A method for simulating the whole process of three-dimensional fracture of mesostructure of multiphase medium material based on space-time sequence prediction

Abstract

The application provides a kind of three-dimensional fracture whole process simulation method of multiphase medium material mesostructure based on space-time sequence prediction, comprising the following steps: S1, sample data acquisition;S2, deep learning model network building and training;S3, model verification: the historical surface crack information of composite material is input into the trained generation network, and the subsequent fracture damage process of composite material can be obtained, the morphology between fracture surfaces is quantified and compared by using Wasserstein distance, and the performance of the model is judged.The application can predict the fracture surface inside the entity according to the historical apparent crack of the composite material without the component distribution information of the composite material;The network model is trained completely from the perspective of observation data, and theoretically, the prediction result can be almost consistent with the observed result in the experiment;The calculation efficiency is high, and the calculation time can be shortened from weeks to seconds;Pixel-level modeling can be realized;The fracture process inside the material can be predicted according to the apparent crack of the material.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of fracture mechanics and computer science, specifically involving a method for simulating the entire three-dimensional fracture process of multiphase medium materials based on spatiotemporal sequence prediction. Background Technology

[0002] Composite materials, as core structural materials in aerospace, new energy, and infrastructure, exhibit highly complex fracture behavior due to their multiphase heterogeneous characteristics (such as the random aggregate distribution in concrete and the fiber-matrix interface in composite materials). Traditional fracture prediction methods face four major bottlenecks: ① They rely on destructive experiments to obtain component data (aggregate gradation and interfacial transition zone properties for concrete, and fiber distribution parameters for composite materials), making it difficult to adapt to material inhomogeneities and manufacturing deviations; ② Finite element and other mechanical simulations are time-consuming (mesh models for concrete require solving water pressure-axial force coupling equations, and composite materials require asymptotic damage iteration), failing to meet the real-time monitoring needs of infrastructure; ③ Mesh generation requires simplifying geometric details (such as the uneven interface of concrete aggregates and the fiber arrangement in composite materials), resulting in distortions in predictions due to localized stress concentrations. Furthermore, image recognition methods can only capture surface cracks and cannot reveal the aggregate-mortar debonding path within concrete or the delamination evolution of composite materials. To address the aforementioned challenges, this invention proposes a general fracture prediction method driven by a three-dimensional spatiotemporal sequence. Based solely on historical surface crack sequences (concrete monitored by DIC, composite materials by optical sensing), and combined with a deep spatiotemporal learning framework, it achieves millisecond-level fracture energy prediction (>10^6 times faster than traditional methods), pixel-level realistic geometric reconstruction (accuracy up to 5 μm, preserving the morphology of concrete aggregates and the microstructure of composite materials), and simultaneously outputs the three-dimensional evolution of internal fracture surfaces (quantifying the fracture path of concrete aggregates and the delamination damage of composite materials). This provides solutions for concrete infrastructure such as bridges and dams, as well as aerospace composite material components, from microscopic failure mechanisms to macroscopic disaster early warning. Summary of the Invention

[0003] The main objective of this invention is to provide a method for simulating the entire three-dimensional fracture process of multiphase medium materials based on spatiotemporal sequence prediction, addressing the aforementioned problems.

[0004] Therefore, the above-mentioned objective of the present invention is achieved through the following technical solution:

[0005] A method for simulating the entire three-dimensional fracture process of multiphase medium materials based on spatiotemporal sequence prediction includes the following steps:

[0006] S1. Sample data acquisition: Sample data includes spatial data on the distribution of composite material components, as well as spatial information such as the location and size of the fracture surface throughout the fracture process;

[0007] S2. Deep learning model network construction and training: The deep learning network framework adopts a conditional generative adversarial network, which consists of two parts: a generator network and a discriminator network.

[0008] S3. Model Validation: Input the historical surface crack information of the composite material into the trained generative network to obtain the subsequent fracture damage process of the composite material. Apply Wasserstein distance quantization and compare the morphology between fracture surfaces to determine whether the fracture surface predicted by the generative network is consistent with the actual result, and evaluate the performance of the model.

[0009] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0010] As a preferred technical solution of the present invention: in step S1, the data acquisition method can be divided into experimental observation and numerical simulation.

[0011] As a preferred technical solution of the present invention: experimental observation is conducted by in-situ fracture test through CT scanning to obtain the distribution of internal components of the composite material and the location of the fracture surface at each stage of the experiment.

[0012] As a preferred technical solution of the present invention: numerical simulation is used to simulate the fracture process of three-dimensional composite materials by means of existing numerical simulation methods, and to obtain the distribution of internal components of composite materials and the location of fracture surfaces at each stage.

