Mechanical property prediction method and system for machine-made sand concrete

Abstract

The invention relates to the technical field of concrete material performance prediction, in particular to a mechanical property prediction method and system for machine-made sand concrete. The method comprises the following steps: obtaining grain shape characteristics and grading characteristics of machine-made sand and hydration potential characteristics of a cementing material; fusing the grain shape feature, the grading feature and the hydration potential feature through a graph neural network to obtain a raw material intrinsic fusion feature; processing microscopic structure image features mapped by a real grain shape by using a knowledge graph to obtain a stress nephogram and a damage evolution feature, and obtaining a physical constraint gradient feature by combining the stress nephogram, the damage evolution feature and an experiment real strength value; quantifying the correlation among the structure image features, the mesoscopic mechanical response features and the mechanical properties of the concrete to obtain strong correlation mesoscopic structure features; and on the basis of the raw material intrinsic fusion features, the physical constraint gradient features and the strong correlation microstructure features, utilizing a multi-modal fusion physical information neural network to predict the mechanical properties of the machine-made sand concrete.

Description

Technical Field

[0001] This invention relates to the field of concrete material performance prediction technology, specifically to a method and system for predicting the mechanical properties of manufactured sand concrete. Background Technology

[0002] Manufactured sand has become an important raw material in concrete preparation, and its properties directly affect the final mechanical properties of concrete. Accurately predicting the mechanical properties of manufactured sand concrete is of great significance for optimizing mix design and ensuring project quality. Currently, the prediction of concrete performance mainly relies on empirical formulas or micromechanical simulation methods based on fixed parameters. However, manufactured sand has characteristics such as complex particle shape, large gradation fluctuations, and high stone powder content. These characteristics and their interaction with the hydration process of cementitious materials are complex nonlinear interactions. Traditional empirical formulas often only consider a limited set of mix proportion parameters and cannot effectively incorporate the micromechanical characteristics of manufactured sand, such as particle shape and gradation. Traditional micromechanical simulation methods mostly use idealized aggregate models and fixed material mechanics parameters, which cannot truly reflect the influence of the unique particle shape, gradation, and stone powder distribution of different batches of manufactured sand on the microstructure, nor can they integrate the nonlinear hydration potential of the cementitious material system. These methods cannot simultaneously and accurately characterize the multidimensional characteristics of manufactured sand and their interactions, resulting in insufficient prediction accuracy and limited generalization ability for the mechanical properties of manufactured sand concrete. Therefore, there is an urgent need for a high-precision prediction method that can deeply integrate the multidimensional intrinsic characteristics, microstructural response characteristics and physical and mechanical constraints of manufactured sand raw materials. Summary of the Invention

[0003] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a method for predicting the mechanical properties of manufactured sand concrete, comprising the following steps: The process involves acquiring the particle shape, gradation, and hydration potential characteristics of manufactured sand, as well as the hydration potential characteristics of the cementitious materials. A graph neural network is used to fuse these characteristics to obtain intrinsic fusion features of the raw materials. A knowledge graph is used to process the microstructural image features mapped from the actual particle shape to obtain stress cloud maps and damage evolution features. These stress cloud maps and damage evolution features are then combined with experimental real strength values ​​to obtain physical constraint gradient features. The correlation between structural image features, micromechanical response features, and concrete mechanical properties is quantified to obtain strongly correlated microstructural features. Based on the intrinsic fusion features of the raw materials, the physical constraint gradient features, and the strongly correlated microstructural features, a multimodal fusion physical information neural network is used to predict the mechanical properties of manufactured sand concrete.

[0004] Optionally, obtaining the particle shape characteristics of manufactured sand includes the following steps: Extract particle contours from particle images of manufactured sand, calculate the curvature of pixels in the particle contours, select pixels with curvature values ​​greater than a preset curvature threshold as corner points, and calculate the core angularity index based on the number of corner points and the average curvature. Box-dimensional fractal calculations are performed on the texture regions inside the particles to obtain the texture fractal dimension, which characterizes the texture complexity of the particle surface. By fusing the nuclear angularity index and the texture fractal dimension, the grain shape feature is obtained.

[0005] Optionally, obtaining the gradation characteristics of manufactured sand includes the following steps: Obtain the cumulative sieve residue data and stone powder content data of manufactured sand; The basic fineness modulus is calculated based on the cumulative sieve residue data. Based on the stone powder content data, the basic fineness modulus is corrected to obtain the stone powder sensitive corrected fineness modulus. The gradation characteristics are generated by combining the stone powder sensitive correction fineness modulus and the sieve residue ratio of each sieve grade.

[0006] Optionally, obtaining the hydration potential characteristics of the cementitious material includes the following steps: Based on raw material measurement data, the measurement data is mapped to a high-dimensional space through kernel principal component analysis in order to capture the nonlinear interactions between raw materials. Based on the measured data of hydration heat, features strongly correlated with the hydration process are selected from the high-dimensional spatial features. The selected features are then subjected to dimensionality reduction processing to obtain the hydration potential features.

[0007] Optionally, the step of fusing the particle shape feature, the gradation feature, and the hydration potential feature through a graph neural network to obtain the intrinsic fused features of the raw materials includes the following steps: A heterogeneous graph is constructed using the particle shape feature, the gradation feature, and the hydration potential feature, wherein the particle shape feature, the gradation feature, and the hydration potential feature are respectively used as nodes, and the edges between nodes are established based on the physical relationship between the features and assigned initial weights. The heterogeneous graph is processed using a graph attention network. The weights of the edges are dynamically adjusted through the attention mechanism, and the fusion representation of the nodes is learned to obtain the intrinsic fusion features of the raw materials.

[0008] Optionally, the step of using knowledge graphs to process the microstructural image features of real grain shape mapping to obtain stress cloud maps and damage evolution features includes the following steps: The characteristics of the microstructure model are matched with the knowledge graph of concrete micromechanics to determine the mechanical parameters of each component phase of the microstructure. Based on the mechanical parameters, a finite element mechanical simulation was performed on the microstructure model to obtain stress cloud diagrams and damage evolution curves; At least one stress distribution feature is extracted from the stress cloud map, and at least one damage development feature is extracted from the damage evolution curve. The stress distribution feature and the damage development feature are then combined to form the stress cloud map and the damage evolution feature.

[0009] Optionally, obtaining the physical constraint gradient characteristics by combining the stress cloud map with the damage evolution characteristics and the actual experimental strength values ​​includes the following steps: Establish a regression model between the stress cloud diagram, damage evolution characteristics, and measured values ​​of concrete mechanical properties; Calculate the SHAP value of each feature in the stress cloud map and damage evolution characteristics for the prediction result of the regression model; Based on the physical mechanism of concrete mechanics, the sign of the SHAP value of each feature is verified, and features that conform to physical laws are selected. The gradient of the SHAP value of each selected feature is used as the physical constraint gradient feature.

[0010] Optionally, obtaining strongly correlated microstructural features by quantifying the correlation between structural image features, micromechanical response features, and concrete mechanical properties includes the following steps: Calculate the signal-to-noise ratio of each feature in the microstructural model and stress cloud map and damage evolution features. The signal-to-noise ratio is the ratio of the absolute value of the correlation coefficient between the feature and the measured value of concrete mechanical properties to the coefficient of variation of the corresponding feature. Features with a signal-to-noise ratio greater than a preset threshold are selected, and the selected features are spliced ​​and normalized to obtain the strongly correlated microstructural features.

[0011] Optionally, the step of predicting the mechanical properties of manufactured sand concrete using a multimodal fusion physical information neural network based on the intrinsic fusion characteristics of the raw materials, the physical constraint gradient characteristics, and the strongly correlated microstructural characteristics includes the following steps: A multimodal fusion physical information neural network is constructed. The loss function of the multimodal fusion physical information neural network includes a data loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the concrete mechanics constitutive equation or the damage evolution equation. The physical constraint gradient features, the intrinsic fusion features of the raw materials, and the strongly correlated microstructure features are respectively input into the physical information neural network, and the multimodal fusion physical information neural network is trained using training samples; The features of the sample to be predicted are input into the trained physical information neural network to obtain the predicted mechanical properties of the manufactured sand concrete.

[0012] Secondly, to efficiently execute the mechanical property prediction method for manufactured sand concrete provided by this invention, this invention also provides a mechanical property prediction system for manufactured sand concrete, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory stores program instructions for the mechanical property prediction method for manufactured sand concrete. The mechanical property prediction system for manufactured sand concrete of this invention has a compact structure and stable performance, and can stably execute the mechanical property prediction method for manufactured sand concrete provided by this invention, further enhancing the overall applicability and practical application capability of this invention.

