A method for generating three-dimensional structures of porous materials through image data processing and its application

Abstract

This invention relates to a method for generating three-dimensional structures of porous materials through image data processing and its application. It belongs to the intersection of 3D reconstruction, deep learning, and material characterization, and solves the problems of high computational resource consumption, functional fragmentation, poor simulation adaptability, strong data dependence, and limited angular features in existing technologies directly applied to the construction of three-dimensional material structures. The method for generating three-dimensional structures of porous materials through image data processing includes the following steps: Step 1: Screening two-dimensional images to obtain images that meet the signal-to-noise ratio and contrast requirements as input microscopic image samples; preprocessing the input microscopic images to generate a training sample set; Step 2: Constructing the algorithm model; Step 3: Training the model, including discriminator model training and generator model training; Step 4: Inputting the sample images into the trained model, and obtaining the corresponding three-dimensional voxel structure based on the output three-dimensional probabilistic voxel structure.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of three-dimensional reconstruction, deep learning, and materials characterization, and particularly to an artificial intelligence-driven method for generating three-dimensional structures from two-dimensional images. It also integrates an automatic porosity calculation function and is applicable to the pore structure characterization and analysis of composite materials with multi-level pore structures in the fields of materials science, catalysis science, and geological engineering. Background Technology

[0002] Three-dimensional characterization of material microstructure is one of the core requirements for material performance analysis. Traditional methods rely on multi-view image acquisition, such as CT tomography and FIB, which have problems such as high equipment cost and complex operation. Some single-view driven structure generation methods have the defects of low three-dimensional structure accuracy and poor feature reproduction.

[0003] Artificial intelligence deep learning has been applied to structure generation, but when used directly for the construction of three-dimensional material structures, the following key problems exist:

[0004] High computational resource consumption: It needs to process high-dimensional voxel data, and the training process requires a lot of GPU memory and computing power, making it difficult to adapt to lightweight computing devices;

[0005] Strong data dependence: training with large-scale 3D labeled data, while obtaining 3D structural labeled data in the field of materials is difficult and costly;

[0006] Functional fragmentation: The generation of the structure and the calculation of key performance parameters such as porosity need to be completed in two steps, and the generated results lack direct simulation adaptability. Third-party software needs to be called for format conversion, which is inefficient and prone to introducing human error and format compatibility issues.

[0007] Limitations of angular features: Relying on single-angle slice data leads to directional bias in the generated 3D structure, resulting in insufficient accuracy in feature reconstruction of non-orthogonal angles; moreover, it does not consider the role of multi-view feature fusion in improving structural integrity, making it difficult to meet the requirement of isotropic 3D structure in simulation analysis. Summary of the Invention

[0008] In view of the above analysis, the present invention aims to provide a method for three-dimensional pore modeling and analysis of porous materials, which can solve at least one of the following technical problems: (1) large consumption of computing resources; (2) functional separation; (3) poor simulation adaptability; (4) strong data dependence; (5) limited angular features.

[0009] The objective of this invention is mainly achieved through the following technical solutions:

[0010] In a first aspect, the present invention provides a method for generating a three-dimensional structure of a porous material through image data processing, comprising the following steps:

[0011] Step 1: Filter the two-dimensional images to obtain images that meet the signal-to-noise ratio and contrast requirements as input microscopic image samples. Preprocess the input microscopic images to generate a training sample set. The preprocessing of the input microscopic images includes: binarization segmentation and data augmentation of the input microscopic images.

[0012] Step 2: Combine the optimization of generator and discriminator network parameters to construct the algorithm model;

[0013] Step 3: Train the model, including training the discriminator model and the generator model;

[0014] Step 4: Input the sample image into the trained model, and obtain the corresponding three-dimensional voxel structure based on the output three-dimensional probabilistic voxel structure.

[0015] Optionally, in step 2, the parameters used when constructing the algorithm model are as follows:

[0016] Generator: Input noise dimension 80-120, transposed convolution kernel size 3×3×3-5×5×5, 3×3×3-4×4×4, 3×3×3-4×4×4 respectively, output channel number 96-160, 48-80, 2-4 respectively, output layer uses activation function to map voxel value to pore / matrix probability distribution;

[0017] Discriminator: Convolutional kernel size 3×3-5×5, input channels 1-2, output channels 48-80, 96-160, 1 respectively, output layer uses The activation function outputs the probability of the structure's "realism" (0-1).

