A cement microstructure design method based on neural network and multi-scale feature constraints

Abstract

The present invention discloses a cement microstructure design method based on neural network and multi-scale feature constraints, including obtaining a two-dimensional image of cement microstructure and establishing a cement microstructure data set; obtaining a hyperparameter group of random steps according to a predetermined number of constraint scales, training a neural network model according to constraint rules and the hyperparameter group, and obtaining a microstructure feature reconstruction operator; establishing a reconstruction algorithm according to the constraint rules, testing the neural network model, selecting and saving a neural network model that meets the preset model; and using a reconstruction algorithm and a microstructure feature reconstruction operator to obtain a cement microstructure after reconstruction with a predetermined number of constraint scales. The present invention uses a reconstruction algorithm and a microstructure feature reconstruction operator to reconstruct cement microstructures, and supports flexible customization of phase distribution in cement microstructures, breaking the technical barriers of neural network methods in the synthesis of the initial state of cement hydration models, significantly saving time and labor costs in the cement manufacturing industry, and reducing carbon emissions.

Description

Technical Field

[0001] The invention relates to the technical field of cement hydration and neural network, and in particular to a cement microstructure design method based on neural network and multi-scale feature constraints. Background Art

[0002] As an important cementitious material, cement plays an extremely important role in the national economic life. In the construction industry, the study of cement hydration is of great significance. Understanding the mechanism of cement hydration process helps to determine the performance of cement by controlling the conditions of cement hydration, thereby optimizing the quality and service life of building materials. Taking the compressive strength of cement as an example, the national standard requires that cement samples be subjected to a compressive test after 28 days of curing. If one wants to study the relationship between the cement hydration process and the compressive strength, the traditional experimental process requires companies to bear the time and labor costs of cement during the 28-day hydration process, and increase carbon emissions.

[0003] In the prior art, there are effective methods for representing the cement hydration process by establishing a physical model of cement hydration. However, these methods still need to receive the cement microstructure as the initial state of evolution as input. Currently, there are several methods for obtaining the initial state.

[0004] Method 1: Scanning real sample representation through scanning technology.

[0005] Although this method is the most intuitive, it still requires grinding, calcining, and particle grading of cement clinker to obtain a suitable particle size ratio, followed by 28 days of curing and vacuum drying, which incurs labor costs, time costs, and carbon emissions.

[0006] For example: The patent with application number 202211502136.3 and invention name "A method for non-destructive quantitative characterization of three-dimensional microstructure of porous materials" uses CT equipment to non-destructively obtain the microstructure characterization of porous materials. It is clearly mentioned in the embodiment that cement slurry with a specific water-cement ratio is required, and the sample is scanned and characterized after 28 days of curing and vacuum drying.

[0007] Method 2: Traditional initial state synthesis method.

[0008] This approach usually uses simple geometries or written algorithms to simulate the phases in the cement microstructure. Although it allows for flexible customization of the distribution of phases in the microstructure, this approach presents characteristic deviations in the representation of the cement microstructure of cement samples.

[0009] For example, the patent application number is 201611160205.1, and the invention name is "A method for characterizing the three-dimensional pore structure characteristics of cement hydration". The method is based on the CEMHYD3D hydration model software. On the basis of the backscattered image, the three algorithms of "internal erosion method", "combustion algorithm" and "random walk method" are used to characterize the three-dimensional pore structure characteristics of cement hydration. However, this method has some limitations. It not only requires the input of energy spectrum images, but also the three algorithms rely on a large number of human-defined rules. The microstructures generated by these rules do not necessarily match the characteristics of the actual cement microstructure in terms of characteristics.

[0010] Method three: neural network method.

[0011] In recent years, deep learning technology has been widely used in various fields due to its powerful feature learning ability. This type of method usually uses a neural network to learn the features present in the cement scanned samples and establish a potential model of the cement microstructure. The cement microstructure reconstructed by the potential model is more consistent with the cement sample in terms of features.

[0012] Although existing neural network methods can more accurately reconstruct cement microstructures through learned features, they still have difficulty in flexibly customizing the distribution of phases in the microstructure like traditional initial state synthesis methods. This is because existing methods focus on constructing cement sample features into a potential model, while ignoring the formal changes in the process of reconstructing the potential model features into cement microstructures. This limitation makes the phase distribution in cement microstructures reconstructed by neural network methods uncontrolled, hindering the research progress of cement hydration.