[0013] As a preferred technical solution of the present invention: In step S2, the generating network is based on the U-net network framework, and multiple three-dimensional convolution operations are used to extract feature vectors of the composition and fracture surface spatial information of the composite material. The self-attention model represented by Transformer is applied to process the feature vectors, and the processed feature vectors are then deconvolutionally processed to generate the composite material spatial information for the next stage.

[0014] As a preferred technical solution of the present invention: In step S2, the discrimination network is different from the traditional two-dimensional convolutional neural network. Three-dimensional convolution operation is used. The concatenation matrix of the output result of the generator network and the real sample result is input to determine whether the two are consistent and a score is given. When they are consistent, the score is 1, and when they are inconsistent, the score is 0.

[0015] As a preferred embodiment of the present invention: Step S2, generating the network specifically includes the following steps:

[0016] S21. After inputting historical spatial data, normalize it.

[0017] S22. Through multiple three-dimensional convolution operations, dimension compression, pooling and other operations, the features of the normalized historical spatial information are extracted.

[0018] S23. When the input historical spatial data exists in multiple stages, the historical spatial information features of each stage are fused according to weights. The weight parameters are continuously optimized as the network is trained.

[0019] S24. Input the fused historical features into a temporal network represented by Transformer for temporal processing to generate the feature vector for the next stage.

[0020] S25. Through multiple operations such as three-dimensional deconvolution, dimension reshaping, skip connections, and feature fusion, the predicted feature vectors are reconstructed into three-dimensional spatial data.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1) This invention can predict the fracture surface inside the composite material based on the historical apparent cracks in the composite material without the need for information on the composition distribution of the composite material. In contrast, numerical models based on traditional mechanical equations require explicit information on the composition distribution to perform simulation calculations.

[0023] 2) Compared to the finite element method based on traditional mechanical equations, which requires the introduction of assumptions and causes calculation errors, this invention trains the network model entirely from the perspective of observation data, and theoretically can achieve prediction results that are almost consistent with the results observed in the experiment.

[0024] 3) Compared with the finite element method based on traditional mechanical equations, the present invention has high computational efficiency and can reduce the calculation time from weeks to seconds;

[0025] 4) In numerical simulation studies based on traditional mechanical equations, it is often necessary to simplify the solid model of composite materials, that is, it is impossible to accurately reproduce the local bumps and defects of composite materials. Overly fine mesh division may lead to strange calculation results, but this invention can achieve pixel-level modeling.

[0026] 5) Composite material fracture prediction methods based on deep learning image processing technology can only simulate the propagation of apparent cracks in the material, ignoring the longitudinal scale effect of the composite material and failing to simulate the internal fracture process of the composite material. However, this invention can predict the internal fracture process of the material based on apparent cracks. Attached Figure Description

[0027] Figure 1 The flowchart shows the three-dimensional fracture simulation method for the microstructure of multiphase media materials based on spatiotemporal sequence prediction provided by this invention.

[0028] Figure 2a This is the original slice image.

[0029] Figure 2bThis is to obtain the compressed slice image after processing.

[0030] Figure 3a The component division diagram is shown when the composite material is sliced ​​in three projection directions.

[0031] Figure 3b A component partitioning diagram when slicing a composite material entity in one projection direction.

[0032] Figure 4 The logical diagram of the cGAN-Transformerm model.

[0033] Figure 5 To generate the network framework diagram of network G.

[0034] Figure 6 The network framework diagram for generating network D.

[0035] Figure 7a This is a slice diagram with three projection directions: xy, yz, and xz.

[0036] Figure 7b This is a contour map of the fracture surface in three directions.

[0037] Figure 7c This is a probability distribution diagram. Detailed Implementation

[0038] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, a method for simulating the entire three-dimensional fracture process of multiphase medium materials based on spatiotemporal sequence prediction specifically includes the following steps:

[0040] S1. Sample data acquisition: Sample data includes spatial data on the distribution of composite material components, as well as spatial information such as the location and size of the fracture surface throughout the fracture process;

[0041] Data acquisition methods can be divided into experimental observation and numerical simulation. Experimental observation involves in-situ fracture tests using CT scans to obtain the distribution of internal components and fracture surface locations of composite materials at each stage of the experiment. Numerical simulation utilizes existing numerical simulation methods to simulate the fracture process of three-dimensional composite materials and obtain the distribution of internal components and fracture surface locations of composite materials at each stage (existing numerical simulation methods have long calculation times and require component distribution information to be simulated).

