution density and the initial stress concentration at the crack starting point are calculated to obtain the initial characteristic parameters of the microcrack. The fusion interface bonding strength parameters, microcrack initial characteristic parameters, interface transition zone porosity and pore filling material distribution are simulated to simulate the pore filling rate and microcrack propagation resistance of the interface area and obtain the interface transition zone performance parameters; Calculate the stress distribution when microcracks propagate at the mortar interface, analyze the microcrack propagation behavior, and combine the crack propagation path curvature and microcrack propagation resistance to calculate the geometric characteristics of the crack propagation path and obtain the microcrack propagation path parameters; The performance parameters of the interface transition zone and the microcrack propagation path parameters are fused, and adaptive training is performed on the chemical activity of the sand surface and the hydration degree of the cement slurry to obtain a strength prediction model whose output is the predicted value of the mortar interface strength.

2. The method according to claim 1, characterized in that The method of obtaining the morphological data of concrete sand particles, extracting the sharpness of the sand particles' edges and corners, the complexity of the surface texture, and the connectivity of the pore space, and obtaining the sand particle morphological feature vector includes: Extracting contour segments using the Sobel operator based on the grayscale image of the sand grains, performing curvature calculation on the contour segments to obtain a first curvature value array, and obtaining initial sand grain angular area data by dividing the area using a preset curvature distinction threshold; The filtered point cloud data is registered with respect to the initial sand grain angular region data to obtain a sand grain angular region point cloud, and the sand grain angular region point cloud is Fourier transformed to obtain a frequency domain energy distribution matrix; A surface texture area point cloud set is obtained by segmenting the surface texture area point cloud set according to the frequency domain energy distribution matrix through a preset frequency domain energy threshold, and a multi-scale texture feature vector is obtained by performing discrete wavelet transform on the surface texture area point cloud set; The point cloud of the sand grain corner area is voxelized to obtain a pore space grid model, the connected domain label matrix is calculated based on the pore space grid model, and the number of connected channels is counted using a breadth-first search algorithm to obtain a connectivity index value; The local curvature radius and angle data are extracted from the point cloud of the sand grain angular area. The density clustering algorithm is used to calculate the angular sharpness characteristic value. The sand grain morphology characteristic vector is constructed by combining the multi-scale texture feature vector and the connectivity index value.

3. The method according to claim 1, characterized in that The simulated interfacial contact behavior between sand and cement slurry is based on the sharpness of the edges and corners and the complexity of the surface texture in the sand morphology feature vector, combined with the cement slurry adhesion and interfacial stress gradient, to calculate the stress distribution in the interface area and obtain the interfacial bonding strength parameters, including: constructing a first sand grain surface mesh according to the angular sharpness data in the sand grain morphology feature vector, and subdividing the first sand grain surface mesh using a tetrahedral mesher to obtain a second subdivided mesh; For the second subdivided grid, a contact stress calculator is used to obtain an interface grid node stress distribution matrix, and the first interface force distribution matrix is obtained by combining the adhesion force matrix obtained by calculating the interfacial tension based on the surface energy. The first interface force distribution matrix is calculated using a Newton-Raphson iterative solver to obtain an interface stress field distribution, and an interface deformation matrix and a first interface bonding coefficient are obtained by performing a finite strain analysis on the interface stress field distribution. A stress-strain constitutive relationship is established according to the first interface bonding coefficient, an interface stress state matrix is obtained through an explicit stress transfer calculator, and a principal value calculation is performed on the interface stress state matrix to obtain an interface bonding strength parameter.

4. The method according to claim 1, wherein If the interface bonding strength parameter is lower than the preset threshold, the initial starting point of the microcrack is analyzed, and the crack distribution density and initial crack stress concentration at the crack starting point are calculated based on the pore space connectivity and the initial width of the microcrack to obtain the initial characteristic parameters of the microcrack, including: A numerical comparison is performed based on the interface bonding strength parameter and a preset threshold value. If the interface bonding strength parameter is less than the preset threshold value, a molecular dynamics calculator is used to obtain the interface interatomic potential energy distribution matrix, and a stress singular value calculator is used to obtain the stress singular distribution matrix. Performing stress peak identification on the stress singular distribution matrix, using a fracture stress calculator to identify the stress peak points in the stress singular distribution matrix, and obtaining an initial crack stress field distribution matrix through a mesh refinement processor; Determining the crack initiation position according to the initial crack stress field distribution matrix, performing stress gradient analysis on the crack initiation position using a stress gradient calculator, and obtaining a crack geometry matrix using a crack boundary identifier; The local crack distribution density is calculated based on the crack geometry matrix, a crack density distribution function is obtained using a fracture morphology analyzer, the crack propagation direction is determined using an energy release rate calculator, and initial characteristic parameters of microcracks are obtained.

5. The method according to claim 1, wherein The fusion interface bonding strength parameters, microcrack initial characteristic parameters, interface transition zone porosity and pore filling material distribution are simulated to simulate the pore filling rate and microcrack propagation resistance of the interface area to obtain the interface transition zone performance parameters, including: A scanning electron microscope image collector is used to acquire a grayscale image of the interface transition region, the grayscale image is enhanced by a histogram equalization enhancer, and the boundary contour of the pore region is extracted to obtain a first pore feature matrix; constructing a three-dimensional reconstructed image based on the CT tomography data, performing density correction on the reconstructed image, and extracting the filling material region using a region marker to obtain a first filling material distribution matrix; Performing feature fusion on the first pore feature matrix and the first filling material distribution matrix, and spatially aligning the fused data using a least squares register to obtain an interface transition zone structure distribution map; A stress field is constructed based on the structural distribution map of the interface transition zone, the stress distribution in the interface area is obtained through a finite element stress calculator, and the performance parameters of the interface transition zone are obtained by extracting local structural features using a deep convolutional network.