[0013] Therefore, the prior art lacks application for designing the initial state microstructure of cement hydration models. Summary of the invention

[0014] In order to solve the above technical problems, the present invention proposes the following technical solutions:

[0015] The embodiment of the present invention provides a cement microstructure design method based on neural network and multi-scale feature constraints, comprising:

[0016] S1, obtain the two-dimensional image of cement microstructure and establish the cement microstructure data set;

[0017] S2, obtaining a hyperparameter group of random steps according to the predetermined number of constraint scales, training a neural network model according to the constraint rules and the hyperparameter group, and obtaining a microstructure feature reconstruction operator;

[0018] S3, establishing a reconstruction algorithm according to the constraint rules, testing the neural network model, selecting a neural network model that meets the preset model and saving it;

[0019] S4, using a reconstruction algorithm and a microstructure feature reconstruction operator, the cement microstructure is obtained after a predetermined number of constrained scale reconstructions.

[0020] Furthermore, in step S1, a two-dimensional image of cement microstructure is obtained to establish a cement microstructure data set, including:

[0021] S11, acquiring a two-dimensional image of cement microstructure to form an original image set;

[0022] S12, randomly selecting a cement microstructure two-dimensional image from the original image set, and randomly cutting out a first image inside the cement microstructure two-dimensional image;

[0023] S13, repeating step S12 to form a first image set with all the cut-out first images.

[0024] Further, in step S2, a hyperparameter group of random steps is obtained according to a predetermined number of constraint scales, and a neural network model is trained according to the constraint rules and the hyperparameter group to obtain a microstructure feature reconstruction operator, including:

[0025] S21, design a neural network model as a microstructure feature reconstruction operator;

[0026] S22, selecting a set of random steps, and obtaining a hyperparameter set of the random steps according to a predetermined number of constraint scales;

[0027] S23, arbitrarily selecting a first image in the first image set, then selecting a random step value at a corresponding position in the hyperparameter group, and then sequentially applying constraints from 1 to the random step value to the first image to obtain a constrained second image, and averaging the entire second image to obtain a third image;

[0028] S24, repeatedly performing the operations in steps S22-S23, forming all the second images into a second image set, and forming all the third images into a third image set;

[0029] S25, sending the third image set and the random step number to the microstructure feature reconstruction operator as input and embedding, respectively, to obtain a fourth image set;

[0030] S26, using back propagation to calculate the distance between the fourth image set and the second image set, and training a microstructure feature reconstruction operator;

[0031] S27, repeat steps S22-S26 until the neural network model training is completed.

[0032] Furthermore, in step S3, a reconstruction algorithm is established according to the constraint rules, including:

[0033] S31, initialize the three-dimensional noise cube;

[0034] S32, firstly, perform the first reconstruction step, separate the three-dimensional noise cube into three first sections in orthogonal directions, apply the microstructure feature reconstruction operator to each first section and then reconstruct it into a first combined three-dimensional body, apply the first After the step constraint, the first combined three-dimensional body is divided into first cubes with a fourth predetermined side length;

[0035] S33, performing the second reconstruction step, separating the first cube into three second cross sections in orthogonal directions, applying a microstructure feature reconstruction operator to each second cross section and then recombining them into a second combined three-dimensional body, applying the first After the step constraint, the second combined three-dimensional body is divided into second cubes with a fifth predetermined side length;

[0036] S34, taking the second cube as the cube to be separated in the next step, repeating step S33 until the last step, without applying constraints, to obtain a hypothetical reconstructed three-dimensional cement microstructure.

[0037] Further, in step S4, a reconstruction algorithm and a microstructure feature reconstruction operator are used to obtain a three-dimensional cement microstructure after a predetermined number of constrained scale reconstructions, including:

[0038] S41, the grass formed by the user drawing the cement microstructure phase distribution at different steps Figure 3 Dimensional body as a constraint;

[0039] S42, using the constraint conditions in step S41 to calibrate the neural network model, combining the reconstruction algorithm and the microstructure feature reconstruction operator to obtain a reconstructed three-dimensional cement microstructure;

[0040] S43, observing the reconstructed three-dimensional cement microstructure and choosing whether to retain it.