[0013] Beneficial Effects: This invention comprehensively and quantitatively captures the key intrinsic properties of raw materials affecting concrete performance by acquiring the particle shape, gradation, and hydration potential characteristics of manufactured sand and cementitious materials. Based on these characteristics, a microstructural model is constructed, generating a microstructure that closely resembles the actual particle shape, gradation, and stone powder distribution of manufactured sand, providing a reliable structural foundation for subsequent mechanical analysis. By performing micromechanical simulation on this model and extracting mechanical features, a direct bridge is established from microstructure to macroscopic mechanical response. Furthermore, physical constraint gradient features are obtained based on mechanical features, ensuring that the extracted features conform to the physical laws of concrete mechanics and quantifying their influence sensitivity. Intrinsic fusion features of raw materials are obtained based on multiple types of raw material features, and the physical correlation between features is captured using a graph structure, achieving deep information fusion rather than simple splicing. Strongly correlated microstructural features are obtained based on physical constraint gradient features, intrinsic fusion features of raw materials, and features from the microstructural model. Redundancy and noise are removed through signal-to-noise ratio filtering, improving feature quality. Ultimately, based on the three types of highly condensed and physically information-rich multimodal features mentioned above, prediction is performed using a physical information neural network that incorporates physical equation constraints. This ensures that the model is not only data-driven but also constrained by physical laws. Therefore, this invention can more accurately simulate the formation process of manufactured sand concrete from raw material properties to microstructure and then to macroscopic mechanical properties, significantly improving prediction accuracy and model generalization ability, and effectively overcoming the deficiency of insufficient prediction accuracy. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a method for predicting the mechanical properties of manufactured sand concrete, provided in an embodiment of the present invention; Figure 2 A framework diagram of a mechanical property prediction system for manufactured sand concrete provided in an embodiment of the present invention. Detailed Implementation

[0015] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0016] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0017] Please see Figure 1 This invention provides a method for predicting the mechanical properties of manufactured sand concrete, comprising the following steps: S1. Obtain the particle shape characteristics, gradation characteristics, and hydration potential characteristics of manufactured sand and cementitious materials.

[0018] In this embodiment, obtaining the particle shape characteristics of manufactured sand includes the following steps: S111. Extract particle contours from the particle image of manufactured sand, calculate the curvature of the pixels of the particle contours, select pixels with curvature values ​​greater than a preset curvature threshold as corner points, and calculate the core angularity index based on the number of corner points and the average curvature.

[0019] First, the acquired color two-dimensional projection image is converted into a grayscale image. By preserving the brightness information of the pixels and removing the differences in color channels, the interference of different colors on the recognition of particle outlines and textures is avoided. The conversion process uses a weighted average method to ensure that the grayscale difference between particles and background remains stable. It can adaptively adjust according to the grayscale distribution of different batches of images, converting grayscale images into black and white binary images, where the particle area is set as the foreground (black) and the background area is set as white, effectively separating particles from the background.

[0020] Then, a 3×3 structuring element is used to perform opening operations on the binary image, first eroding and then dilating. This prioritizes removing tiny noise points smaller than 5 pixels in the image (such as dust and light spot interference during image acquisition) while preserving the original edge contours of the particles to the greatest extent possible, avoiding the loss of angular features caused by excessive edge smoothing.

[0021] Furthermore, the Canny edge detection algorithm is adopted, which extracts the complete contour of particles through Gaussian filtering for noise reduction, calculation of gradient magnitude and direction, non-maximum suppression, dual threshold detection, and edge connection.

[0022] Next, the curvature of each particle outline pixel is calculated. Based on the extracted particle outline pixel coordinates, the curvature value of each pixel is calculated using the numerical differentiation method. The larger the curvature value, the more obvious the curvature of the point, that is, the sharper the particle edge. According to the crushing characteristics of manufactured sand particles, the curvature threshold is set to 0.8. Pixels with curvature greater than this threshold are selected as "corner points" and smooth edge points with small curvature are removed to ensure that the selected corner points can truly reflect the sharpness of manufactured sand particles and avoid misjudging smooth edges as corners.

[0023] Based on this, taking the particle's centroid as the origin, the polar angle distribution density of the corner points is statistically analyzed to calculate the "core angle index": number of corner points / total number of outline pixels × average corner curvature, which characterizes the density and sharpness of the corners. Specifically, the centroid position of each particle is first calculated using the outline pixel coordinates. A polar coordinate system is established with the centroid as the origin, and the polar angle of each corner point relative to the centroid is calculated. The number of corner points in different polar angle intervals is statistically analyzed to obtain the polar angle distribution density of the corner points, reflecting the uniformity of the corner distribution on the particle surface. Then, the core angle index is calculated, which consists of two parts: first, the ratio of the number of corner points to the total number of outline pixels, which characterizes the density of the corners; the larger the ratio, the more corners there are on the particle surface; second, the average curvature value of all corner points, which characterizes the sharpness of the corners; the larger the average value, the sharper the corners. The two are multiplied to obtain the core angle index, realizing the three-dimensional quantification of the "number + sharpness + distribution" of the corners of manufactured sand particles.

[0024] S112. Perform box-dimensional fractal calculations on the texture region inside the particle to obtain the texture fractal dimension that characterizes the texture complexity of the particle surface.

[0025] The box size is set from 1 pixel to 1 / 10 of the smallest outer circle radius of the particle, and the fractal dimension of the texture region is calculated. Specifically, the internal region of the particle is first segmented from the extracted particle outline, the outline edge is removed, and the focus is on the fine texture of the particle surface (such as scratches, pits, etc. formed during the crushing process). The box dimension method is used for fractal calculation. By setting square boxes of different sizes to cover the texture region inside the particle, the number of boxes containing texture pixels under each box size is counted. The box size is set from 1 pixel to 1 / 10 of the smallest outer circle radius of the particle, which takes into account both texture detail and calculation efficiency, avoiding the loss of texture details due to the box size being too large, or the excessive amount of calculation and increased error due to the box size being too small. The fractal dimension of the texture region is calculated by fitting double logarithmic coordinates. Its value is controlled between 1.2 and 1.8. The larger the fractal dimension, the more complex and irregular the texture of the particle surface, and vice versa.

[0026] S113. The core angularity index and the texture fractal dimension are fused to obtain the grain shape feature.

[0027] The min-max normalization method is used to map the two types of feature values ​​to the 0-1 interval, eliminating the influence of dimensions and ensuring the fairness of the fusion.

[0028] Based on the influence mechanism of the mechanical properties of manufactured sand concrete, the influence of angular features on the interfacial bond strength and compressive strength of concrete is greater than that of texture features. Therefore, the weight of the core angularity index is set to 0.6 and the weight of the texture fractal dimension is set to 0.4. The two types of features are fused by weighted summation. The resulting particle shape feature vector contains information on the density and sharpness of particle angularity, as well as information on the complexity of particle surface texture. This achieves full-dimensional and accurate quantification of the particle shape features of manufactured sand, providing core particle shape feature support for subsequent microstructure modeling and mechanical property prediction.

[0029] In this embodiment, obtaining the gradation characteristics of manufactured sand includes the following steps: S121. Obtain the cumulative sieve residue data and stone powder content data of the manufactured sand.

[0030] The cumulative sieve residue data must be obtained using standard sand and gravel sieving test methods. The standard sieve aperture specifications follow the requirements for construction sand, and are 4.75mm, 2.36mm, 1.18mm, 0.6mm, 0.3mm, 0.15mm, and 0.075mm respectively. Each sieve grade requires three parallel sieve tests, and the cumulative sieve residue value is taken as the average of the three tests, with the error controlled within ±2% to avoid the influence of single test errors on the accuracy of the gradation characteristics. The actual measured value of stone powder content is determined using the water washing method. The manufactured sand sample... The mixture is placed in a sand washing container, and an appropriate amount of water is added and stirred thoroughly to separate the stone powder (particles with a diameter <0.075mm) from the sand particles. After filtration through a 0.075mm standard sieve and drying, the remaining stone powder is weighed, and the stone powder content is calculated (stone powder mass / total mass of manufactured sand × 100%). Three parallel tests are conducted, and the average value is taken as the final measured value of stone powder content to ensure that the stone powder content data can truly reflect the actual situation of the manufactured sand and provide accurate data support for the subsequent construction of stone powder sensitive gradation characteristics.

[0031] S122. Calculate the basic fineness modulus based on the cumulative sieve residue data, and correct the basic fineness modulus based on the stone powder content data to obtain the stone powder sensitive corrected fineness modulus.

[0032] Calculate the foundation fineness modulus according to the standard formula in GB / T14684-2011 "Construction Sand". Then, multiple sets of manufactured sand samples with different stone powder contents were selected, covering the common range of 0%-15%, and evenly distributed in a gradient of 2% (e.g., 0%, 2%, 4%...14%, 15%) to ensure that the samples can fully cover the fluctuation of stone powder content in actual manufactured sand production. Each set of samples used the same crushing process and raw materials to avoid interference from other factors (such as aggregate hardness and crushing method) on the test results. Each set of samples underwent three parallel sieving tests, strictly following the standard sieving process. The cumulative sieve residue data of each sieve grade (4.75mm to 0.075mm) was recorded in detail, and the actual stone powder content of the corresponding samples was also recorded to establish a correlation dataset of "stone powder content - cumulative sieve residue of each sieve grade".

