AI-driven cement hydration microstructure evolution method and online virtual experiment platform

Abstract

The present invention relates to the field of image generation technology, and specifically to an AI-driven cement hydration microstructure evolution method and an online virtual experiment platform. The method comprises: obtaining a scanned image of a cement sample, and extracting vertical features, horizontal features, and neighborhood features of the image through a model feature extraction module. After splicing the horizontal features with the neighborhood features, they are input into a gated mask convolution module together with the vertical features for feature fusion to generate new horizontal features, which are then cyclically input into the module with the neighborhood features, and finally the output of the model is obtained through an activation layer and a convolution layer. Based on the output loss, the gradient is calculated and the model parameters are optimized through back propagation, so as to train an intelligent evolution model that is close to reality. The virtual experiment platform adopts a Web architecture, which not only realizes the online real-time evolution simulation of cement hydration microstructures, but also supports functions such as performance estimation, providing experimenters with efficient and low-cost cement microstructure simulation evolution and performance evaluation tools.

Description

Technical Field

[0001] The present invention relates to the technical field of image generation, and specifically to an AI-driven cement hydration microstructure evolution method and an online virtual experiment platform. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The hydration reaction of cement is the process of chemical reaction between cement and water. In this process, hydration products are gradually formed and connected to each other, so that the cement paste gradually hardens and gains strength. The hydration process of cement is a long evolutionary process, which usually takes at least 28 days from the preparation of cement samples to the initial hydration reaction and then to the stabilization of its main properties. This long cycle consumes huge experimental costs and human resources. If the process is simulated by computer, it can reduce costs and improve efficiency.

[0004] However, the cement hydration reaction involves complex physical and chemical reactions. The current understanding of the cement hydration mechanism is not thorough, which makes the modeling of the cement hydration process challenging. Some existing simulation evolution models or software (such as HYMOSTRUC, CEMHYD3D, DuCOM, VCCTL, etc.) usually rely on empirical rules and assumptions under ideal conditions (such as HYMOSTRUC assumes that cement particles are standard spheres). Although the amount of simulation calculations is reduced, these simplifications also limit the accuracy of the model's description of the real cement microstructure. In addition, the model will also ignore the impact of environmental changes on cement under actual curing conditions, which will lead to poor adaptability of the model and make it difficult to accurately predict the cement microstructure evolution process under complex environments such as different formulations and different curing conditions.

[0005] Some simulation evolution models use deep learning technology in artificial intelligence (AI) to learn the evolution law of cement material microstructure from experimental data and achieve simulation by simulating the complex reaction process of the material. However, these AI models cannot accurately model the local correlation in the evolution of cement microstructure. During the cement hydration process, the evolution of microstructure is affected by both local neighborhood information and global conditions. AI simulation evolution models using deep learning technology have obvious deficiencies in handling the fusion of local and global information. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides an AI-driven cement hydration microstructure evolution method and an online virtual experiment platform, and proposes a neighborhood condition gated network simulation evolution intelligent model. Through the neighborhood condition gated mechanism, the local consistency and global logical rationality are taken into account in the process of capturing the local structure of the image, so that the generated results have strong correlation and detail coherence in space, thereby more realistically reflecting the microstructure changes in the cement hydration process. Based on this evolution model, an online virtual experiment platform for cement based on Web architecture is constructed, which can be used for experiments such as evolution simulation of cement hydration microstructure. The platform also integrates other intelligent reasoning models, such as intelligent models such as cement compressive strength prediction, element distribution prediction, and microstructure generation based on deep learning technology. Based on these models, the platform can conduct virtual experiments online, thereby simplifying the experimental process, shortening the experimental time and reducing the complexity of operation, so that the experimenter can quickly iterate the experimental scheme in a virtual environment to improve research efficiency. The platform supports browser access, and the experimenter does not need expensive computing equipment, and can conduct experiments anytime and anywhere through a browser. The virtual experiment platform provides an efficient and low-cost solution for experimental research.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] The first aspect of the present invention provides an AI-driven cement hydration microstructure evolution method, comprising the following steps:

[0009] Obtaining scanned images of cement samples and preprocessing them;

[0010] The preprocessed image is processed by vertical mask convolution, horizontal mask convolution and neighborhood mask convolution respectively to obtain the vertical features, horizontal features and neighborhood features of the image; among them, the horizontal mask convolution processes all pixels on the left side of the row where the current pixel is located, the vertical mask convolution processes all pixels above the current pixel, and the neighborhood mask convolution processes the surrounding area of ​​the current pixel;

[0011] The horizontal features and vertical features are used together as the input of the gated mask convolution module. After the new horizontal features are generated by feature fusion, they are concatenated and fused with the neighborhood features. The horizontal features of the cyclic gating module pass through the output module to obtain the simulated evolution output results that are consistent with the input image size.

[0012] Furthermore, the preprocessed image is passed through a vertical mask convolution module, a horizontal mask convolution module, and a neighborhood mask convolution module in the feature input module to obtain vertical features, horizontal features, and neighborhood features of the image, respectively.