[0041] Furthermore, the method is executed using the following system, including: a neural network model training module, a reconstruction algorithm establishment module, a neural network model testing module and a cement microstructure reconstruction module, wherein the neural network model training module is used to obtain a microstructure feature reconstruction operator based on a cement microstructure data set and constraint rules, the reconstruction algorithm establishment module is used to establish a reconstruction algorithm based on the constraint rules, the neural network model testing module is used to test the neural network model based on the reconstruction algorithm, and the cement microstructure reconstruction module is used to reconstruct the cement microstructure using the reconstruction algorithm and the microstructure feature reconstruction operator.

[0042] Furthermore, establishing the cement microstructure data set includes:

[0043] S11, acquiring a two-dimensional image of cement microstructure to form an original image set;

[0044] S12, randomly selecting a cement microstructure two-dimensional image from the original image set, and randomly cutting out a first image inside the cement microstructure two-dimensional image;

[0045] S13, repeating step S12 to form a first image set with all the cut-out first images.

[0046] Furthermore, the neural network model training module obtains a microstructure feature reconstruction operator according to the cement microstructure data set and constraint rules, including:

[0047] S21, design a neural network model as a microstructure feature reconstruction operator;

[0048] S22, selecting a set of random steps, and obtaining a hyperparameter set of the random steps according to a predetermined number of constraint scales;

[0049] S23, arbitrarily selecting a first image in the first image set, then selecting a random step value at a corresponding position in the hyperparameter group, and then sequentially applying constraints from 1 to the random step value to the first image to obtain a constrained second image, and averaging the entire second image to obtain a third image;

[0050] S24, repeatedly performing the operations in steps S22-S23, forming all the second images into a second image set, and forming all the third images into a third image set;

[0051] S25, sending the third image set and the random step number to the microstructure feature reconstruction operator as input and embedding, respectively, to obtain a fourth image set;

[0052] S26, using back propagation to calculate the distance between the fourth image set and the second image set, and training a microstructure feature reconstruction operator;

[0053] S27, repeat steps S22-S26 until the neural network model training is completed.

[0054] Furthermore, the reconstruction algorithm establishment module establishes a reconstruction algorithm according to the constraint rules, including:

[0055] S31, initialize the three-dimensional noise cube;

[0056] S32, firstly, perform the first reconstruction step, separate the three-dimensional noise cube into three first sections in orthogonal directions, apply the microstructure feature reconstruction operator to each first section and then reconstruct it into a first combined three-dimensional body, apply the first After the step constraint, the first combined three-dimensional body is divided into first cubes with a fourth predetermined side length;

[0057] S33, performing the second reconstruction step, separating the first cube into three second cross sections in orthogonal directions, applying a microstructure feature reconstruction operator to each second cross section and then recombining them into a second combined three-dimensional body, applying the first After the step constraint, the second combined three-dimensional body is divided into second cubes with a fifth predetermined side length;

[0058] S34, taking the second cube as the cube to be separated in the next step, repeating step S33 until the last step, without applying constraints, to obtain a hypothetical reconstructed three-dimensional cement microstructure.

[0059] Furthermore, the cement microstructure reconstruction module uses a reconstruction algorithm and a microstructure feature reconstruction operator to reconstruct the cement microstructure, including:

[0060] S41, the grass formed by the user drawing the cement microstructure phase distribution at different steps Figure 3 Dimensional body as a constraint;

[0061] S42, using the constraint conditions in step S41 to calibrate the neural network model, combining the reconstruction algorithm and the microstructure feature reconstruction operator to obtain a reconstructed three-dimensional cement microstructure;

[0062] S43, observing the reconstructed three-dimensional cement microstructure and choosing whether to retain it.

[0063] Beneficial effects:

[0064] (1) The cement microstructure design method based on neural network and multi-scale feature constraints proposed in the embodiment of the present invention uses a reconstruction algorithm and a microstructure feature reconstruction operator to reconstruct the cement microstructure and supports flexible customization of the phase distribution in the cement microstructure, breaking the technical barriers of the neural network method in the initial state synthesis of the cement hydration model, significantly saving time and labor costs in the cement manufacturing industry and helping to reduce carbon emissions.