[0033] With the "influence coefficient of stone powder content on concrete strength" as the optimization objective (calibrated through previous small-batch concrete test blocks), random forest regression was used to optimize the weight of each sieve grade (especially the 0.075mm sieve grade).

[0034] Specifically, the influence coefficient was determined through preliminary small-batch concrete test blocks: concrete test blocks were prepared using manufactured sand with different stone powder contents. The mix proportion parameters, such as water-cement ratio and sand ratio, were kept constant. After 28 days of standard curing, the compressive strength of the test blocks was measured. The ratio of the concrete strength under different stone powder contents to the benchmark strength (the strength when the stone powder content is 0%) was calculated. This ratio is the influence coefficient of stone powder content on concrete strength. The smaller the influence coefficient, the greater the negative impact of stone powder on concrete strength. Subsequently, a random forest regression algorithm was used, with the cumulative sieve residue of each sieve grade as the input feature and the influence coefficient of stone powder content on concrete strength as the output target. The regression model was trained, and the weights of each sieve grade were optimized through model iteration. The weight of the 0.075mm sieve grade was given special emphasis (because this sieve grade directly corresponds to stone powder particles and has the strongest correlation with stone powder content). This ensured that the weight allocation accurately reflected the degree of correlation between each sieve grade and the influence of stone powder, providing a scientific basis for subsequent fineness modulus correction.

[0035] Furthermore, based on the stone powder content data, the basic fineness modulus is corrected to obtain a stone powder-sensitive corrected fineness modulus, which satisfies: , Indicates the sensitive correction fineness modulus of stone powder. This is a correction factor determined based on the stone powder content data. For the first Optimization weights for each sieve stage For the first The cumulative percentage of residue for each sieve gradation. This represents the total number of sieve grades. The value can be linearly or non-linearly corrected based on the stone powder content. It can be obtained by optimizing a large number of samples through algorithms such as random forest, so that the generated gradation features can more sensitively reflect the influence of stone powder.

[0036] In one embodiment, the stone powder correction factor The design aligns with the actual characteristics of manufactured sand. When the stone powder content is ≤5%, the stone powder can act as a micro-aggregate filler with minimal impact on concrete performance. Therefore, a linear correction formula is adopted. For stone powder content, the correction factor increases by 0.08 for every 1% increase in stone powder content, appropriately amplifying the impact of stone powder on the fineness modulus. When the stone powder content exceeds 5%, excessive stone powder leads to decreased workability and strength of concrete. At this point, linear correction can no longer accurately reflect the negative impact of stone powder, and a nonlinear correction method is required. Specifically, a correction factor is introduced into the original linear formula. ( ),Right now This makes the correction coefficient increase non-linearly with the increase of stone powder content, which is more in line with the actual gradation change law when stone powder is in excess.

[0037] S123. The gradation characteristics are generated by combining the stone powder sensitive correction fineness modulus and the sieve residue ratio of each sieve grade.

[0038] First, the cumulative sieve residue data for each sieve grade is normalized, mapping the cumulative sieve residue value of each sieve grade to the 0-1 interval, eliminating the dimensional differences in the cumulative sieve residue of different sieve grades, and ensuring the balanced weight of the characteristics of each sieve grade. Then, the sieve residue ratio of each sieve grade is calculated, that is, the ratio of the cumulative sieve residue value of a certain sieve grade to the sum of the cumulative sieve residues of all sieve grades, which represents the proportion of particles of that sieve grade in the manufactured sand. Finally, the corrected fineness modulus and the normalized sieve residue ratio of each sieve grade are concatenated in sequence to form a "stone powder sensitive gradation feature vector". This feature vector contains both the coarseness information of the manufactured sand and accurately incorporates the influence of stone powder content, which can comprehensively and accurately characterize the gradation characteristics of the manufactured sand, providing core gradation feature support for subsequent microstructure modeling and mechanical property prediction.

[0039] In another embodiment, obtaining the hydration potential characteristics of a cementitious material includes the following steps: S131. Based on the raw material measurement data, the measurement data is mapped to a high-dimensional space through kernel principal component analysis in order to capture the nonlinear interaction between raw materials.

[0040] First, normalize and process missing values ​​for the raw material measurement data, including cement (type, dosage), water (dosage), admixtures (type, dosage), and fly ash / mineral powder (dosage), to ensure that the data format is consistent and there are no anomalies.

[0041] Then, kernel principal component analysis (KPCA) is used to map the low-dimensional raw material measurement data to a high-dimensional space to capture the nonlinear interactions between raw materials (such as the interaction term of cement dosage × admixture dosage).

[0042] Specifically, dimensions are determined based on the type of raw materials (such as five core dimensions: cement usage, water usage, admixture dosage, fly ash usage, and mineral powder usage, as well as categorical variables such as cement type and admixture type); subsequently, a Gaussian kernel function is selected as the kernel function for KPCA, and the kernel parameters are... The Gaussian kernel function, set to 0.8-1.2 (optimized through 5-fold cross-validation to ensure optimal mapping), effectively captures the nonlinear relationships between raw materials, avoiding the limitation of linear kernel functions in representing complex interactions. KPCA maps low-dimensional data to a 15-20 dimensional high-dimensional space. Each feature in the high-dimensional space corresponds to a nonlinear combination of low-dimensional raw material data, encompassing not only the dosage information of individual raw materials but also, more importantly, capturing the interactions between them. Examples include the product of cement dosage and admixture dosage (reflecting the regulatory effect of admixtures on cement hydration), the ratio of water dosage to total cementitious material dosage (reflecting the influence of the water-cement ratio on hydration), and the interaction between fly ash / mineral powder dosage and cement dosage (reflecting the synergistic / inhibitory effect of admixtures on cement hydration), achieving a comprehensive quantification of raw material interactions.

[0043] S132. Based on the measured data of hydration heat, features strongly correlated with the hydration process are selected from the high-dimensional spatial features. The selected features are then subjected to dimensionality reduction processing to obtain the hydration potential features.

[0044] Based on measured data of heat of hydration, the Pearson correlation coefficient between high-dimensional spatial features and heat of hydration was calculated. Features with a correlation coefficient > 0.7 were selected, and redundant high-dimensional spatial features that were irrelevant to the hydration potential were eliminated to ensure that the retained features could accurately characterize the hydration properties of cementitious materials.

[0045] In this embodiment, two key parameters were selected: the cumulative heat of hydration over 72 hours and the peak rate of heat of hydration. These two parameters reflect the total degree and intensity of the hydration reaction, respectively, and can comprehensively characterize the hydration potential of cementitious materials. Subsequently, the Pearson correlation coefficient between each feature in the high-dimensional space and these two heat of hydration indicators was calculated. The Pearson correlation coefficient ranges from -1 to 1. The larger the absolute value, the stronger the correlation between the feature and the heat of hydration. A positive correlation indicates that the heat of hydration increases when the feature increases (such as features related to cement dosage), while a negative correlation indicates that the heat of hydration decreases when the feature increases (such as features related to retarder admixtures). A correlation coefficient threshold of >0.7 was set, and high-dimensional space features with a correlation coefficient exceeding this threshold for any heat of hydration indicator were selected. Features with a correlation coefficient ≤0.7 were removed (these features have a weak correlation with the hydration potential and are redundant features). This ensures that the selected high-dimensional space features are closely related to the hydration process of cementitious materials, providing an accurate feature basis for the subsequent generation of hydration potential features.

[0046] Furthermore, the selected high-dimensional spatial features are reduced to 5 dimensions using PCA to serve as the "hydration potential feature vector of cementitious materials" to characterize the degree of hydration, hydration rate, and distribution of hydration products.

[0047] Specifically, the selected high-dimensional spatial features are first standardized (re-normalized to the 0-1 interval) to eliminate dimensional differences between different high-dimensional spatial features. Then, the covariance matrix of the high-dimensional spatial features is calculated, and eigenvalues ​​and eigenvectors are obtained through eigenvalue decomposition. The top 5 principal components with eigenvalues ​​greater than 1 (cumulative contribution rate ≥ 90%) are selected to ensure that more than 90% of the core information in the high-dimensional spatial features is retained after dimensionality reduction, avoiding information loss. These 5 principal components are named and defined, corresponding to the 5 core dimensions of hydration potential: the first... The first principal component characterizes the total degree of hydration (correlated with the cumulative heat of hydration over 72 hours), the second principal component characterizes the hydration rate (correlated with the peak rate of heat of hydration), the third principal component characterizes the moderating effect of admixtures (correlated with the interaction between admixtures and cement), the fourth principal component characterizes the synergistic effect of admixtures (correlated with the interaction between fly ash / mineral powder and cement), and the fifth principal component characterizes the effect of the water-cement ratio (correlated with the interaction between water usage and total amount of cementitious materials). This results in a 5-dimensional "hydration potential characteristic vector of cementitious materials," which comprehensively and accurately characterizes the hydration properties of cementitious materials.