[0042] Because the data packets obtained from CT in-situ experiments or 3D numerical simulations are large in size and occupy a lot of computer space, this stage also involves data compression and storage technologies. In this patent, the original sample data is converted into slice images, and further image processing techniques such as noise reduction, enhancement, threshold segmentation, and interference removal are used to compress the original slice images into smaller slice images for storage. In the slice images used for storage, different components are distinguished based on pixel values. In subsequent use, the component information marked between slice images and pixel values ​​is combined to interpolate and reconstruct the internal 3D spatial information of the composite material (the above image storage process is as follows). Figures 2a-2b As shown, the spatial information reconstruction process is as follows: Figures 3a-3b As shown, concrete is used as an example. It should be noted that the composite material fracture prediction method described later in this invention is also applicable to uncompressed sample data.

[0043] How to extract the original sliced ​​image (e.g.) Figure 2a The compressed slice image (as shown) is obtained after processing. Figure 2b As shown), the specific steps include:

[0044] S111. Denoising and enhancing sliced ​​images: Reduce image noise interference by spatial domain filtering or frequency domain filtering, and enhance the edge and texture details between components by adjusting contrast.

[0045] S112. Threshold segmentation of the image: Differentiate components by controlling the threshold of pixel values. For example, set pixel values ​​<100 as component 1 and pixel values ​​≥100 as component 2.

[0046] S113. Remove small-area noise and isolated interference points through operations such as corrosion and expansion.

[0047] How to reconstruct compressed sliced ​​images into three-dimensional spatial data can be divided into two cases based on the slice dimension, and the specific steps are as follows:

[0048] S121. When slices in the three projection directions xy, yz, and xz are available, such as Figure 3a As shown, based on the component information at the intersection of the slices, the component division surface of the composite material is determined, and then the spatial information of the three-dimensional solid of the composite material is reconstructed.

[0049] S122. When only a slice in one projection direction can be obtained, such as when only a slice in the yz projection direction can be obtained, Figure 3b As shown, after determining the reference point, the maximum volume enclosed by connecting all the pixels of a single component is the spatial distribution information of that component.

[0050] Meanwhile, the spatial data in this invention is stored in the computer as a four-dimensional vector, consisting of [length, width, height, component type], which can be represented as [x, y, z, v]. The length, width, and height correspond to the x, y, and z values ​​of the composite material particle in the spatial coordinate system, respectively, while the component information v corresponds to the component information of the composite material at that particle. Taking concrete as an example, the component information can be divided into aggregate, mortar, pores, fracture surfaces, etc.

[0051] S2. Deep learning model network construction and training: The deep learning network framework adopts a conditional generative adversarial network, which consists of two parts: a generator network and a discriminator network.

[0052] like Figure 4 As shown, the generative network is based on the U-net network framework. It uses multiple three-dimensional convolution operations to extract feature vectors of the composition and fracture surface spatial information of the composite material. It applies a self-attention model represented by Transformer to process the feature vectors, and then applies deconvolution operations to reconstruct the composite material spatial information for the next stage.

[0053] like Figure 6 As shown, the discriminant network D differs from traditional two-dimensional convolutional neural networks by employing three-dimensional convolution operations. It takes the concatenated matrix of the generator network's output and the real sample results as input, determines whether the two are consistent, and assigns a score: 1 for consistency and 0 for inconsistency. This score is fed back to the generator network to help it optimize its parameters.

[0054] The cGAN-Transformer model proposed in this invention uses a 3D Conditional Generative Adversarial Network (cGAN) as its main framework to process spatial information (including fracture surfaces and component distribution information within composite materials). It also introduces a self-attention model, represented by Transformer, to establish temporal relationships between spatial data. Notably, the temporal algorithm in this invention can be adjusted to other temporal algorithms such as LSTM and RNN based on practical results. The cGAN consists of a generator network G and a discriminator network D. The generator network reads the spatial information of the previous stage or several stages of the composite material and generates the spatial information of the next stage through convolution and deconvolution operations. The discriminator network determines whether the output of the generator network is consistent with the results of the real samples. The generator and discriminator networks are trained together, simultaneously optimizing parameters and improving performance.

[0055] like Figure 5 As shown, the network framework for generating network G includes the following steps:

[0056] S21. After inputting historical spatial data, normalize it.

[0057] S22. Through multiple three-dimensional convolution operations, dimension compression, pooling and other operations, the features of the normalized historical spatial information are extracted.

[0058] S23. When the input historical spatial data exists in multiple stages, the historical spatial information features of each stage are fused according to weights. The weight parameters are continuously optimized as the network is trained.

[0059] S24. Input the fused historical features into a temporal network represented by Transformer for temporal processing to generate the feature vector for the next stage.

[0060] S25. Through multiple operations such as three-dimensional deconvolution, dimension reshaping, skip connections, and feature fusion, the predicted feature vectors are reconstructed into three-dimensional spatial data.

[0061] It should be noted that the spatial temporal model proposed in this invention is not limited to the network framework shown in the figure. The number of convolutional layers, the size of the convolutional kernels, the arrangement of functional layers, etc., can be adjusted according to practical conditions.