[0065] (2) The cement microstructure design method based on neural network and multi-scale feature constraints proposed in the embodiment of the present invention effectively realizes the content customization from macro to micro at different particle sizes. The cement microstructure obtained by the neural network model faithfully restores the original characteristics of cement, making the user-customized initial hydration state available and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0067] Figure 2 A system structure block diagram provided for an embodiment of the present invention;

[0068] Figure 3Part of the data used for training a neural network provided by an embodiment of the present invention;

[0069] Figure 4 This is an example of a customized result provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The present solution is described below in conjunction with the accompanying drawings and specific implementation methods.

[0071] The embodiment of the present invention proposes a cement microstructure design method based on neural network and multi-scale feature constraints, which is oriented to the study of cement hydration mechanism and used for the initial state microstructure customization of cement hydration model. Features are captured in existing cement scanning sample images, and cement microstructure is obtained based on feature reconstruction to serve as the initial state input of the hydration model.

[0072] In order to make the method flexible and customizable, during the reconstruction process, the neural network does not directly reconstruct the cement microstructure from the input noise, but gradually reconstructs it under the constraints of rules through a multi-step process. Constraint rules include: in each step, the intermediate product is divided into different "supervoxels", that is, cubes composed of several pixels. All pixels in the "supervoxel" are required to represent the average color within the supervoxel at each step. As the method progresses, the side length of the cube decreases and the number of cubes increases. Users can flexibly customize the cement microstructure according to the constraint rules, such as replacing "supervoxels" on different constrained scales through sketches to achieve constraints of different strengths.

[0073] Figure 1 The method flow chart provided in the embodiment of the present invention includes the following steps:

[0074] S1, obtain the two-dimensional image of cement microstructure and establish the cement microstructure data set;

[0075] S2, obtaining a hyperparameter group of random steps according to the predetermined number of constraint scales, training a neural network model according to the constraint rules and the hyperparameter group, and obtaining a microstructure feature reconstruction operator;

[0076] S3, establishing a reconstruction algorithm according to the constraint rules, testing the neural network model, selecting a neural network model that meets the preset model and saving it;

[0077] S4, using a reconstruction algorithm and a microstructure feature reconstruction operator, the cement microstructure is obtained after a predetermined number of constrained scale reconstructions.

[0078] The two-dimensional image in the cement microstructure data set of the embodiment method of the present invention is obtained by grinding and grading cement clinker, mixing it with water, curing, drying and scanning equipment, or it can be an existing two-dimensional image of cement microstructure.

[0079] In step S1, the scanning device may be -CT or SEM, collects one or more two-dimensional images through scanning equipment, and the working parameters of the scanning equipment can be obtained through but not limited to the supporting software and manual of the working instrument.

[0080] In step S1, obtaining a two-dimensional image of cement microstructure and establishing a cement microstructure data set includes:

[0081] S11, acquiring a two-dimensional image of cement microstructure to form an original image set;

[0082] S12, randomly selecting a cement microstructure two-dimensional image from the original image set, and randomly cutting out a first image inside the cement microstructure two-dimensional image;

[0083] S13, repeating step S12 to form a first image set with all the cut-out first images.

[0084] In step S12, the side length of the cement microstructure two-dimensional image is set to a first length , the first image is set to have a side length of the first predetermined side length of squares, where .

[0085] In step S13, all the cut first images are grouped into a batch The first image set .

[0086] In step S2, a hyperparameter group of random steps is obtained according to the predetermined number of constraint scales, and a neural network model is trained according to the constraint rules and the hyperparameter group to obtain a microstructure feature reconstruction operator, including the following process:

[0087] S21, design a neural network model as a microstructure feature reconstruction operator;

[0088] S22, selecting a set of random steps, and obtaining a hyperparameter set of the random steps according to a predetermined number of constraint scales;

[0089] S23, arbitrarily selecting a first image in the first image set, then selecting a random step value at a corresponding position in the hyperparameter group, and then sequentially applying constraints from 1 to the random step value to the first image to obtain a constrained second image, and averaging the entire second image to obtain a third image;

[0090] S24, repeatedly performing the operations in steps S22-S23, forming all the second images into a second image set, and forming all the third images into a third image set;

[0091] S25, sending the third image set and the random step number to the microstructure feature reconstruction operator as input and embedding, respectively, to obtain a fourth image set;

[0092] S26, using back propagation to calculate the distance between the fourth image set and the second image set, and training a microstructure feature reconstruction operator;

[0093] S27, repeat steps S22-S26 until the neural network model training is completed.