[0048] S2. The intrinsic fusion characteristics of raw materials are obtained by fusing the particle shape characteristics, the gradation characteristics and the hydration potential characteristics through a graph neural network.

[0049] In this embodiment, the step of fusing the particle shape feature, the gradation feature, and the hydration potential feature through a graph neural network to obtain the intrinsic fused features of the raw materials includes the following steps: S21. Construct a heterogeneous graph using the particle shape feature, the gradation feature, and the hydration potential feature, wherein the particle shape feature, the gradation feature, and the hydration potential feature are respectively used as nodes, and the edges between nodes are established based on the physical relationship between the features and assigned initial weights.

[0050] First, using "particle shape characteristics," "gradation characteristics," and "hydration potential characteristics" as nodes, each node embeds the core information of its corresponding feature vector (e.g., particle shape feature nodes embed the nuclear angularity index and texture fractal dimension; gradation feature nodes embed the corrected fineness modulus and stone powder sieve residue ratio; hydration potential feature nodes embed the eigenvalues ​​of the five principal components), ensuring that each node can fully represent the core information of its corresponding feature. The construction of edges is strictly based on the physical mechanism of the performance formation of manufactured sand concrete, specifically including three types of core associations: First, there is the directional correlation between particle shape and gradation. The angularity of manufactured sand particles (a core parameter of particle shape characteristics) directly affects the aggregate bulk density. The sharper the angularity and the denser the distribution, the larger the gaps when the aggregate is packed, which in turn changes the rationality of the gradation. For example, when the core angularity index increases from 0.3 to 0.8, the aggregate bulk density decreases by 8%-12%. This correlation was calibrated through previous aggregate packing tests.

[0051] Second, there is a directional correlation between gradation and hydration. The stone powder content in the gradation (a core parameter of stone powder-sensitive gradation characteristics) directly affects the hydration rate. As a micro-aggregate, an appropriate amount of stone powder (≤5%) can promote cement hydration, while an excessive amount of stone powder (>5%) will coat cement particles and inhibit hydration. For example, when the stone powder content increases from 3% to 8%, the peak hydration heat rate decreases by 15%-20%.

[0052] Thirdly, there is a directional correlation between particle shape and hydration. The complexity of the aggregate surface texture (texture fractal dimension) affects the generation of interfacial hydration products. The more complex the texture, the larger the contact area between the aggregate and the mortar, and the richer the interfacial hydration products, which in turn improves the interfacial bond strength. For example, when the texture fractal dimension increases from 1.2 to 1.8, the content of interfacial hydration products increases by 25%-30%. All correlations are combined with previous experimental data and concrete mechanics principles to ensure that the construction of the edge has a clear physical meaning.

[0053] Furthermore, each edge is assigned a weight (based on the correlation determined in previous experiments, such as a weight of 0.4 for particle shape-hydration and a weight of 0.5 for gradation-hydration), with the weight value ranging from 0 to 1. This quantifies the strength of the correlation between each feature and provides a reasonable initial value for the subsequent dynamic optimization of the weights in the GAT model.

[0054] Based on extensive orthogonal experimental data from previous studies, the initial edge weights were determined by calculating the Pearson correlation coefficients of the two types of features and combining the influence weights of the features on the mechanical properties of concrete. In the example, the correlation between gradation features and hydration potential features was the strongest (stone powder content directly affects the hydration process, and gradation rationality affects the distribution of hydration products), therefore the edge weight for gradation → hydration was set to 0.5; the correlation between particle shape features and hydration potential features was the second strongest (texture and edges affect hydration through interfacial contact), therefore the edge weight for particle shape → hydration was set to 0.4; the correlation between particle shape features and gradation features was relatively weak (edges mainly affect bulk density and have limited impact on the overall gradation distribution), therefore the edge weight for particle shape → gradation was set to 0.3 (to complete the weight system and ensure the closed-loop logic of the heterogeneous diagram weights).

[0055] In other embodiments, to avoid excessive deviation in the initial weight setting, all initial weights can be normalized to ensure that the sum of the weights of the three edges is 1.2 (which is close to the actual proportion of the feature association strength). This lays the foundation for the subsequent GAT model to dynamically correct the weights through the attention mechanism, ensuring that the initial weights conform to the physical mechanism and have room for optimization.

[0056] S22. The heterogeneous graph is processed using a graph attention network. The weights of the edges are dynamically adjusted through the attention mechanism, and the fusion representation of the nodes is learned to obtain the intrinsic fusion features of the raw materials.

[0057] By adopting the GAT (Graph Attention Network) model and inputting heterogeneous graph data, the association weights between features are automatically learned through the attention mechanism (correcting the initial weights set manually), thereby achieving dynamic adaptation of the association relationships between features and improving the rationality and accuracy of feature fusion.

[0058] Specifically, the constructed heterogeneous graph data (node ​​features, edges, and initial weights) is input into the input layer of the GAT model. The node features are the feature vectors normalized in the early stage, and the edge weights are the initial weights set manually. Next, the core parameters of the GAT model are set: the number of attention heads is set to 4 (optimized through 5-fold cross-validation to balance feature learning accuracy and computational efficiency), the hidden layer dimension is set to 32, and the ReLU function is used as the activation function (to avoid gradient vanishing and improve the model's learning ability). During model training, the attention mechanism calculates the attention coefficient between each node and its neighboring nodes. This coefficient is based on the similarity of node features and the initial edge weights. The higher the similarity and the larger the initial weights, the larger the attention coefficients, thus dynamically adjusting the edge weights. For example, when the stone powder content in a batch of samples is high, the correlation strength between gradation and hydration is enhanced, and the model will automatically adjust the edge weight from 0.5 to 0.55-0.6, while the correlation strength between particle shape and gradation is weakened, and the weight is adjusted from 0.3 to 0.25-0.28, ensuring that the weight allocation can accurately match the actual feature correlation patterns of the samples and correct the limitations of manually set weights.

[0059] Furthermore, pooling operations are performed on the high-dimensional spatial features output by the GNN, using the Global Average Pooling (GAP) method. First, the high-dimensional spatial features output by the GAP model have a dimension of 32 (hidden layer dimension), containing the correlation information of the three types of features and their respective core feature information. Through global average pooling, the average value is taken for each feature channel of the high-dimensional spatial features, compressing the 32-dimensional high-dimensional spatial features to 10 dimensions, ensuring that the cumulative contribution rate of the features after dimensionality reduction is ≥92%, avoiding the loss of core correlation information. During the dimensionality reduction process, redundant feature dimensions with a variance less than 0.01 are simultaneously removed (these features have no obvious discriminative power and do not have a positive effect on subsequent modeling), further optimizing the feature quality, ultimately obtaining a 10-dimensional high-dimensional fusion feature. This feature not only contains the core information of the three types of features—granularity, gradation, and hydration potential—but also incorporates the dynamic correlation relationships among them.

[0060] In other embodiments, the high-dimensional fusion features after dimensionality reduction are normalized to obtain the "raw material intrinsic fusion feature vector". The final generated "raw material intrinsic fusion feature vector" is 10-dimensional, and each dimension corresponds to the correlation information or core parameters of the three types of features: particle shape, gradation and hydration potential. It can comprehensively and accurately characterize the intrinsic properties of the raw materials of manufactured sand concrete, which not only breaks through the limitations of a single feature, but also realizes the deep correlation and fusion of the three types of features.

[0061] S3. Use knowledge graphs to process the microstructural image features of real particle shape mapping to obtain stress cloud maps and damage evolution features. Combine the stress cloud maps and damage evolution features with the experimental real strength values ​​to obtain physical constraint gradient features.

[0062] In this embodiment, the step of using knowledge graphs to process the microstructural image features of real grain shape mapping to obtain stress cloud maps and damage evolution features includes the following steps: S311. Match the features of the microstructure model with the knowledge graph of concrete micromechanics to determine the mechanical parameters of each component phase of the microstructure.

[0063] Traditional random aggregate models only input particle size distribution, failing to reflect the particle shape characteristics and stone powder influence of manufactured sand. This results in significant differences between the generated aggregate and actual manufactured sand, thus affecting the accuracy of micromechanical simulations. This invention, based on the traditional model, adds three types of key feature parameters for input, enabling precise correction of the aggregate model and ensuring that the model closely matches the actual properties of manufactured sand concrete. First, the core angle index in the particle shape characteristics directly controls the sharpness of the aggregate edges. The larger the core angle index (approaching 1), the sharper and more numerous the aggregate edges will be, and vice versa. The specific correspondence is determined through the correlation test between the particle shape characteristics and the actual aggregate edges in the early stage, to ensure that the index and the actual edge state are accurately matched.