[0013] Furthermore, the feature input module obtains the spliced ​​horizontal features and neighborhood features, and inputs them together with the vertical features into the recurrent gating module. The recurrent gating module has multiple gated masked convolution modules, each of which has a vertical masked convolution module and a horizontal masked convolution module. The gated masked convolution modules are called layer by layer in a cycle, and the receptive field range is controlled by using different expansion coefficients to obtain the spliced ​​horizontal features and neighborhood features.

[0014] Furthermore, the horizontal features of the recurrent gating module are passed through the output module to obtain the simulation evolution output results consistent with the input image size, including:

[0015] The input vertical feature v_stack is convolved through the vertical mask convolution module to generate the vertical stack output feature map v_stack_feat, which is divided into two parts along the channel dimension and multiplied by the corresponding activation function to generate the gated vertical feature output;

[0016] The spliced ​​horizontal features and neighborhood features are used as the updated horizontal features h_stack to input into the horizontal mask convolution module, and h_stack_feat is generated through convolution. The context information of v_stack_feat is introduced into h_stack_feat by adding it to the projected v_stack_feat. After feature block division and corresponding activation function and multiplication, the gated horizontal feature output is obtained.

[0017] The gated vertical feature output and horizontal feature output are the vertical and horizontal features updated by the gated mask convolution module.

[0018] Furthermore, the horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result consistent with the input image size, which also includes: the gated horizontal feature output is further convolved and superimposed with the input horizontal feature h_stack through channel compression to ensure the continuity of the features.

[0019] Furthermore, the horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result that is consistent with the input image size, and also include: convolving the neighborhood features through a neighborhood mask convolution layer to generate a new conditional feature map to help introduce contextual information into the output result.

[0020] Furthermore, the horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result consistent with the input image size, which also includes: in the vertical and horizontal features updated by the gated mask convolution module, the vertical features are returned to the gated mask convolution module, and the horizontal feature output is concatenated with the conditional features in the channel dimension to form a combined feature, which is returned to the gated mask convolution module.

[0021] Furthermore, the horizontal features of the cyclic gating module pass through the output module to obtain a simulated evolution output result consistent with the input image size, and also include: after the gated mask convolution module executes the cycle a set number of times, the output combined features are output as new horizontal features.

[0022] Furthermore, the horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result that is consistent with the input image size, and also include: the combined features are passed through the activation layer as a new horizontal feature map, and then the number of channels is changed through the convolution layer to generate a simulated evolution result that is consistent with the input image size.

[0023] The second aspect of the present invention provides an online virtual experiment platform for implementing the above-mentioned AI-driven cement hydration microstructure evolution method. The platform is oriented to various performance predictions and microstructure simulation evolution experiments of cement. The platform is based on a Web architecture and includes:

[0024] Visualization front-end, the visualization front-end consists of two parts: the user interaction part and the background management part. The user interaction part provides an interactive interface between the user and the system, supporting experimenters to upload cement formulas, cement microstructure image data, experimental data and initiate simulation evolution experiment requests. It is also responsible for displaying the dynamic results and simulation evolution result output during the experiment, and supports interactive result browsing and analysis. The background management part is used to receive user-uploaded data (such as image data, experimental data) and experimental requests, and is responsible for sending these requests to the simulation reasoning backend. At the same time, it is also responsible for saving user data, supporting task status management and returning results to the visualization front-end. The background part has permission management functions to ensure data security and user privacy protection;

[0025] The simulation inference backend is responsible for receiving and processing experimental requests such as simulation evolution from the visualization frontend. Its main functions include preprocessing the received experimental data, including steps such as image denoising, cropping, and standardization, as well as data formatting and normalization. The preprocessed data is ready for experiments such as simulation evolution calculations. It also includes assigning the corresponding intelligent model through a proxy server to perform the corresponding experiment. When conducting a simulation experiment on cement microstructure evolution, the simulation evolution model used is based on a gated pixel convolutional network, which can accurately simulate the formation process of cement hydration microstructure. The simulation evolution model performs inference calculations through efficient computing methods and generates simulation results. Finally, after the simulation calculation is completed, the inference backend transmits the simulation evolution results back to the visualization frontend, and the experimenter can view the simulation evolution results in real time for subsequent analysis.

[0026] Compared with the prior art, one or more of the above technical solutions have the following beneficial effects:

[0027] 1. Each pixel in the cement microstructure image represents the material state or material properties and local environment at that location. Although these pixels are discrete in the image, the physical or chemical states they reflect are continuous and interrelated in reality. As the distance increases, the mutual influence between pixels usually gradually weakens, and this influence can even be negligible at a longer distance. This phenomenon is particularly significant in the reaction environment of cement hydration microstructure. Therefore, the conditional probability relationship between adjacent pixels in the image is modeled pixel by pixel to capture the local dependence in the evolution of the microstructure. After the neighborhood information is extracted through convolution features and spliced ​​to each module output of the model in the form of channels, channel information fusion is performed to enhance the model's ability to capture the spatial correlation between adjacent pixels; the information flow is improved by designing mask convolution and introducing a gating mechanism, thereby reducing the influence of the mask convolution blind field, so as to more effectively learn the complex spatial dependencies of the image, and then learn the evolution law of the cement hydration microstructure.