[0094] When training a neural network model as a microstructure feature reconstruction operator, the number of required constraint scales is determined in advance, which is recorded as the predetermined number of constraint scales .

[0095] In step S21, the neural network model is used as a microstructure feature reconstruction operator, and the neural network model has an input and embed .

[0096] The microstructure feature reconstruction operator can be of different structures, such as embedded multi-layer perceptrons, convolutional neural networks, recurrent neural networks, transformers, and their permutations and / or combinations after adding additional connections. It is recommended to use a fully convolutional or a structure with similar properties to a fully convolutional as a neural network model to achieve better reconstruction quality and enable the microstructure feature reconstruction operator obtained after training to flexibly control the output size by changing the input size.

[0097] In step S22, a set of random steps is selected and recorded as , according to the number of pre-defined constraint scales , set the number of random steps Each value in Get a random number of steps between The hyperparameter group.

[0098] In step S23, a first image set is randomly selected The first image in , and then select the random step value of the corresponding position in the hyperparameter group, recorded as , then, for the first image Apply the first Step constraints, get the second image after constraints , and then for the second image Take the average of the whole area to get the third image , A solid color cube.

[0099] In step S23, the first image Apply the first Step constraints include:

[0100] S231, for the first image Apply the first step constraint and use the predetermined formula to transform the first image Dividing into a fifth image, setting the color value of each pixel in the fifth image to the average color of the pixels in the fifth image;

[0101] S232, applying the constraint of step 2 to the fifth image, merging the fifth image into a sixth image according to a predetermined combination rule, and setting the color value of each pixel in the sixth image to the average color of the pixels in the sixth image;

[0102] S233, taking the sixth image as the image to be constrained in the next step, and repeating step S232 until the sixth image is obtained. Second image after step constraint .

[0103] In step S231, the fifth image is set to the second predetermined side length of squares, where .

[0104] In step S232, the sixth image is set to be a square with a third predetermined side length.

[0105] In step S233, the side length is The sixth image is the The image where the constraint is applied in step 1; When the step constraint is used, the merged sixth image has a side length of Then, we change the side length to The sixth image is the The image to which the constraints are imposed.

[0106] In step S24, the second image set is recorded as , and the third image set is denoted as .

[0107] In step S25, the third image set and random steps They are respectively sent to the microstructure feature reconstruction operator as input and embed , get the fourth image set .

[0108] In step S3, a reconstruction algorithm is established according to the constraint rules, the neural network model is tested, and a neural network model that meets the preset model is selected and saved.

[0109] In step S3, establishing a reconstruction algorithm according to the constraint rules includes:

[0110] S31, initialize the three-dimensional noise cube;

[0111] S32, firstly, perform the first reconstruction step, separate the three-dimensional noise cube into three first sections in orthogonal directions, apply the microstructure feature reconstruction operator to each first section and then reconstruct it into a first combined three-dimensional body, apply the first After the step constraint, the first combined three-dimensional body is divided into first cubes with a fourth predetermined side length;

[0112] S33, performing the second reconstruction step, separating the first cube into three second cross sections in orthogonal directions, applying a microstructure feature reconstruction operator to each second cross section and then recombining them into a second combined three-dimensional body, applying the first After the step constraint, the second combined three-dimensional body is divided into second cubes with a fifth predetermined side length;

[0113] S34, taking the second cube as the cube to be separated in the next step, repeating step S33 until the last step, without applying constraints, to obtain a hypothetical reconstructed three-dimensional cement microstructure.

[0114] In step S32, the fourth predetermined side length of the first cube is set to ,in, In step S233 of Consistent.

[0115] In step S33, the previous step is recorded as Step 1 reconstruction, the second cube is used as the first After separation and reorganization, the Step constraint, divided into edge length of cube.

[0116] The neural network model trained in step S2 is combined with the reconstruction algorithm in step S31 to obtain a hypothetical reconstructed three-dimensional cement microstructure. Through expert observation, or evaluation using image similarity indicators, or analysis of cement porosity, connectivity, permeability and other parameters, the two-dimensional image in the cement microstructure data set and the cross-section of the hypothetical reconstructed three-dimensional cement microstructure are compared to determine whether the neural network model meets the preset model, and the neural network model that meets the preset model is selected and saved.