[0064] Secondly, the fractal dimension of the texture in the particle shape characteristics controls the complexity of the aggregate surface texture. The larger the fractal dimension (approaching 1.8), the denser the texture such as scratches and pits on the aggregate surface, and vice versa, the smoother the texture, which conforms to the surface texture differences formed during the crushing process of manufactured sand.

[0065] Thirdly, the stone powder correction coefficient in the gradation characteristics controls the filling state of fine aggregates. The larger the stone powder correction coefficient, the higher the stone powder content and the greater the filling density of fine aggregates. This adjusts the size of the aggregate gaps, providing a basis for the subsequent stone powder distribution model and achieving accurate matching between the aggregate model and the actual manufactured sand gradation and particle shape.

[0066] Furthermore, based on the stone powder-sensitive gradation characteristics and the aggregate particle size range in the mix proportion parameters, multiple continuous aggregate particle size sub-ranges (such as 0.075mm-0.3mm, 0.3mm-0.6mm, 0.6mm-1.18mm, 1.18mm-2.36mm, 2.36mm-4.75mm, 4.75mm-20mm) are divided. The aggregate proportion in each sub-range strictly follows the sieve residue proportion in the stone powder-sensitive gradation characteristics to ensure that the aggregate gradation is consistent with the actual manufactured sand. For the aggregate in each particle size sub-range, the idealized geometric shapes such as circles and ellipses used in traditional random aggregate models are abandoned. "Irregular polygonal aggregates" are generated based on particle shape fusion characteristics. The number of edges and corners is directly determined by the core angularity index. For every 0.1 increase in the core angularity index, the number of edges and corners of the aggregate increases by 2-3. At the same time, the sharpness of the edges and corners is controlled by the average curvature value in the core angularity index. The larger the curvature value, the sharper the edges and corners. The edge texture of the aggregate is controlled by the texture fractal dimension. Irregular edge textures that match the fractal dimension are generated through fractal algorithms. The higher the fractal dimension, the more complex and coarser the edge texture. This perfectly restores the irregular particle shape characteristics of manufactured sand caused by the crushing process, solving the core problem of the large difference between the "ideal aggregate" of the traditional model and the actual manufactured sand.

[0067] Furthermore, based on the gradation characteristics sensitive to stone powder, the proportion of residue on a 0.075mm sieve was extracted to determine the total number of stone powder particles (number of stone powder particles = total aggregate quantity × proportion of stone powder residue × correction coefficient; the correction coefficient is adjusted according to the water-cement ratio of the mix proportion; the higher the water-cement ratio, the closer the correction coefficient is to 1.2, and vice versa). The particle size of the stone powder particles was controlled within the range of <0.075mm, and different particle size sub-intervals were divided according to a normal distribution (e.g., 0.02mm-0.04mm, 0.04mm, etc.). (0.06mm-0.06mm, 0.06mm-0.075mm) to ensure that the stone powder particle size distribution is close to reality; the distribution location is preferentially selected in the gaps between aggregates. At the same time, combined with the filling state of fine aggregates, the number of stone powder particles is appropriately increased in areas with dense aggregate gaps, and the number of stone powder particles is reduced in areas with uniform aggregate distribution and large gaps. This simulates the real state of stone powder filling the gaps between aggregates in actual manufactured sand concrete, avoids simulation deviations caused by uniform stone powder distribution, and provides a real stone powder distribution basis for subsequent micromechanical simulation.

[0068] Then, based on the concrete mix proportion parameters, and according to the sand ratio and water-cement ratio, the proportions and spatial distributions of the four phases—aggregate, mortar, interfacial transition zone (ITZ), and stone powder—in the microstructure are determined to ensure that the microstructure is consistent with the actual manufactured sand concrete. Specifically, the aggregate proportion is determined by the sand ratio and the total aggregate content in the mix proportion, with the coarse aggregate proportion controlled at 60%-70%, the fine aggregate proportion controlled at 30%-40%, and the stone powder proportion determined by the stone powder content in the stone powder-sensitive gradation characteristics (0%-15%). The mortar proportion is adjusted according to the total mix proportion allowance. The interfacial transition zone (ITZ), as the transition area between aggregate and mortar, has a thickness controlled within a certain range. (Adjusted according to the sharpness of the aggregate edges; the sharper the edges, the greater the ITZ thickness); In terms of spatial distribution, the aggregates are evenly distributed according to their gradation characteristics to avoid local aggregation. Mortar fills the gaps between the aggregates and stone powder, the ITZ coats the aggregate surface, and stone powder fills the gaps between the fine and coarse aggregates; The resolution of the microstructure image is set to ≥500×500 pixels to ensure that the detailed features of the four phases (such as ITZ thickness, stone powder distribution, aggregate texture, etc.) can be clearly presented. The image is processed with grayscale, and different phases are distinguished by different grayscale values ​​(aggregate grayscale value 80-120, mortar grayscale value 150-180, ITZ grayscale value 120-150, stone powder grayscale value 50-80) to facilitate subsequent image feature extraction. The generated microstructure image needs to be verified by preliminary experiments to ensure that the similarity with the actual microstructure of manufactured sand concrete is ≥90%.

[0069] Then, the detailed image is preprocessed, including grayscale conversion, denoising, and enhancement, to eliminate the interference of image noise on feature extraction. Subsequently, the gray-level co-occurrence matrix (GLCM) analysis method is used, with the gray level set to 256 levels, the distance factor set to 1-3 pixels, and the angle set to 0°, 45°, 90°, and 135°, to extract four core texture features: contrast, entropy, correlation, and energy.

[0070] Simultaneously, using image segmentation algorithms, the aggregate, ITZ, and stone powder regions are segmented separately, and the aggregate distribution uniformity, ITZ thickness mean, and stone powder aggregation degree are calculated. Finally, the four texture features and three distribution features are normalized (0-1 interval), and concatenated in the order of "texture features + distribution features" to form a 7-dimensional "microstructure image feature vector". This comprehensively and accurately represents the characteristics of the microstructure, providing core structural feature support for subsequent knowledge graph matching and micromechanical simulation.

[0071] The prior knowledge graph of concrete micromechanics is constructed based on a large amount of experimental data on the micromechanics of manufactured sand concrete and industry standards. It includes four entities: aggregate, mortar, interfacial transition zone (ITZ), and stone powder. Each entity has a clear relationship and attribute: the relationship between entities covers strength correlation, elastic modulus matching, bond force, and the influence of hydration degree, etc., while the entity attributes precisely correspond to the microstructural characteristics (such as aggregate particle size and angularity, stone powder content and aggregation degree, ITZ thickness, etc.). At the same time, it incorporates the unique mechanical laws of manufactured sand concrete (such as the influence of stone powder on mortar strength and the influence of aggregate angularity on ITZ bond force). The knowledge graph is stored in a graph database, which supports rapid retrieval and matching of entities, relationships, and attributes, ensuring that the subsequent matching process is efficient and accurate.

[0072] Based on this, specific parameters in the microstructure image features (such as aggregate edge angle, stone powder aggregation degree, and average ITZ thickness) are used as search keywords to retrieve corresponding entity attributes (such as the "edge angle" attribute of aggregate entities and the "aggregation degree" attribute of stone powder entities) in the knowledge graph. Secondly, based on the preset entity relationships and attribute association rules in the knowledge graph, the mechanical parameters of each phase are dynamically determined. For example, the elastic modulus of ITZ is negatively correlated with the aggregate edge angle. When the aggregate edge angle (after normalization) increases from 0.3 to 0.8, the elastic modulus of ITZ linearly decreases from 35 GPa to 22 GPa. This correlation was calibrated through a large number of micromechanical experiments in the early stage and entered into the knowledge graph. Furthermore, the higher the stone powder aggregation degree, the lower the mortar bonding force. When the stone powder aggregation degree (after normalization) is > 0.6, the mortar bonding force decreases by 15%-25%. This ensures that the matched mechanical parameters can accurately reflect the influence of microstructure features on mechanical properties, providing realistic parameter support for subsequent micromechanical simulations.

[0073] S312. Based on the mechanical parameters, perform finite element mechanical simulation on the microstructure model to obtain stress cloud diagrams and damage evolution curves.

[0074] Based on the dynamic mechanical parameters of each phase obtained by knowledge graph matching, the finite element method (FEM) is used to carry out micromechanical simulation, realizing the transformation of the geometric features of the microstructure into mechanical response features.