[0018] Optionally, in step 1, the two-dimensional images are filtered to find images that meet the requirements of a signal-to-noise ratio of 1.5-3.0 and a contrast ratio of 0.3-0.7 as input microscopic image samples.

[0019] Optionally, filtering a two-dimensional image includes the following steps: the algorithm automatically calculates the average intensity of entity pixels in the image. Average pixel intensity of aperture The signal-to-noise ratio (SNR) and contrast ratio (Cw) are calculated using the above parameters.

[0020] The formula for calculating the signal-to-noise ratio (SNR) is as follows:

[0021] ;

[0022] The formula for calculating contrast ratio Cw is as follows:

[0023] .

[0024] Optionally, in step 3, the discriminator training includes the following steps:

[0025] a. Real sample input: Take the reference 3D features from the training set, input them into the discriminator, and obtain the real sample probability. ;

[0026] b. Sample Input Generation: The generator receives a random noise vector and a 2D image feature vector, and generates pseudo-3D structural features. Input the data into the discriminator to obtain the true probability of the generated sample. ;

[0027] c. Calculate the discriminator loss:

[0028] ;

[0029] Where λ is the weight coefficient of the feature matching loss, L D The discriminator loss measures the discriminator's ability to distinguish between "real structures" and "generated structures," where N is the batch size. Let be the probability of the discriminator's accuracy for the i-th true 3D structural feature. Let be the probability of the discriminator's assessment of the authenticity of the false 3D structural features output by the i-th generator. This represents the feature vector extracted from the real 3D structural features by the intermediate convolutional layers. L1 is the feature vector extracted by the intermediate convolutional layer of the discriminator from the pseudo-3D structural features output by the generator; L1 is the loss function, which calculates the difference between the two input features.

[0030] d. Backpropagation and parameter update: With generator parameters fixed, perform backpropagation on... The derivative is calculated, and the Adam optimizer is used to update the parameters of the convolutional and fully connected layers of the discriminator. The learning rate is adjusted according to the scheduling strategy.

[0031] Generator training includes the following steps:

[0032] a. Generating fake samples: The generator receives a random noise vector and a 2D image feature vector again to generate new fake 3D structural features. ;

[0033] b. Discriminator scoring: Input the data into the discriminator and obtain the true probability output by the discriminator. ;

[0034] c. Calculate the generator loss:

[0035] ;

[0036] Among them, L GThe generator loss is N, where N is the batch size. For the discriminator to generate pseudo-3D structural features again for the i-th generator The probability of authenticity is given by λ, where λ is the weight coefficient of the feature matching loss, L1 is the loss function, and the difference between the two input features is calculated. λ represents the feature vector extracted from the true 3D structural features by the intermediate convolutional layer. For intermediate convolutional layers to pseudo-3D structures Feature vectors extracted from features To calculate the structural consistency loss, the porosity deviation between the generated structure and the input 2D image is calculated; the larger the deviation, the greater the loss. We assign weights ([0.2-0.4]) to ensure that the porosity of the generated structure matches that of the input image;

[0037] d. Backpropagation and parameter update: With the discriminator parameters fixed, for Take the derivative, update the generator parameters, and complete one round of training.

[0038] Secondly, the present invention also provides the application of the above-mentioned method in the pore structure characterization and analysis of composite materials with hierarchical pore structures, comprising the following steps:

[0039] Step i: Calculate the porosity and pore size distribution of the three-dimensional voxel structure obtained by the above method;

[0040] Step ii: Perform material property analysis based on the three-dimensional voxel structure and the porosity and pore size distribution results obtained in step i.

[0041] Optionally, in step i, calculating the porosity of the three-dimensional voxel structure includes the following steps:

[0042] After binarizing the generated three-dimensional voxel structure, the voxel proportion is statistically analyzed to obtain the porosity result;

[0043] Calculating the pore size distribution of a three-dimensional voxel structure involves the following steps:

[0044] Based on the binarized three-dimensional voxel structure, an integrated analysis of aperture distribution is achieved by using a process of extracting equivalent diameters from connected holes and calculating distribution statistics.

[0045] Optionally, the following steps may be included between step i and step ii:

[0046] Step i-1: Perform simulation format compatibility conversion;

[0047] Step i-2: Verify the multi-dimensional model.