[0117] In step S4, a reconstruction algorithm and a microstructure feature reconstruction operator are used to obtain a three-dimensional cement microstructure after a predetermined number of constrained scale reconstructions, including:

[0118] S41, the grass formed by the user drawing the cement microstructure phase distribution at different steps Figure 3 Dimensional body as a constraint;

[0119] S42, using the constraint conditions in step S41 to calibrate the neural network model, combining the reconstruction algorithm and the microstructure feature reconstruction operator to obtain a reconstructed three-dimensional cement microstructure;

[0120] S43, observing the reconstructed three-dimensional cement microstructure and choosing whether to retain it.

[0121] In step S4, the user can select any number of steps as the random number of steps I, and impose constraints according to the constraint rules to achieve customization of the required cement microstructure.

[0122] In step S41, the grass Figure 3 The three-dimensional body can be an incomplete three-dimensional body, or some steps can be omitted. After testing, it is found that for macro customization that tends to be statistical, the grass should be Figure 3 The dimension is applied to the earlier steps, and the opposite is true for the details. It is not recommended to refactor in the last step, i.e. When applying a sketch, it is equivalent to directly modifying the 3D body.

[0123] In step S42, Before refactoring, use the Figure 3 The reconstruction algorithm will execute the constraints of this step according to the constraints by replacing the values ​​of the corresponding positions in the combined three-dimensional bodies (including the first combined three-dimensional body and the first combined three-dimensional body) in steps S32-S34.

[0124] Figure 2 The system structure block diagram provided for the embodiment of the present invention includes: a neural network model training module, a reconstruction algorithm establishment module, a neural network model testing module and a cement microstructure reconstruction module. The neural network model training module is used to obtain a microstructure feature reconstruction operator according to a cement microstructure data set and constraint rules. The reconstruction algorithm establishment module is used to establish a reconstruction algorithm according to the constraint rules. The neural network model testing module is used to test the neural network model according to the reconstruction algorithm. The cement microstructure reconstruction module is used to reconstruct the cement microstructure using the reconstruction algorithm and the microstructure feature reconstruction operator.

[0125] The establishment of cement microstructure dataset includes:

[0126] S11, acquiring a two-dimensional image of cement microstructure to form an original image set;

[0127] S12, randomly selecting a cement microstructure two-dimensional image from the original image set, and randomly cutting out a first image inside the cement microstructure two-dimensional image;

[0128] S13, repeating step S12 to form a first image set with all the cut-out first images.

[0129] The neural network model training module obtains the microstructure feature reconstruction operator based on the cement microstructure data set and constraint rules, including:

[0130] S21, design a neural network model as a microstructure feature reconstruction operator;

[0131] S22, selecting a set of random steps, and obtaining a hyperparameter set of the random steps according to a predetermined number of constraint scales;

[0132] S23, arbitrarily selecting a first image in the first image set, then selecting a random step value at a corresponding position in the hyperparameter group, and then sequentially applying constraints from 1 to the random step value to the first image to obtain a constrained second image, and averaging the entire second image to obtain a third image;

[0133] S24, repeatedly performing the operations in steps S22-S23, forming all the second images into a second image set, and forming all the third images into a third image set;

[0134] S25, sending the third image set and the random step number to the microstructure feature reconstruction operator as input and embedding, respectively, to obtain a fourth image set;

[0135] S26, using back propagation to calculate the distance between the fourth image set and the second image set, and training a microstructure feature reconstruction operator;

[0136] S27, repeat steps S22-S26 until the neural network model training is completed.

[0137] The reconstruction algorithm establishment module establishes the reconstruction algorithm according to the constraint rules, including:

[0138] S31, initialize the three-dimensional noise cube;

[0139] S32, firstly, perform the first reconstruction step, separate the three-dimensional noise cube into three first sections in orthogonal directions, apply the microstructure feature reconstruction operator to each first section and then reconstruct it into a first combined three-dimensional body, apply the first After the step constraint, the first combined three-dimensional body is divided into first cubes with a fourth predetermined side length;

[0140] S33, performing the second reconstruction step, separating the first cube into three second cross sections in orthogonal directions, applying a microstructure feature reconstruction operator to each second cross section and then recombining them into a second combined three-dimensional body, applying the first After the constraints, the second combined three-dimensional body is divided into second cubes of a fifth predetermined side length;

[0141] S34, taking the second cube as the cube to be separated in the next step, repeating step S33 until the last step, without applying constraints, to obtain a hypothetical reconstructed three-dimensional cement microstructure.