[0062] It should be noted that the three-dimensional spatial data in this invention includes information on the internal component distribution of the composite material and the location of the fracture surface. However, this method does not require the user to input all the three-dimensional spatial data. In fact, the biggest advantage of this method is that it can predict the internal fracture surface and surface cracks of the composite material even based solely on historical cracks on the surface of the composite material.

[0063] S3. Model Validation: Inputting historical surface crack information of the composite material into the trained generative network yields the subsequent fracture damage process of the composite material (including the surface and interior). During validation, the feature projection method is applied to compress the predicted spatial information into projected contour lines, which are then compared with the actual projected contour lines. Wasserstein distance quantization is applied to compare the morphology between fracture surfaces, determining whether the predicted fracture surfaces of the generative network match the actual results and evaluating the model's performance. Figures 7a-7c As shown.

[0064] How to quantitatively evaluate the accuracy of the predicted internal fracture surface of composite materials, that is, how to quantify and compare the fracture surface morphology, specifically includes the following steps:

[0065] S31. The simulated fracture surface of the composite material is fabricated into slices along the xy, yz, and xz projection directions (e.g., ...). Figure 7a As shown), the contour lines of the fracture surface in three directions are obtained (e.g. Figure 7b (as shown)

[0066] S32. Discretize each contour line by pixel to obtain the corresponding probability distribution map (e.g., Figure 7cAs shown), the Wasserstein distance between the two contour lines of the fracture surface can be obtained as follows:

[0067] ;

[0068] In the formula, This represents the probability distribution corresponding to the generated contour lines of the fracture surface. It can be divided into n segments, and the probability density of each segment is... ,satisfy ; This represents the probability distribution corresponding to the contour lines of the fracture surface in the target sample. It can be divided into m segments, and the probability density of each segment is... ,satisfy ; and Euler distance It can be represented as a norm. and satisfy , , ; This represents the combined distribution of two fracture surfaces.

[0069] By comparing all the contour lines, the differences between the generated fracture surface and the fracture surface of the target sample can be comprehensively evaluated.

[0070] The technical solution of the present invention has been described in conjunction with the specific experimental procedures shown in the accompanying drawings. However, the scope of protection of the present invention is not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for simulating the whole process of three-dimensional fracture of the mesostructure of multiphase medium material based on space-time sequence prediction, characterized in that, Includes the following steps: S1. Sample data acquisition: Sample data includes spatial data on the distribution of composite material components, as well as spatial information on the location and size of the fracture surface throughout the fracture process; S2. Deep learning model network construction and training: The deep learning network framework adopts a conditional generative adversarial network, which consists of two parts: a generator network and a discriminator network. S3. Model Validation: Input the historical surface crack information of the composite material into the trained generative network to obtain the subsequent fracture damage process of the composite material. Apply Wasserstein distance quantization and compare the morphology between fracture surfaces to determine whether the fracture surface predicted by the generative network is consistent with the actual result and to evaluate the performance of the model. In step S2, the generating network is based on the U-net network framework. It uses multiple three-dimensional convolution operations to extract feature vectors of the composition and fracture surface spatial information of the composite material. It applies a self-attention model represented by Transformer to process the feature vectors, and then applies deconvolution operations to the processed feature vectors to generate the composite material spatial information for the next stage. In step S2, the discrimination network is different from the traditional two-dimensional convolutional neural network. Three-dimensional convolution operation is used. The concatenation matrix of the output result of the generator network and the real sample result is input to determine whether the two are consistent and give a score. When they are consistent, the score is 1, and when they are inconsistent, the score is 0. Step S2, generating the network specifically includes the following steps: S21. After inputting historical spatial data, normalize it. S22. Through multiple three-dimensional convolution operations, dimension compression, and pooling, feature extraction of historical spatial information is performed on the normalized historical spatial information. S23. When the input historical spatial data exists in multiple stages, the historical spatial information features of each stage are fused according to weights. The weight parameters are continuously optimized as the network is trained. S24. Input the fused historical features into a temporal network represented by Transformer for temporal processing to generate the feature vector for the next stage. S25. Through multiple 3D deconvolutions, dimension reshaping, skip connections, and feature fusion, the predicted feature vectors are reconstructed into 3D spatial data.

2. The method according to claim 1, characterized in that: In step S1, data acquisition methods can be divided into experimental observation and numerical simulation.

3. The method according to claim 2, characterized in that: Experimental observations were conducted using CT scans to obtain the distribution of internal components and the location of fracture surfaces in the composite material at various stages of the experiment.

4. The method according to claim 2, characterized in that: Numerical simulation is used to simulate the fracture process of three-dimensional composite materials using existing numerical simulation methods, and to obtain the distribution of internal components and the location of fracture surfaces at each stage.

Images