[0075] Specifically, the microstructure image is meshed using a quadrilateral structured mesh. Next, matched mechanical parameters (such as aggregate elastic modulus, mortar bond strength, ITZ strength, and stone powder influence coefficient) are assigned to the corresponding microphases. The aggregate elastic modulus is dynamically adjusted based on particle size and edge angle (larger particle size and higher edge angle result in a larger elastic modulus), while the stone powder influence is indirectly reflected by adjusting the mortar mechanical parameters. Subsequently, simulation boundary conditions and loading methods are set, employing uniaxial compression loading to simulate the actual stress process of concrete. During loading, stress changes and damage development in each mesh unit are recorded in real time. Finally, a stress cloud map and damage evolution curve are generated. The stress cloud map uses a color gradient to represent the stress magnitude at each location (red indicates high-stress areas, blue indicates low-stress areas), clearly showing the stress concentration locations (mostly at aggregate edges and ITZ regions). The damage evolution curve uses loading time as the horizontal axis and damage value as the vertical axis, accurately representing the damage development law of the concrete microstructure during loading (the complete process from initial microcrack formation to final failure).

[0076] S313. Extract at least one stress distribution feature from the stress cloud map, extract at least one damage development feature from the damage evolution curve, and fuse the stress distribution feature and the damage development feature to form the stress cloud map and damage evolution feature.

[0077] Extract at least one stress distribution feature from the generated stress cloud map to characterize the stress distribution state of the microstructure, including: First, the maximum principal stress, which is the maximum stress value that occurs in the microstructure during loading, reflects the stress bearing limit of the microstructure. The larger the maximum principal stress, the stronger the stress resistance of the microstructure. When extracting, abnormal extreme points need to be excluded and the maximum value of the remaining stress value should be taken. Second, the stress standard deviation characterizes the dispersion of stress values ​​of each grid element in the stress cloud diagram. The larger the standard deviation, the more uneven the stress distribution and the more obvious the stress concentration phenomenon. The statistical standard deviation of stress values ​​of all grid elements is used in the calculation to reflect the stress differences in each region of the microstructure. Third, the area ratio of stress concentration regions, which is the ratio of the area of ​​regions where the stress value exceeds 80% of the maximum principal stress to the total area of ​​the microstructure, characterizes the severity of stress concentration. The larger the ratio, the more areas in the microstructure are prone to damage. During extraction, stress concentration regions are identified through image segmentation algorithms, and their area ratio is accurately calculated.

[0078] At least one damage development feature is extracted from the damage evolution curve to characterize the damage development law of the microstructure during loading and reflect the damage resistance of the microstructure: The first is the initial damage threshold, which is the loading time and corresponding stress value when the microstructure begins to show microcracks and the damage value first exceeds 0.05. It characterizes the microstructure's resistance to initial damage. The larger the threshold, the more difficult it is for the microstructure to generate initial microcracks. Second, the damage development rate, i.e. the slope of the damage evolution curve, is calculated using a linear fitting method. The higher the rate, the faster the damage develops in the microstructure during loading and the weaker its damage resistance. When fitting, the damage value is selected in the range of 0.1 to 0.8 to ensure the accuracy of the fitting results. Thirdly, the limit damage value, which is the maximum damage value (range 0-1) when the microstructure fails, characterizes the microstructure's resistance to failure. The closer the limit damage value is to 1, the more complete the microstructure failure and the weaker the resistance to failure. When extracting the limit damage value, the damage value when the microstructure stress suddenly drops is used as the limit damage value.

[0079] The extracted stress cloud map features (maximum principal stress, stress standard deviation, and area ratio of stress concentration region) and damage evolution features (damage initial threshold, damage development rate, and ultimate damage value) are normalized respectively, and all feature values ​​are uniformly mapped to the 0-1 interval to eliminate the dimensional differences between different features. After normalization, the features are spliced ​​in the order of "stress cloud map features + damage evolution features" to form a 6-dimensional "stress cloud map-damage evolution feature vector". This feature vector contains both the stress distribution information of the microstructure and the damage development law, realizing the accurate transformation of the geometric features of the microstructure into mechanical response features.

[0080] In another embodiment, obtaining the physical constraint gradient characteristics by combining the stress cloud map with the damage evolution characteristics and the actual experimental strength values ​​includes the following steps: S321. Establish a regression model between the stress cloud diagram, damage evolution characteristics, and measured values ​​of concrete mechanical properties.

[0081] A preliminary regression model, such as XGBoost, is constructed. The input is a stress cloud map and damage evolution characteristics, and the output is the predicted strength value of the mechanical properties of manufactured sand concrete. A preliminary correlation between the micromechanical response characteristics and the macromechanical properties is established.

[0082] The model construction process requires parameter optimization to ensure good model fit and avoid overfitting: First, the input features and measured strength values ​​are divided into training and test sets in a 7:3 ratio. The training set is used for model training, and the test set is used for model performance validation. Second, the core parameters of the XGBoost model are set: the initial learning rate is set to 0.01, the number of decision trees is set to 100, the maximum tree depth is set to 5, and the minimum number of sample splits is set to 3. The parameters are optimized through 5-fold cross-validation to improve the model's goodness of fit (R²) on the training set. 2R ≥ 0.85 on the test set 2 A value ≥0.8 is used to ensure that the model can capture the correlation between features and intensity well. After training, all input features are substituted into the model to obtain the predicted intensity value for each sample.

[0083] S322. Calculate the SHAP value of each feature in the stress cloud map and damage evolution features for the prediction result of the regression model.

[0084] In this embodiment, the SHAP (Shapley Additive Interpretation) value is calculated using the TreeExplainer interpreter. During the calculation process, the SHAP value of each feature corresponding to each sample is retained, and the average value of all sample SHAP values ​​is taken as the global contribution of the feature. The positive or negative sign of the SHAP value directly represents the direction of the feature's influence on the strength. A positive SHAP value indicates that when the feature increases, the strength prediction value increases accordingly (e.g., the larger the maximum principal stress, the positive SHAP value, indicating that it has a positive contribution to the concrete strength). A negative SHAP value indicates that when the feature increases, the strength prediction value decreases accordingly (e.g., the larger the area ratio of the stress concentration region, the negative SHAP value, indicating that it has a negative contribution to the concrete strength).

[0085] S323. Based on the physical mechanism of concrete mechanics, the sign of the SHAP value of each feature is verified, and features that conform to physical laws are selected. The gradient of the SHAP value of each selected feature is used as the physical constraint gradient feature.

[0086] The specific verification process is based on the unique mechanical properties of manufactured sand concrete. For example, according to the principles of concrete mechanics, an increase in the area of ​​stress concentration regions leads to excessively high local stress, which easily generates microcracks and reduces the strength of concrete. Therefore, its SHAP value should be negative. If the SHAP value of this feature in a sample is positive and the absolute value is small, it is judged to violate physical laws. For another example, the larger the maximum principal stress, the stronger the stress resistance of the microstructure, and the higher the macroscopic strength of the concrete. Its SHAP value should be positive. If a negative SHAP value appears without a reasonable reason, it is judged to be abnormal. A threshold for the absolute value of SHAP value is set (e.g., ≥0.05). Redundant features with an absolute value of SHAP value <0.05 (small contribution) and a sign that violates physical laws are removed to ensure that the retained features have both significant contribution and conform to the physical mechanism of concrete mechanics.

[0087] For features retained after physical verification, calculate the gradient of their SHAP values, quantify the sensitivity of the features to concrete strength, and clarify the changing pattern of the strength prediction contribution when the features undergo small changes. Gradient calculation employs numerical differentiation. Specifically, the value of a single retained feature is finely adjusted by 1% within its normalized range (0-1 interval) (e.g., adjusting the feature value from 0.5 to 0.505), while keeping other feature values ​​unchanged. The adjusted SHAP value of the feature is then recalculated, and the difference between the two is the gradient value of that feature. The sign of the gradient value indicates the direction of the feature's influence on sensitivity. A positive gradient value means that for every 1% increase in the feature's value, its positive contribution to intensity increases (or its negative contribution decreases), while a negative gradient value means that for every 1% increase in the feature's value, its positive contribution to intensity decreases (or its negative contribution increases). The absolute value of the gradient value represents the strength of the influence on sensitivity; the larger the absolute value, the more sensitive the feature is to the intensity. For example, if the gradient value of the area ratio of the stress concentration region is -0.02, it means that for every 1% increase in the feature's value, its negative contribution to intensity increases by 0.02, indicating that the feature has a high sensitivity to negative influences on intensity.

[0088] The calculated gradient values ​​are categorized and organized according to their corresponding physical meanings, and then concatenated to form a "physical constraint gradient feature vector". The classification is based on the composition phases of the microstructure and the mechanical response mechanism, specifically divided into three categories: First, there are aggregate-related gradient characteristics, including characteristic gradients related to aggregate properties, such as the gradient value of the maximum principal stress, which is associated with the aggregate's bearing capacity. Second, stone powder-related gradient characteristics, including characteristic gradients related to stone powder distribution and content, such as the gradient value of stress standard deviation, which are related to the influence of stone powder on the uniformity of microstructure. Thirdly, there are ITZ-related gradient features, including feature gradients related to the characteristics of the interface transition zone, such as the gradient value of the damage development rate, which are associated with the bond strength of the ITZ.