[0028] 2. The intelligent model of cement hydration microstructure evolution obtained by this scheme moves the computing tasks to the server side, so that the virtual experimental platform formed reduces the dependence on local device performance, allowing users to run complex evolution tasks on conventional devices. At the same time, the virtual experimental platform is based on the Web architecture and integrates other intelligent reasoning models to realize a variety of predictive experiments (such as strength prediction, etc.), allowing experimenters to quickly iterate experimental schemes in a virtual environment, thereby improving research efficiency. The platform supports browser access, and experimenters do not need expensive computing equipment, but can conduct virtual experiments anytime and anywhere through a browser. The platform provides an efficient and low-cost solution for cement-related experimental research. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0030] Figure 1 is a schematic diagram of a neighborhood conditional gated pixel convolutional neural network model provided by one or more embodiments of the present invention;

[0031] Figure 2 Schematic diagram of a gated mask convolution module in a neighborhood conditional gated pixel convolutional neural network model provided by one or more embodiments of the present invention;

[0032] Figure 3 It is a result schematic diagram of a cement hydration microstructure evolution simulation intelligent model provided by one or more embodiments of the present invention;

[0033] Figure 4is a functional schematic diagram of a cement online virtual experiment platform provided by one or more embodiments of the present invention;

[0034] Figure 5 It is a schematic diagram of the architecture of the cement online virtual experiment platform provided by one or more embodiments of the present invention;

[0035] Figure 6 is a schematic diagram of setting cement curing parameters in a cement online virtual experiment platform provided by one or more embodiments of the present invention;

[0036] Figure 7 is a schematic diagram showing the hydration progress in the cement online virtual experiment platform provided by one or more embodiments of the present invention;

[0037] Figure 8 It is a schematic diagram showing the simulation results of cement hydration microstructure evolution experiment in the cement online virtual experiment platform provided by one or more embodiments of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0040] The following embodiments provide an AI-driven cement hydration microstructure evolution method and an online virtual experimental platform. Considering that cement hydration has significant differences in local hydration due to different local environments, the influence of neighborhood information is crucial. GPCNN (Gated Pixel-CNN, gated pixel convolutional neural network) only performs strong constraint association modeling on global pixel information, better integrating the feature information around the previous pixel into the network layer by layer, and further strengthening the influence of neighborhood information. Therefore, based on GPCNN, the NC-GPCNN model is designed based on the influence of neighborhood information caused by different local evolution of cement, namely, the Neighborhood-Conditioned GatedPixel Convolutional Neural Networks (NC-GPCNN). This operation not only expands the dimension of the feature, but also integrates local neighborhood information and global information, providing the model with more comprehensive information to predict the generation of the next pixel.

[0041] Through the neighborhood condition gating mechanism, the model takes into account both local consistency and global logical rationality in the process of capturing the local structure of the image, so that the generated results have strong spatial correlation and detail coherence, thereby more realistically reflecting the microstructural changes in the cement hydration process.

[0042] The online virtual experiment platform adopts a Web architecture and is mainly composed of the following three modules: visualization front-end, simulation reasoning back-end and intelligent model module (including intelligent models such as NC-GPCNN simulation evolution). Each module cooperates with each other to provide high-precision and high-efficiency real-time online simulation evolution experiment services for cement microstructure, which can more realistically reflect the microstructure changes in the cement hydration process.

[0043] Embodiment 1:

[0044] The AI-driven cement hydration microstructure evolution method includes the following steps:

[0045] Obtaining scanned images of cement samples and preprocessing them;

[0046] The preprocessed image uses vertical mask convolution, horizontal mask convolution and neighborhood mask convolution respectively to obtain the vertical features, horizontal features and neighborhood features corresponding to each pixel; among them, the horizontal mask convolution processes all pixels on the left side of the row where the current pixel is located, the vertical mask convolution processes all pixels above the current pixel, and the neighborhood mask convolution processes the surrounding area of ​​the current pixel;

[0047] The horizontal features and vertical features are used together as the input of the gated mask convolution module. After the new horizontal features are generated by feature fusion, they are concatenated and fused with the neighborhood features. The horizontal features of the cyclic gating module pass through the output module to obtain the simulated evolution output results that are consistent with the input image size.

[0048] Taking into account the impact of different types of silicate cement, the data set of this embodiment is composed of three different formulations of silicate cement. By adjusting the water-cement ratio of cement to water, four cement samples are configured and prepared and cured. Samples are taken out on a specified number of days, and then micro-CT is used to scan and image cement samples with different proportions at different ages and save the data. After scanning, the samples are cured and micro-CT imaging is performed again after curing to the specified age. Thus, time-series CT data of cement samples with different proportions at ages of 2, 3, 4, 5, 6, 7, 14, 21, and 28 days are obtained. The CT images of each cement sample are cropped to obtain the domain of interest and produce a data set. The data set is randomly divided into a training set and a test set at a ratio of 9:1, where the training set is used for model training and the test set is used to verify the effectiveness of the model.

[0049] The resulting image shows the microstructure of cement during hydration. Each "pixel" in the cement microstructure image represents the material state or material properties (e.g., chemical composition, density, porosity, etc.) and local environment (e.g., temperature, pH, ion concentration, etc.) at that location. Although these pixels are discrete on the image, the physical or chemical states they reflect are continuous and interrelated in reality. Therefore, the state evolution of the local area has spatial autocorrelation, which means that the state of each location depends not only on the direct influence of the current neighborhood, but also on the chain indirect influence. Such characteristics are consistent with some characteristics of Markov chains, so it is reasonable and effective to use spatial autoregressive modeling methods to capture this local dependence and continuity.