[0142] The cement microstructure reconstruction module uses the reconstruction algorithm and microstructure feature reconstruction operator to reconstruct the cement microstructure, including:

[0143] S41, the grass formed by the user drawing the cement microstructure phase distribution at different steps Figure 3 Dimensional body as a constraint;

[0144] S42, using the constraint conditions in step S41 to calibrate the neural network model, combining the reconstruction algorithm and the microstructure feature reconstruction operator to obtain a reconstructed three-dimensional cement microstructure;

[0145] S43, observing the reconstructed three-dimensional cement microstructure and choosing whether to retain it.

[0146] In the embodiment of the method and system of the present invention, in the synthesis embodiment of cement hydration initial state, the relevant parameters are set as follows.

[0147] In step S12, the side length of the cement microstructure two-dimensional image is , the side length of the first cropped image .

[0148] In step S13, all the cropped first images are grouped into batches The first image set , set the batch number is 1.

[0149] In step S2, a fully convolutional neural network with a U-Net structure is used as a microstructure feature reconstruction operator, and a predetermined number of constraint scales is set. ,in:

[0150] In step S21, sine-cosine position coding is used as the embedding method:

[0151] ;

[0152] in, represents the embedded value, is the embedding dimension of the microstructure feature reconstruction operator, and the embedded value is encoded by the function Then it is added to the output of the latent space of the neural network model.

[0153] In step S231, the first image Divide into side length The number of squares (i.e., the fifth image) is , the formula is as follows:

[0154] ;

[0155] satisfy:

[0156] ;

[0157] In step S26, the loss function is used Training microstructure feature reconstruction operator:

[0158] ;

[0159] In step S31, a 2D image in the cement microstructure data set is randomly selected, and the image color average is calculated, and the image color is used as the color of the initialized 3D noise cube. The side length of the 3D noise cube is set to , that is, the required size is The microstructure of cement.

[0160] In step S41, black spheres with a radius of 24 are set at the eight corners of the three-dimensional blank image, and the centers of the spheres are: (0,0,0), (255,0,0), (0,255,0), (0,0,255), (255,255,0), (0,255,255), (255,0,255), (255,255,255), and the three-dimensional image is used as the grass. Figure 3 The constraints of the dimension body act on the reconstruction process.

[0161] In step S43, it is observed whether the constraint is effective.

[0162] Finally, a sketch of the target cement microstructure is drawn and applied to the reconstruction process again, and the reconstruction results are input into the hydration model for study.

[0163] Figure 3 The partial data for training a neural network provided in an embodiment of the present invention is specifically a first image obtained from a first image set, and a portion of the image after being cropped. Figure 4 This is an example of a customized result provided by an embodiment of the present invention.

[0164] The cement microstructure design method based on neural network and multi-scale feature constraints proposed in the embodiment of the present invention sequentially constrains the features in the three-dimensional microstructure of cement at different scales to be consistent with the features in the cement sample or the constraints given by the user, thereby realizing the fine-grained customization of the initial state microstructure of the cement hydration model and significantly reducing the experimental time and resources required for studying cement hydration.

[0165] The cement microstructure design method based on neural network and multi-scale feature constraints proposed in the embodiment of the present invention has an optional scale that allows users to choose the strength of the constraint and customize the required content in the cement microstructure from macro to micro at different particle sizes.

[0166] The constraints imposed by the user at a larger scale are subsequently matched to the characteristics of a given cement sample at a smaller scale. The synthesized microstructure not only conforms to the user-customized content, but also faithfully restores the original characteristics of the cement, making the user-customized initial hydration state usable and reliable.