[0089] After classification, the gradient values ​​of each type are normalized (mapped to the 0-1 interval) to eliminate the dimensional differences between different gradient values. They are then concatenated in the order of "aggregate-related + stone powder-related + ITZ-related" to form a "physical constraint gradient feature vector" with the same dimension as the number of retained features. This vector includes both the sensitivity of the feature to the strength and the physical constraints of concrete mechanics.

[0090] S4. Quantify the correlation between structural image features, micromechanical response features and concrete mechanical properties to obtain strongly correlated microstructural features.

[0091] In this embodiment, the correlation between the quantified structural image features, the micromechanical response features, and the mechanical properties of concrete is used to obtain strongly correlated microstructural features, which includes the following steps: S41. Calculate the signal-to-noise ratio of each feature in the microstructural model and stress cloud diagram and damage evolution features. The signal-to-noise ratio is the ratio of the absolute value of the correlation coefficient between the feature and the measured value of the concrete mechanical properties to the coefficient of variation of the corresponding feature.

[0092] For each input feature (13 dimensions including microstructural image features, stress cloud map, and damage evolution features), its "signal" and "noise" are calculated respectively. The effectiveness of the feature is quantified by the signal-to-noise ratio (SNR), providing a quantitative basis for subsequent feature selection.

[0093] In this embodiment, "signal" is defined as the absolute value of the Pearson correlation coefficient between the feature and the measured values ​​of concrete mechanical properties (compressive / flexural strength), ranging from 0 to 1. The closer the absolute value is to 1, the stronger the correlation between the feature and the mechanical properties, and the more significant the signal. During the calculation, the correlation coefficient between each feature and the compressive strength and flexural strength is calculated separately, and the maximum value of the two is taken as the "signal" value of the feature, ensuring that the degree of correlation between the feature and the mechanical properties can be fully reflected.

[0094] "Noise" is defined as the coefficient of variation of the feature, which is the ratio of the standard deviation of the feature to the mean. The value ranges from 0 to ∞. The smaller the coefficient of variation, the less dispersion of the feature data, the lower the noise, and the better the stability of the feature. When calculating, it is necessary to calculate the sample mean and standard deviation based on the feature value of all samples, and then obtain the noise value by "coefficient of variation = standard deviation / mean" to avoid the noise calculation bias caused by single sample data.

[0095] The signal-to-noise ratio (SNR) satisfies the following condition: SNR = signal / noise. The larger the SNR value, the stronger the signal and the lower the noise of the feature, the greater its contribution to the prediction of mechanical performance, and the more valuable it is to retain. If the SNR value is smaller, it indicates that the feature has too much noise or a weak correlation with mechanical performance, and is a redundant feature that needs to be removed later.

[0096] S42. Filter features with a signal-to-noise ratio greater than a preset threshold, and then splice and normalize the filtered features to obtain the strongly correlated microstructure features.

[0097] Based on the actual characteristics of the microstructure of manufactured sand concrete, a SNR threshold of ≥1.5 was determined through 5-fold cross-validation optimization. Features with SNR ≥ the threshold were selected, while redundant features with SNR < 1.5 were eliminated, ensuring that the retained features possess both strong correlation and low noise. The core basis for threshold setting is the noise characteristics of the microstructure of manufactured sand concrete. Microstructure image features are easily affected by imaging accuracy and segmentation errors, while microstructure mechanical response features are easily affected by small fluctuations in simulation parameters, and their noise is generally higher than that of ordinary concrete features. Through 5-fold cross-validation, different thresholds (1.0, 1.2, 1.5, 1.8, 2.0) were substituted into the feature selection, and the threshold corresponding to the highest model validation accuracy (1.5) was selected as the final selection criterion to ensure the rationality and generalization of the selection results.

[0098] The microstructure image features and micromechanical response features, after being filtered by the SNR threshold, are concatenated in the order of "microstructure image features + micromechanical response features" and normalized to obtain a "strongly correlated microstructure feature vector". This feature vector retains only the microstructure features that are strongly correlated with the mechanical properties of concrete and have low noise, which reduces the dimensionality and computational load of subsequent modeling, improves the feature quality, and effectively avoids the impact of microstructure feature noise on prediction accuracy.

[0099] S5. Based on the intrinsic fusion characteristics of the raw materials, the physical constraint gradient characteristics, and the strongly correlated microstructure characteristics, the mechanical properties of the manufactured sand concrete are predicted using a multimodal fusion physical information neural network.

[0100] In one embodiment, the method of predicting the mechanical properties of manufactured sand concrete using a multimodal fusion physical information neural network based on the intrinsic fusion characteristics of the raw materials, the physical constraint gradient characteristics, and the strongly correlated microstructural characteristics includes the following steps: S51. Construct a multimodal fusion physical information neural network. The loss function of the multimodal fusion physical information neural network includes a data loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the concrete mechanics constitutive equation or damage evolution equation.

[0101] A multimodal fusion physical information neural network is constructed, including: building three independent input branches, each receiving three types of features; setting an independent fully connected layer in each branch; and extracting feature information within each branch to achieve independent enhancement and preliminary extraction of features from different modalities. The three input branches correspond to the "physical constraint gradient feature branch," the "raw material intrinsic fusion feature branch," and the "strongly correlated mesostructure feature branch," respectively. The branch design is tailored to the dimensions and characteristics of each feature type, ensuring the targeted nature of feature extraction. First, there is a physical constraint gradient feature branch with an input dimension of 4-5. Two fully connected layers are set, with 16 neurons in the first layer and 8 neurons in the second layer. The activation function is the ReLU function. The focus is on strengthening the information related to "physical constraints" in the gradient features, highlighting the sensitivity of the features to the intensity, and avoiding the weakening of the physical meaning of the gradient features.

[0102] Second, the intrinsic fusion feature branch of raw materials has an input dimension of 10 and two fully connected layers. The first layer has 32 neurons and the second layer has 16 neurons. The activation function is ReLU. The focus is on extracting the correlation information between particle shape, gradation and hydration potential, and retaining the core details of the intrinsic characteristics of raw materials.

[0103] Third, the strongly correlated microstructural feature branch has an input dimension of 8-10 dimensions and two fully connected layers. The first layer has 24 neurons and the second layer has 12 neurons. The activation function is ReLU, which focuses on strengthening the strong correlation between microstructure features and mechanical properties and further suppressing residual noise.

[0104] Dropout layers are added to the fully connected layers of each branch to prevent overfitting. At the same time, the He initialization method is used to initialize the weights of the fully connected layers to ensure gradient stability during model training and avoid gradient vanishing or gradient exploding problems.

[0105] The outputs of the three branches are concatenated in the order of "physical constraint gradient feature branch output + raw material intrinsic fusion feature branch output + strongly correlated mesostructure feature branch output" to obtain a 24-dimensional fusion feature vector, which is then input into the fusion layer to achieve the initial integration of the three types of multimodal features.

[0106] During the splicing process, it is ensured that the dimensions of the output features of each branch are consistent (all have been normalized to 0-1) to avoid feature weight imbalance caused by differences in dimensions. At the same time, the spliced ​​fused feature vector is standardized (mean normalization, so that the feature mean is 0 and the standard deviation is 1) to further optimize the feature distribution, improve the feature learning efficiency of the fusion layer, and ensure that the fused features can comprehensively cover the three core types of information: physical constraints, raw material intrinsics, and mesostructure, so as to achieve the initial aggregation of multi-dimensional information.

[0107] By introducing concrete mechanical physical constraints (such as the compressive strength constitutive equation and the damage evolution equation) into the fusion layer, and using the physical equation as part of the loss function (PINN loss = data loss + physical loss), a dual constraint of "data-driven + physical mechanism" is achieved.

[0108] Specifically, considering the unique mechanical properties of manufactured sand concrete, two types of core physical equations are selected as constraints: First, the constitutive equation for the compressive strength of concrete adopts a damage-plastic constitutive model, and its expression is as follows: ,in To resist compressive stress, The initial elastic modulus, In response, As the damage variable, this equation is used to constrain the relationship between the compressive strength and strain predicted by the model, ensuring that it conforms to the laws of concrete compressive strength mechanics.

[0109] Second, the damage evolution equation adopts an exponential damage evolution model, expressed as follows: ,in and These are model parameters (calibrated through prior micromechanical experiments) used to constrain the model's fit to the damage development law, ensuring consistency with the actual micromechanical damage evolution process.

[0110] In the loss function design, the mean squared error (MSE) is used for data loss to measure the deviation between the model's predicted values ​​and the measured intensity values. The expression is: , These are measured values. For predicted values, The sample size is used; the physical loss is the sum of squares of the residuals from the physical equations, which measures the deviation between the model's predictions and the physical equations, and is expressed as follows: , For the physical equation residuals, The physical constraint is the number of samples; the final total loss of PINN is The weight , (Determined through 5-fold cross-validation optimization).

[0111] S52. Input the physical constraint gradient features, the intrinsic fusion features of the raw materials, and the strongly correlated microstructure features into the physical information neural network, and train the multimodal fusion physical information neural network using training samples.