[0050] As the distance increases, the mutual influence between pixels usually gradually weakens, and this influence can even be ignored at a longer distance. This phenomenon is particularly significant in the reaction environment of cement hydration microstructure, because the influence range of chemical reactions and physical effects is usually limited. Considering the characteristic that this influence decreases with distance, an effective modeling strategy is introduced to use neighborhood information as a conditional constraint in the model. This method allows the model to accurately capture local interactions while reducing interference to non-neighborhood areas by assigning lower weights to distant areas. This strategy can help the model more accurately simulate the dynamic changes of the microstructure while maintaining the spatial autocorrelation characteristics.

[0051] The core design concept of the pixel convolutional network (Pixel-CNN) is to simulate the joint distribution of image pixels in an autoregressive manner. By constructing explicit associations between pixels in space, it is transformed into a sequential decision problem. Pixel-CNN can capture long-distance dependencies between pixels and learn complex image distributions. Causal convolution is used in Pixel-CNN. This convolution design ensures that future pixel information is not used when calculating the current pixel. However, causal convolution is implemented in a masked manner, so that only the pixels on the left and above can be used when reasoning about each pixel. This structural design prevents the pixel from using information on the right and below. This limits the model's ability to capture spatial relationships in images, and may cause the generated image to appear spatially unnatural in some parts.

[0052] To solve this problem, the Gated Pixel-CNN improves information flow by redesigning mask convolution and introducing a gating mechanism, thereby reducing the impact of mask convolution blind field. In order to more effectively learn and generate images with complex spatial dependencies, and compared with the original Pixel-CNN, it has improved training efficiency and capacity. However, for the evolution of cement hydration microstructure, the local environment has a huge impact on the evolution results. Its modeling information not only needs to come from the influence of global information, but also depends on the capture and modeling of local neighborhood information.

[0053] This embodiment uses a neighborhood conditional gated pixel convolutional network (NC-GPCNN). By extracting neighborhood information through convolutional features and splicing it to each module output of the Gated PixelCNN model in the form of channels, and using 1×1 convolution for channel information fusion, this strategy effectively enhances the model's ability to capture the spatial correlation between adjacent pixels. In this way, NC-GPCNN can take into account both local consistency and global logical rationality when capturing image structure, so that the generated results have strong spatial correlation and detail coherence, thereby significantly improving the model's ability to process image details and local structures.

[0054] Considering that cement evolution has significant differences in local hydration reactions due to different local environments, the influence of neighborhood information is crucial. However, GPCNN only models the global pixel information with strong constraints, which better integrates the feature information around the previous pixel into the network layer by layer, further strengthening the influence of neighborhood information. Therefore, based on GPCNN, the NC-GPCNN model is designed to capture the influence of neighborhood information on the local evolution of cement. This operation not only expands the dimension of the features, but also integrates local neighborhood information and global information, providing the model with more comprehensive information to predict the next pixel.

[0055] Considering that Gated Pixel-CNN only performs strong constraint association modeling on global pixel information and does not fully utilize the information of the neighborhood around the pixel, compared with the original Gated pixel-CNN model, this model integrates the information around the current pixel extracted from the neighborhood feature channel of the cement microstructure image into the network layer by layer, further strengthening the influence of the neighborhood information. In this way, the model can capture the detailed spatial relationship around each pixel, thereby capturing the evolution law of cement hydration microstructure. The input of the model is a cement CT microstructure image at a certain age, and the output of the model is a microstructure image at the next age. The model is trained using the above data set, the loss function mentioned in the embodiment is selected as the optimization target, the Adam optimizer is used to optimize the neural network parameters, and the gradient back-propagation algorithm is used to train the neighborhood conditional gated pixel convolutional network.

[0056] The model is mainly composed of three modules, namely, feature input module, recurrent gating module and output module. The model structure diagram is as follows: Figure 1 The specific implementation details of the model are as follows.

[0057] The feature input module consists of three parts: the vertical mask convolution (VerticalStackConv) module, the horizontal mask convolution (HorizontalStackConv) module and the neighborhood mask convolution module. The three modules process the initial input image x simultaneously.

[0058] The initial input x passes through three initial feature processing modules at the same time to obtain three feature maps, namely, the feature output v_stack obtained by vertical mask convolution, h_stack obtained by horizontal mask convolution, and c_stack obtained by neighborhood mask convolution. Then, the h_stack obtained by horizontal mask convolution and the c_stack obtained by neighborhood mask convolution are channel-joined, and then this output is convolved by 1x1 to obtain a new feature output h_stack. The above operations realize the extraction of initial features.

[0059] Among them, consistent with the implementation of Gated Pixel-CNN, horizontal mask convolution and vertical mask convolution are inherited from mask convolution. The implementation of mask convolution is to set some positions in the convolution kernel to 0 and other positions to 1, so as to ensure that when generating each pixel, it can only rely on the generated pixels on the left and above the current pixel. Mask convolution is divided into two categories. The mask convolution of category "A" does not contain the information of the current pixel. Assuming that the size of the convolution kernel is 3 (the size of the convolution kernel in this embodiment is 3), the mask convolution kernel of "category" A is , which is generally used in the first layer of the network, while the mask convolution of type "B" considers the information of the current pixel, and its convolution kernel is as follows , used in all layers after the first layer of the network. Since the central area of ​​the first layer has been masked, the pixel information of the central prediction area has not been leaked into the network, so the central area of ​​all subsequent layers does not need to be masked.