6. The method according to claim 1, characterized in that The calculation of the stress distribution of microcracks when they propagate at the mortar interface, the analysis of the microcrack propagation behavior, the calculation of the geometric characteristics of the crack propagation path in combination with the crack propagation path curvature and the microcrack propagation resistance, and the acquisition of the microcrack propagation path parameters include: constructing a first crack distribution map according to the crack distribution density in the microcrack characteristic parameters, obtaining a crack tip stress intensity factor through the first crack distribution map, and obtaining a first stress distribution matrix using a stress analyzer; Calculating the interface deformation field using an elastic-plastic mechanics solver for the first stress distribution matrix, performing curvature analysis on the deformation field using a grid curvature calculator, and obtaining a first path curvature matrix using a minimum energy path planner; Establishing a crack extension criterion based on the first path curvature matrix, obtaining atomic bond energy distribution through a molecular dynamics calculator, and obtaining a first extension driving force matrix using a bond energy gradient analyzer; A fracture path tracker is used to obtain the crack propagation trajectory for the first extension driving force matrix, the geometric features of the crack propagation trajectory are extracted through a deep convolutional network, and a feature fuser is used to obtain the microcrack propagation path parameters.

7. The method according to claim 1, characterized in that The feature fusion of the interface transition zone performance parameters and the microcrack propagation path parameters is performed, and adaptive training is performed on the sand surface chemical activity and cement slurry hydration degree to obtain a strength prediction model that outputs a predicted value of the mortar interface strength, including: Constructing a first characteristic matrix based on the interface transition zone performance parameters and the microcrack propagation path parameters, performing interval mapping on the first characteristic matrix using a maximum and minimum value normalizer, and extracting characteristic principal components using a principal component analyzer to obtain first characteristic principal component data; The surface functional group distribution data of the sand particles is obtained by using a Raman spectrometer, the cement hydration degree data is obtained by using a nuclear magnetic resonance calculator, and the interface reaction ion distribution data is obtained by using an ion concentration analyzer to obtain the first reaction characteristic data. Establishing a residual neural network structure based on the first characteristic principal component data and the first reaction characteristic data, iteratively updating the network structure through a backpropagation optimizer, and calculating the prediction error using a cross entropy loss function calculator to obtain a first network parameter matrix; An intensity prediction function is constructed according to the first network parameter matrix, the prediction function is dynamically adjusted through a learning rate adaptor, and the prediction function is optimized using a gradient descent optimizer to obtain an intensity prediction model.

8. The method according to claim 1, characterized in that The method further includes: obtaining a prediction accuracy index by calculating the deviation between the predicted intensity and the actual intensity, and if the prediction accuracy index does not reach a preset threshold, updating the intensity prediction model parameters to obtain an updated intensity prediction model; The updated strength prediction model was verified, and the strength prediction results were calculated for the quantitative morphological feature vectors, microcrack propagation path parameters and cement paste hydration degree of various concrete material systems to obtain the strength prediction accuracy.

9. The method according to claim 8, characterized in that The method of obtaining a prediction accuracy index by calculating the deviation between the predicted intensity and the actual intensity and updating the intensity prediction model parameters to obtain an updated intensity prediction model if the prediction accuracy index does not reach a preset threshold is as follows: Constructing a first validation data set based on the gravel morphology feature vector and the microcrack propagation path parameters, and performing group calculation on the validation data set using a K-fold cross-validator to obtain first predicted strength data; performing an error calculation on the first predicted intensity data and the actual intensity value, obtaining a prediction deviation value through a root mean square error calculator, and obtaining a first prediction accuracy index; If the first prediction accuracy index is less than a preset accuracy threshold, constructing a first gradient matrix based on the interface stress gradient data, and adjusting the neural network parameters through a weight optimizer to obtain a first optimized weight matrix; The parameters of the intensity prediction model are updated according to the first optimized weight matrix, and the updated parameters are trained through a batch gradient descent to obtain an updated intensity prediction model.

10. The method according to claim 8, characterized in that The updated strength prediction model is verified to calculate strength prediction results based on the quantitative morphological feature vectors, microcrack propagation path parameters, and cement paste hydration degree of various concrete material systems to obtain strength prediction accuracy, including: According to the gravel morphology feature vector of the concrete material system, the first characteristic principal component matrix is obtained by normalizing it through the maximum and minimum value normalizer; For the first characteristic principal component matrix, a cluster analyzer is used to group the material systems, and a feature mapper is used to extract microcrack propagation path parameters to obtain first path characteristic data; Based on the first path feature data, the features are combined and processed by a multidimensional data fuser, and spatiotemporal registration is performed using a data aligner to obtain a first training data matrix; For the first training data matrix, the concrete strength value is calculated by using a deep residual network, and the predicted value is statistically analyzed by a prediction result analyzer to obtain a prediction result matrix; The prediction result matrix is compared with the measured strength value, and the prediction deviation value is obtained through the error calculator to obtain the strength prediction accuracy.

Images