[0112] In this embodiment, the Adam optimizer is used to train the model, and the learning rate is adjusted (initially 0.001, gradually decreasing) to optimize the model parameters, ensuring stable model convergence and high fitting accuracy, while avoiding overfitting and improving the model's generalization ability.

[0113] The input data is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for updating model parameters, the validation set is used to monitor overfitting and adjust hyperparameters during training, and the test set is used for final model performance evaluation. During training, the training and validation losses are monitored in real time to ensure that they decrease synchronously. When the training loss is below 0.001 and the validation loss is below 0.002, the model is considered to have converged. At this point, the model's goodness of fit (Rfit) is... 2 A value ≥ 0.92 is used to ensure the model has good fitting performance and generalization ability. Simultaneously, model parameters are periodically saved during training, and the model with the smallest loss on the validation set is selected as the final training model, providing reliable model support for subsequent predictions.

[0114] S53. Input the features of the sample to be predicted into the trained physical information neural network to obtain the predicted mechanical properties of the manufactured sand concrete.

[0115] After training, three types of features of new samples (physical constraint gradient features, intrinsic fusion features of raw materials, and strongly correlated microstructure features) are input. After data preprocessing (normalization and outlier verification) consistent with the training process, the data is input into the trained multimodal fusion PINN model. The model extracts features through three independent input branches, and the fusion layer combines physical constraints to perform feature fusion and inference. Finally, it outputs the predicted mechanical properties (compressive / flexural strength) of manufactured sand concrete, achieving accurate and reliable performance prediction.

[0116] During the prediction process, the prediction results can be automatically corrected by combining physical constraint equations to avoid prediction values ​​that violate the laws of concrete mechanics (such as negative predicted strength or abnormal correlation between strength and microstructure). At the same time, the confidence interval of the prediction value (95% confidence level) is output to evaluate the reliability of the prediction results. The smaller the confidence interval, the higher the prediction accuracy.

[0117] After the prediction is completed, error analysis is performed on the prediction results to calculate the relative error between the predicted value and the subsequent measured value. The relative error is ensured to be ≤5%. If the error exceeds the threshold, the feature data and model parameters of the new sample need to be re-verified to further optimize the prediction accuracy.

[0118] It should be noted that the specific implementation methods mentioned above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in the memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art according to actual needs.

[0119] Please see Figure 2In an embodiment, to efficiently execute the mechanical property prediction method for manufactured sand concrete provided by the present invention, the present invention also provides a mechanical property prediction system for manufactured sand concrete, comprising: an input device 1, an output device 2, a processor 3, and a memory 4, wherein the input device 1, output device 2, processor 3, and memory 4 are interconnected, and the memory 4 stores program instructions for executing the steps of the mechanical property prediction method for manufactured sand concrete. The mechanical property prediction system for manufactured sand concrete of the present invention has a compact structure and stable performance, and can stably execute the mechanical property prediction method for manufactured sand concrete of the present invention, further improving the overall applicability and practical application capability of the present invention.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A method for predicting the mechanical properties of manufactured sand concrete, characterized in that, Includes the following steps: To obtain the particle shape characteristics, gradation characteristics, and hydration potential characteristics of manufactured sand and cementitious materials; The intrinsic fusion characteristics of raw materials are obtained by fusing the particle shape characteristics, the gradation characteristics, and the hydration potential characteristics using a graph neural network. By using knowledge graphs to process the microstructural image features of real particle shape mapping, stress cloud maps and damage evolution features are obtained. The physical constraint gradient features are then obtained by combining the stress cloud maps and damage evolution features with the actual experimental strength values. The correlation between quantitative structural image features, micromechanical response features and concrete mechanical properties is obtained to obtain strongly correlated microstructural features; Based on the intrinsic fusion characteristics of the raw materials, the physical constraint gradient characteristics, and the strongly correlated microstructure characteristics, a multimodal fusion physical information neural network is used to predict the mechanical properties of machined sand concrete.

2. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The process of obtaining the particle shape characteristics of manufactured sand includes the following steps: Extract particle contours from particle images of manufactured sand, calculate the curvature of pixels in the particle contours, select pixels with curvature values ​​greater than a preset curvature threshold as corner points, and calculate the core angularity index based on the number of corner points and the average curvature. Box-dimensional fractal calculations are performed on the texture regions inside the particles to obtain the texture fractal dimension, which characterizes the texture complexity of the particle surface. By fusing the nuclear angularity index and the texture fractal dimension, the grain shape feature is obtained.

3. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, Obtaining the gradation characteristics of manufactured sand includes the following steps: Obtain the cumulative sieve residue data and stone powder content data of manufactured sand; The basic fineness modulus is calculated based on the cumulative sieve residue data. Based on the stone powder content data, the basic fineness modulus is corrected to obtain the stone powder sensitive corrected fineness modulus. The gradation characteristics are generated by combining the stone powder sensitive correction fineness modulus and the sieve residue ratio of each sieve grade.

4. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, Obtaining the hydration potential characteristics of cementitious materials includes the following steps: Based on raw material measurement data, the measurement data is mapped to a high-dimensional space through kernel principal component analysis in order to capture the nonlinear interactions between raw materials. Based on the measured data of hydration heat, features strongly correlated with the hydration process are selected from the high-dimensional spatial features. The selected features are then subjected to dimensionality reduction processing to obtain the hydration potential features.

5. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The method of fusing the particle shape feature, the gradation feature, and the hydration potential feature through a graph neural network to obtain the intrinsic fused features of the raw materials includes the following steps: A heterogeneous graph is constructed using the particle shape feature, the gradation feature, and the hydration potential feature, wherein the particle shape feature, the gradation feature, and the hydration potential feature are respectively used as nodes, and the edges between nodes are established based on the physical relationship between the features and assigned initial weights. The heterogeneous graph is processed using a graph attention network. The weights of the edges are dynamically adjusted through the attention mechanism, and the fusion representation of the nodes is learned to obtain the intrinsic fusion features of the raw materials.

6. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The method of using knowledge graphs to process the microstructural image features of real grain shape mapping to obtain stress cloud maps and damage evolution features includes the following steps: The characteristics of the microstructure model are matched with the knowledge graph of concrete micromechanics to determine the mechanical parameters of each component phase of the microstructure. Based on the mechanical parameters, a finite element mechanical simulation was performed on the microstructure model to obtain stress cloud diagrams and damage evolution curves; At least one stress distribution feature is extracted from the stress cloud map, and at least one damage development feature is extracted from the damage evolution curve. The stress distribution feature and the damage development feature are then combined to form the stress cloud map and the damage evolution feature.

7. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The process of obtaining the physical constraint gradient characteristics by combining the stress cloud map, damage evolution characteristics, and experimental true intensity values ​​includes the following steps: Establish a regression model between the stress cloud diagram, damage evolution characteristics, and measured values ​​of concrete mechanical properties; Calculate the SHAP value of each feature in the stress cloud map and damage evolution characteristics for the prediction result of the regression model; Based on the physical mechanism of concrete mechanics, the sign of the SHAP value of each feature is verified, and features that conform to physical laws are selected. The gradient of the SHAP value of each selected feature is used as the physical constraint gradient feature.

8. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The correlation between the quantified structural image features, micromechanical response features, and concrete mechanical properties is used to obtain strongly correlated microstructural features, including the following steps: Calculate the signal-to-noise ratio of each feature in the microstructural model and stress cloud map and damage evolution features. The signal-to-noise ratio is the ratio of the absolute value of the correlation coefficient between the feature and the measured value of concrete mechanical properties to the coefficient of variation of the corresponding feature. Features with a signal-to-noise ratio greater than a preset threshold are selected, and the selected features are spliced ​​and normalized to obtain the strongly correlated microstructural features.

9. The method for predicting the mechanical properties of manufactured sand concrete according to claim 1, characterized in that, The method for predicting the mechanical properties of manufactured sand concrete using a multimodal fusion physical information neural network based on the intrinsic fusion characteristics of the raw materials, the physical constraint gradient characteristics, and the strongly correlated microstructural characteristics includes the following steps: A multimodal fusion physical information neural network is constructed. The loss function of the multimodal fusion physical information neural network includes a data loss term and a physical constraint loss term. The physical constraint loss term is constructed based on the concrete mechanics constitutive equation or the damage evolution equation. The physical constraint gradient features, the intrinsic fusion features of the raw materials, and the strongly correlated microstructure features are respectively input into the physical information neural network, and the multimodal fusion physical information neural network is trained using training samples; The features of the sample to be predicted are input into the trained physical information neural network to obtain the predicted mechanical properties of the manufactured sand concrete.

10. A system for predicting the mechanical properties of manufactured sand concrete, characterized in that, The mechanical property prediction system for manufactured sand concrete includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions, which are used to execute the mechanical property prediction method for manufactured sand concrete according to any one of claims 1-9.

Images