[0060] The horizontal mask convolution only processes all pixels to the left of the row where the current pixel is located. The convolution kernel of type "A" is , and the convolution kernel of type “B” is .

[0061] The vertical mask convolution only processes all pixels above the current pixel, and its "A" type convolution kernel is , and the convolution kernel of type “B” is , the combination of vertical and horizontal mask convolution can achieve the desired receptive field and avoid the occurrence of blind areas.

[0062] For neighborhood convolution, the neighborhood around the pixel is taken into account, and the "A" type convolution kernel is , and the convolution kernel of type “B” is .

[0063] In the recurrent gating module, GateMaskedConv is called layer by layer through the for layer in self.conv_layers loop, and the receptive field is gradually expanded through convolution operations with different dilation coefficients to extract local and global features. The following is the detailed implementation of each step:

[0064] self.conv_layers contains 7 GateMaskedConv (Gated Masked Convolution Module) layers, each GateMaskedConv consists of vertical masked convolution and horizontal masked convolution, aiming to capture the features in the vertical and horizontal directions. Each layer has a different expansion coefficient (1, 2, 4) to control the range of the receptive field.

[0065] The specific structure of the gated mask convolution module is as follows Figure 2 As shown in the figure, the input vertical feature v_stack is first convolved through the vertical mask convolution module to generate the output feature map v_stack_feat of the vertical stack. The output channel is 2*in_channels, which is used to distinguish the value channel and the gated channel. v_stack_feat is divided into two parts along the channel dimension (dim=1), v_val and v_gate, and the gated vertical output v_stack_out is generated by tanh and sigmoid activation respectively. At the same time, the input horizontal feature is processed by horizontal stack convolution, that is, the input horizontal feature h_stack is convolved to generate h_stack_feat. The h_stack_feat is added with the v_stack_feat projected by conv_vert_to_horiz (a convolution layer that maps 2*in_channels input channels to the same number of output channels), and the context information of v_stack_feat is introduced into h_stack_feat. The processed h_stack_feat is activated and multiplied by tanh and sigmoid to obtain the gated horizontal feature output. The gated horizontal stack features are further convolved through conv_vert_1x1 to compress the feature channels back to in_channels and superimpose them with the original horizontal features h_stack to ensure feature coherence. The outputs v_stack and h_stack of each layer of GateMaskedConv are the updated vertical and horizontal features, which enhance the feature expression layer by layer.

[0066] Neighborhood masked convolution (CurrentStackConv) processing: c_stack is convolved using the neighborhood masked convolution layer CurrentStackConv (mask type "B"). This operation generates a new conditional feature map to help introduce contextual information in the generation process.

[0067] Next, we can further perform channel splicing and channel compression, as shown in the following example. Figure 1 The convolution feature fusion module shown is implemented as follows: In the vertical and horizontal features updated by GateMaskedConv (gated mask convolution module), the vertical features are returned to GateMaskedConv (gated mask convolution module), and the horizontal features h_stack and conditional features c_stack are concatenated in the channel dimension to form a combined feature with a channel number of 2*channels. Finally, the number of concatenated feature channels is compressed back to GateMaskedConv (gated mask convolution module) through 1x1 convolution to ensure that the number of model parameters does not increase and the feature information integration is effective.

[0068] Output module: The horizontal feature map h_stack is processed by the ELU layer so that the negative area is smoothly mapped to a small negative value, enhancing the nonlinear expression ability of the model. The feature map after the ELU activation layer is passed to the 1x1 convolution layer conv_out, which compresses the number of channels from channels to out_channels (for example, 3 channels of RGB images) and generates an output feature map out with the same size as the original image. This step ensures that the final output of the model has the shape and number of channels of the target image.

[0069] The loss function of the gated pixel convolutional network is based on the principle of maximum likelihood estimation (MLE). Its core goal is to optimize the model parameters to maximize the probability that the model-generated image matches the training data. Specifically, the model adopts an autoregressive structure, that is, the generation of each pixel depends on the values ​​of all previous pixels. The definition of the loss function is based on the conditional probability of a single pixel value. The conditional probability of each pixel value depends on its previous pixel values, and the model provides a conditional probability distribution for each pixel. The loss function is designed to be the negative log likelihood loss (NLL), which is to negate the logarithm of all conditional probabilities and sum them to measure the degree of fit of the model to the true data distribution. By minimizing the negative log likelihood loss, the model can better learn the dependencies between pixels in the data distribution, thereby improving the quality of image generation. At the same time, neighborhood information is introduced to ensure that the model can effectively use neighborhood information for image generation. A method based on this pixel neighborhood information is also introduced. Here, It can be regarded as an additional input to enhance the model's understanding of the surrounding neighborhood information of the current pixel that needs to be predicted.