[0167] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

Claims

1. A cement microstructure design method based on neural network and multi-scale feature constraints, characterized in that: The method comprises: S1, obtain the two-dimensional image of cement microstructure and establish the cement microstructure data set; S2, obtaining a hyperparameter group of random steps according to the predetermined number of constraint scales, training a neural network model according to the constraint rules and the hyperparameter group, and obtaining a microstructure feature reconstruction operator; S3, establishing a reconstruction algorithm according to the constraint rules, testing the neural network model, selecting a neural network model that meets the preset model and saving it; S4, using a reconstruction algorithm and a microstructure feature reconstruction operator, the cement microstructure is obtained after a predetermined number of constrained scale reconstructions; In step S2, a hyperparameter group of random steps is obtained according to the predetermined number of constraint scales, and a neural network model is trained according to the constraint rules and the hyperparameter group to obtain a microstructure feature reconstruction operator, including: S21, design a neural network model as a microstructure feature reconstruction operator; S22, selecting a set of random steps, and obtaining a hyperparameter set of the random steps according to a predetermined number of constraint scales; S23, arbitrarily selecting a first image in the first image set, then selecting a random step value at a corresponding position in the hyperparameter group, and then sequentially applying constraints from 1 to the random step value to the first image to obtain a constrained second image, and averaging the entire second image to obtain a third image; S24, repeatedly performing the operations in steps S22-S23, forming all the second images into a second image set, and forming all the third images into a third image set; S25, sending the third image set and the random step number to the microstructure feature reconstruction operator as input and embedding, respectively, to obtain a fourth image set; S26, using back propagation to calculate the distance between the fourth image set and the second image set, and training a microstructure feature reconstruction operator; S27, repeating steps S22-S26 until the neural network model training is completed; In step S3, a reconstruction algorithm is established according to the constraint rules, including: S31, initialize the three-dimensional noise cube; S32, firstly, perform the first reconstruction step, separate the three-dimensional noise cube into three first sections in orthogonal directions, apply the microstructure feature reconstruction operator to each first section and then reconstruct it into a first combined three-dimensional body, apply the first After the step constraint, the first combined three-dimensional body is divided into first cubes with a fourth predetermined side length; S33, performing the second reconstruction step, separating the first cube into three second cross sections in orthogonal directions, applying a microstructure feature reconstruction operator to each second cross section and then recombining them into a second combined three-dimensional body, applying the first After the step constraint, the second combined three-dimensional body is divided into second cubes with a fifth predetermined side length; S34, taking the second cube as the cube to be separated in the next step, repeating step S33 until the last step, without applying constraints, to obtain a hypothetical reconstructed three-dimensional cement microstructure.

2. The cement microstructure design method based on neural network and multi-scale feature constraints according to claim 1 is characterized in that: In step S1, a two-dimensional image of cement microstructure is obtained to establish a cement microstructure data set, including: S11, acquiring a two-dimensional image of cement microstructure to form an original image set; S12, randomly selecting a cement microstructure two-dimensional image from the original image set, and randomly cutting out a first image inside the cement microstructure two-dimensional image; S13, repeating step S12 to form a first image set with all the cut-out first images.

3. The cement microstructure design method based on neural network and multi-scale feature constraints according to claim 1 is characterized in that: In step S4, a reconstruction algorithm and a microstructure feature reconstruction operator are used to obtain a three-dimensional cement microstructure after a predetermined number of constrained scale reconstructions, including: S41, a three-dimensional sketch of the cement microstructure phase distribution at different steps drawn by the user is used as a constraint condition; S42, using the constraint conditions in step S41 to calibrate the neural network model, combining the reconstruction algorithm and the microstructure feature reconstruction operator to obtain a reconstructed three-dimensional cement microstructure; S43, observing the reconstructed three-dimensional cement microstructure and choosing whether to retain it.

4. The cement microstructure design method based on neural network and multi-scale feature constraints according to any one of claims 1 to 3, characterized in that: The method is performed by the following system, which includes: a neural network model training module, a reconstruction algorithm establishment module, a neural network model testing module and a cement microstructure reconstruction module, wherein the neural network model training module is used to obtain a microstructure feature reconstruction operator according to a cement microstructure data set and constraint rules, the reconstruction algorithm establishment module is used to establish a reconstruction algorithm according to the constraint rules, the neural network model testing module is used to test the neural network model according to the reconstruction algorithm, and the cement microstructure reconstruction module is used to reconstruct the cement microstructure according to the reconstruction algorithm and the microstructure feature reconstruction operator.

Images