[0070] The loss function of this embodiment is as follows:

[0071] ;

[0072] in, 、 and Respectively represent , The CT image results at the time and the model prediction output, the distribution of the model prediction is and ， represents the model parameters, where and is a weighted coefficient used to balance the contribution of the two loss functions in the total loss. Represents the final loss function.

[0073] In this embodiment, the experimental results of the neighborhood conditional gated pixel convolutional network model (NC-GPCNN) are as follows: Figure 3 shown.

[0074] The NC-GPCNN simulation evolution model in the above process is based on the local modeling of the neighborhood condition gated pixel convolution network, which can generate simulation evolution results that are more consistent with the real cement microstructure without relying on traditional empirical rules. Through the neighborhood condition gating mechanism, local consistency and global logical rationality are taken into account in the process of capturing the local structure of the image, so that the generated results have strong correlation and detail coherence in space, thereby more realistically reflecting the microstructural changes in the cement hydration process. The principle of NC-GPCNN is well adapted to the local characteristics of cement microscale reactions. It captures the local dependence in the microstructural evolution process by modeling the conditional probability relationship between adjacent pixels in the image pixel by pixel.

[0075] Embodiment 2:

[0076] Online virtual experiment platform, including:

[0077] The visualization front end includes a user interaction part and a background management part. The user interaction part is used to interact with users and display experimental results such as simulation evolution. The background management is used to receive and save image data, experimental data, simulation and other experimental requests uploaded by users and send them to the simulation reasoning back end.

[0078] The simulation reasoning backend receives simulation evolution and other experimental requests from the visualization frontend, preprocesses the data of the experimental request, and returns the experimental results. It forwards the data to the available intelligent model through the proxy server to perform the corresponding tasks (such as simulation evolution, performance estimation, etc.), and sends the obtained experimental results to the frontend.

[0079] The visualization front-end is divided into two parts, namely the front-end responsible for user interaction and real-time display of simulation evolution and other experimental results, and the front-end back-end, which can receive data from the front-end and simulation evolution and other experimental requests, and also includes user management, model request, security control and database support. The user management module is responsible for the management of user permissions and information; the model request module is responsible for processing task requests such as simulation evolution submitted by users; the security module ensures the security of the system; the database module is used to store user data and experimental results.

[0080] The simulation reasoning backend is responsible for receiving data requests and experimental requests from the visualization frontend, preprocessing data, and managing experimental tasks. The reasoning backend includes a proxy server, which is used to process and forward experimental requests and result feedback such as simulation evolution. Experimental requests such as simulation evolution from the visualization frontend will be sent to the corresponding intelligent model end for processing through the proxy server, and the results will be returned to the visualization frontend after the experiment is completed. In addition, the simulation reasoning backend also includes a data server, a management server, a model scheduling server, and an authentication server to ensure the management and control of the experimental process.

[0081] In this embodiment, the visualization front end provides a convenient data upload interface and supports multiple data formats. Users can upload the initial cement microstructure image and experimental parameters. At the same time, the front-end interface provides a user-friendly experimental parameter configuration module, including the selection of cement curing conditions (temperature, humidity, etc.), which supports users to adjust the experimental conditions according to specific needs. During the simulation evolution experiment, users can view the simulation evolution results in real time through the front end. The visualization front end uses technologies such as WebGL and three.js to present 2D or 3D images, allowing users to intuitively observe the dynamic changes in the cement microstructure. Experimental data supports cloud storage, which is convenient for viewing and comparing historical records.

[0082] To realize the Web-based online cement virtual experiment platform, the visualization front end of this embodiment is a typical C / S architecture, such as Figure 4 and Figure 5 As shown in the figure, the front-end module is based on the Spring Boot framework and the Thymeleaf template engine, providing an intuitive user interface that supports user data input, parameter configuration, real-time experimental results display, and interactive feedback.

[0083] The backend of the front-end module is developed based on the Spring Boot framework to achieve efficient communication with the backend. SpringBoot is combined with the Thymeleaf template engine and bootstrap to build dynamic web pages, which improves page rendering speed and user interaction experience. To display the three-dimensional microstructure of cement hydration, Three.js is also used. It is a WebGL-based 3D engine running in the browser. This engine is used to create and display various three-dimensional objects. In addition, the ECharts visualization tool is also used to display dynamic and smooth charts. The front end communicates with the simulation reasoning backend through RESTful API and WebSocket to ensure real-time transmission and display of data.

[0084] like Figure 4 and Figure 5 As shown in the figure, the simulation reasoning backend receives and processes the raw data uploaded by the user, including data standardization and format conversion, to ensure that the input meets the requirements of the model. The backend uses Django's RESTful API interface to achieve efficient data transmission with the frontend. Each experimental task such as simulation evolution submitted by the user is carried out as an independent process, and the backend is scheduled according to the task priority to ensure efficient use of computing resources. In order to cope with complex experimental tasks, the simulation reasoning backend supports a distributed computing architecture, which can distribute tasks to multiple servers or GPU clusters for processing, improving the system's processing power and computing efficiency.

[0085] In the online cement virtual experiment platform, experimenters can upload initial cement data through the visual front-end interface, such as the phase composition and distribution of cement hydration, etc. These data will be uploaded to the background database of the platform's visual front-end. It supports three-dimensional tensor formats such as nrrd files and npz to meet different research needs. It supports data viewing and reuse. In order to allow users to easily adjust experimental parameters, the front-end provides a setting interface for cement curing parameters, including temperature, humidity, etc., through sliders, input boxes and other controls. When the user completes data uploading and parameter setting, he can click the "Start Hydration" button through the interface to start the simulation evolution experiment. Figure 6 and Figure 7 shown.

[0086] During the simulation evolution experiment, the visualization front-end can display the evolution of cement microstructure in real time. Users can observe the microstructural state of cement through 2D or 3D views, including changes in key parameters such as porosity and crack distribution. The front-end is based on graphics rendering technologies such as WebGL and three.js to achieve efficient image and data rendering. Users can rotate, zoom, and other operations to view the evolution of cement microstructure from multiple angles. In the three-dimensional microstructure display interface of cement, sections in the X, Y, and Z directions are displayed. Users can drag the slider to view any microstructure section, and there is a preview two-dimensional image of the section below. To help users understand the progress of the experiment, the front-end also provides real-time display of the simulation evolution progress bar, such as Figure 8 After the simulation reasoning backend completes data processing, the evolution results are returned to the visualization frontend, which updates the data in real time to ensure that users can see the latest simulation evolution results as soon as possible. Figure 8 The simulation evolution results are presented in the form of dynamic charts and data tables, and users can view different parameter changes in the view, such as cement strength growth curve, porosity change, etc.

[0087] To ensure the convenience and experience of user operation, the front-end interface adopts a responsive design that can adapt to the display of different resolutions and devices, ensuring smooth use on PCs, tablets and mobile phones. The simple and intuitive interface layout and color matching effectively improve the user's operating efficiency and experience.

[0088] The simulation reasoning backend is developed based on the Django framework. Django provides an efficient Web server architecture and task management functions to ensure the stability and scalability of the backend service. In this virtual experiment platform, Django is mainly responsible for receiving the initial cement data and maintenance parameter requests from the front end, calling the intelligent model for reasoning, and sending the reasoning results to the visualization front end for display. Django's asynchronous processing capabilities support high-concurrency task execution, enabling the simulation reasoning backend to simultaneously process experimental requests such as simulation evolution from multiple users, improving the response efficiency of the system.

[0089] After the user uploads the initial data and sets the simulation parameters, the simulation reasoning backend will preprocess the data to ensure that the data format matches the model input requirements. Data standardization and format conversion are automatically completed in this module. The communication between the simulation backend and the visualization frontend uses the RESTful API protocol to ensure the reliability and efficiency of data transmission. The backend server receives and responds to data requests from the visualization frontend through Django's REST interface, ensuring real-time result feedback.

[0090] The simulation reasoning backend uses a task scheduling mechanism to manage user requests for simulation evolution and other experiments. Each task is executed independently, and the Django framework schedules it according to task priority and system resources to ensure efficient task management. In order to cope with complex simulation evolution and other experimental calculations and large-scale data processing, the simulation reasoning backend supports a distributed computing architecture. The system can distribute simulation evolution and other experimental tasks to multiple servers or GPU clusters for processing, call intelligent simulation models for simulation evolution and other experiments, and ensure stability and fast response under high load. By integrating Django with the distributed computing platform, the backend can realize dynamic allocation of computing resources, improving the scalability and maintainability of the virtual experiment platform.

[0091] For the microstructure evolution intelligent simulation model, when the reasoning is completed, the simulation reasoning backend will send the evolution results of the microstructure back to the visualization frontend in the form of batch data for users to observe in real time. The result of each simulation evolution reasoning contains the dynamic change information of the microstructure (such as the distribution of reactants, the distribution of products, etc.), and these data are presented in the user interface in the form of charts and 3D visualization.

[0092] To ensure the security of simulation data, the simulation reasoning backend integrates user authentication and permission management modules. Each user's operations and data in the platform are strictly controlled to prevent unauthorized data access. The user management function provided by Django ensures data isolation and platform security, and can prevent simulation data from being leaked between different users.

[0093] The simulation reasoning backend module adopts modular design, following the principle of high cohesion and low coupling. Each functional module is independent of each other, which is convenient for subsequent maintenance and functional expansion. Through the scalability design of Django, the cement online virtual experiment platform can quickly adapt to the new cement hydration microstructure evolution simulation evolution intelligent model to ensure that it can adapt to changes in different experimental conditions and research needs.

[0094] Taking the online cement microstructure evolution simulation experiment as an example, the workflow of the platform is as follows:

[0095] Step 1: User data input; the experimenter uploads the initial image of cement microstructure and experimental parameters through the visualization front end, and the front end transmits the data to the back end for processing;

[0096] Step 2: The cement hydration microstructure evolution simulation task is started; after the experimenter configures the curing parameters in the front-end console, he clicks the "Start Hydration" button, and the back-end Django server receives the task request and starts the NC-GPCNN simulation evolution intelligent model for reasoning;

[0097] Step 3: NC-GPCNN model reasoning and result transmission; The NC-GPCNN simulation evolution intelligent model performs dynamic prediction of cement microstructure based on autoregression. Each time the model completes a prediction, the simulation backend packages the results and transmits them to the front end for display;

[0098] Step 4: Real-time result display and interaction: After the front end receives the simulation evolution results, the experimenter can observe the dynamic evolution of the cement microstructure in real time, adjust the observation angle and zoom ratio, and obtain more detailed information;

[0099] Step 5: Data storage and export; users can choose to save experimental records or download simulation evolution results as files in a specified format for further analysis and research.

[0100] The above-mentioned virtual experiment platform moves the computing tasks to the server side, reducing the system's dependence on local device performance, allowing users to run complex simulation evolution experiments on ordinary devices.

[0101] At the same time, the platform is based on the Web architecture and integrates other intelligent reasoning models to realize a variety of prediction experiments (such as compressive strength prediction, etc.), allowing experimenters to quickly iterate experimental plans in a virtual environment, thereby improving research efficiency. The platform supports remote access, and experimenters do not need expensive computing equipment, but can conduct virtual experiments anytime and anywhere through a browser. The platform provides an efficient and low-cost solution for experimental research.

[0102] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. AI-driven cement hydration microstructure evolution method, characterized in that: The following steps are involved: Obtaining scanned images of cement samples and preprocessing them; The preprocessed image is processed by vertical mask convolution, horizontal mask convolution and neighborhood mask convolution respectively to obtain the vertical features, horizontal features and neighborhood features of the image; among them, the horizontal mask convolution processes all pixels on the left side of the row where the current pixel is located, the vertical mask convolution processes all pixels above the current pixel, and the neighborhood mask convolution processes the surrounding area of ​​the current pixel; The horizontal features and vertical features are used as the input of the gated mask convolution module. After the new horizontal features are generated by feature fusion, they are spliced ​​and fused with the neighborhood features. The horizontal features of the recurrent gating module pass through the output module to obtain the simulated evolution output result consistent with the input image size. The horizontal features of the recurrent gating module pass through the output module to obtain the simulation evolution output results consistent with the input image size, including: Among the vertical and horizontal features updated by the gated mask convolution module, the vertical feature is returned to the gated mask convolution module, and the horizontal feature output is concatenated with the conditional feature in the channel dimension to form a combined feature, which is returned to the gated mask convolution module; After the gated mask convolution module executes the set number of cycles, it outputs the combined features as the new horizontal features.

2. The AI-driven cement hydration microstructure evolution method according to claim 1, characterized in that: The preprocessed image passes through the vertical mask convolution module, the horizontal mask convolution module and the neighborhood mask convolution module in the feature input module to obtain the vertical features, horizontal features and neighborhood features of the image respectively.

3. The AI-driven cement hydration microstructure evolution method according to claim 2, characterized in that: The feature input module obtains the spliced ​​horizontal features and neighborhood features, and inputs them together with the vertical features into the recurrent gating module. The recurrent gating module has multiple gated masked convolution modules, each of which has a vertical masked convolution module and a horizontal masked convolution module. The gated masked convolution modules are called layer by layer in a cycle, and the receptive field range is controlled by using different expansion coefficients to obtain the spliced ​​horizontal features and neighborhood features.

4. The AI-driven cement hydration microstructure evolution method according to claim 1, characterized in that: The horizontal features of the recurrent gating module pass through the output module to obtain the simulation evolution output results consistent with the input image size, including: The input vertical feature v_stack is convolved through the vertical mask convolution module to generate the vertical stack output feature map v_stack_feat, which is divided into two parts along the channel dimension and multiplied by the corresponding activation function to generate the gated vertical feature output; The spliced ​​horizontal features and neighborhood features are used as the updated horizontal features h_stack to input into the horizontal mask convolution module, and h_stack_feat is generated through convolution. The context information of v_stack_feat is introduced into h_stack_feat by adding it to the projected v_stack_feat. After feature block division and corresponding activation function and multiplication, the gated horizontal feature output is obtained. The gated vertical feature output and horizontal feature output are the vertical and horizontal features updated by the gated mask convolution module.

5. The AI-driven cement hydration microstructure evolution method according to claim 1, characterized in that: The horizontal features of the recurrent gating module pass through the output module to obtain the simulated evolution output results that are consistent with the input image size, and also include: the gated horizontal feature output is further convolved and superimposed with the input horizontal feature h_stack through channel compression to ensure the continuity of the features.

6. The AI-driven cement hydration microstructure evolution method according to claim 1, characterized in that: The horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result that is consistent with the input image size. It also includes: convolving the neighborhood features through a neighborhood mask convolution layer to generate a new conditional feature map to help the output result introduce context information.

7. The AI-driven cement hydration microstructure evolution method according to claim 1, characterized in that: The horizontal features of the recurrent gating module pass through the output module to obtain a simulated evolution output result that is consistent with the input image size, and also include: the combined features are used as a new horizontal feature map through the activation layer, and then the number of channels is changed through the convolution layer to generate a simulated evolution result that is consistent with the input image size.

8. An online virtual experiment platform for implementing the method according to any one of claims 1 to 7, characterized in that: include: The visualization front end includes a user interaction part and a background management part. The user interaction part is used to interact with users and display experimental results. The background management part is used to receive and save image data, experimental data and experimental requests uploaded by users and send them to the simulation reasoning back end. The simulation reasoning backend receives the experimental request from the visualization frontend, preprocesses the data of the experimental request and returns the experimental results. It executes the simulation task through the intelligent model loaded by the proxy server and sends the obtained experimental results to the visualization frontend.

Images