[0048] Optionally, step i-1, performing simulation format compatibility conversion, includes the following steps:

[0049] Step 1: First, filter out isolated small pores, exclude pores with a volume of less than 5-20 voxels, and extract the pore connectivity network to preserve details;

[0050] Step II: Using the improved Marching Cubes algorithm, the voxel structure is converted into a triangular mesh, which is then exported in STL format.

[0051] Optionally, in step i-2, the multi-dimensional model verification includes comparing the porosity and pore size distribution results calculated by the model with the measured values ​​to verify the consistency between the three-dimensional structure generated by the model and the actual structure of the sample.

[0052] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0053] (1) By optimizing the network parameters of the generator and discriminator, and combining the quantitative screening mechanism of signal-to-noise ratio and contrast, invalid images with low contrast and high noise are automatically removed, ensuring the consistency and effectiveness of training samples, greatly reducing the memory and computing power consumption of high-dimensional voxel data processing, and can be adapted to lightweight computing devices.

[0054] (2) It integrates the full-process functions of “3D structure generation → porosity calculation → pore size distribution analysis → simulation format conversion → multi-dimensional verification”, without the need to call third-party software for step-by-step processing. The whole process is automated without human intervention, avoiding human operation errors and greatly improving data processing efficiency.

[0055] (3) The simulation format compatibility conversion method of the present invention has high channel connectivity restoration degree, no manual intervention throughout the process, avoids the error introduced by manual operation of third-party format conversion tools, and does not require additional format conversion or manual parameter input when importing into software such as Fluent, ABAQUS, COMSOL, and LAMMPS. It has good simulation adaptability and can directly carry out finite element analysis, multiphysics simulation and molecular dynamics simulation.

[0056] (4) Expand sample diversity by using data augmentation techniques such as random mirroring, rotation, and grayscale perturbation, and ensure sample effectiveness by combining signal-to-noise ratio and contrast quantification screening mechanisms, thereby reducing reliance on large-scale 3D labeled data and lowering data acquisition costs.

[0057] (5) By using multi-view feature fusion design, the problem of "directional bias caused by single angle dependence" in traditional methods is solved. The porosity deviation of the three dimensions X, Y and Z is ≤2%-4%, and the pore size distribution characteristics are consistent, which meets the requirements of simulation analysis for the anisotropy of three-dimensional structures.

[0058] (6) By comparing the pore size distribution in the BET experiment and verifying the porosity of the water absorption method, the three-dimensional verification system is ensured to be consistent with the physical properties of the generated structure and the real sample, providing reliable data support for the analysis of the microscopic properties such as the material's permeability and mechanical properties.

[0059] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages will become apparent from the description or may be learned by practicing the invention. Attached Figure Description

[0060] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0061] Figure 1 This is a dual-beam electron microscope image of the sample from Example 1;

[0062] Figure 2 This is a binarized image of the dual-beam electron microscope image of the sample from Example 1 after preprocessing;

[0063] Figure 3 This is a three-dimensional pore structure diagram generated by inputting the sample image into the trained model in Example 1;

[0064] Figure 4(a) shows the calculation results of the aperture distribution in Example 1 using the probability density function;

[0065] Figure 4(b) shows the pore size distribution calculation results of Example 1 using the cumulative density function;

[0066] Figure 4(c) shows the calculated pore size distribution of Example 1 as a bar chart;

[0067] Figure 5(a) is a visualization of the three-dimensional voxel structure converted to a simulation format compatible form using the present invention, and the output STL format file in COMSOL software.

[0068] Figure 5(b) is a visualization of the structure of the STL format file output using the existing Fiji software;

[0069] Figure 6 To verify the model using the method of the present invention. Detailed Implementation

[0070] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which constitute a part of the present invention and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0071] In a first aspect, the present invention provides an artificial intelligence-driven method for generating three-dimensional structures from two-dimensional images, comprising the following steps:

[0072] Step 1: Filter the two-dimensional images to obtain images that meet the signal-to-noise ratio and contrast requirements as input microscopic image samples. Preprocess the input microscopic images to generate a training sample set. The preprocessing of the input microscopic images includes: binarization segmentation and data augmentation of the input microscopic images.

[0073] Step 2: Combine the optimization of generator and discriminator network parameters to construct the algorithm model;

[0074] Step 3: Train the model, including training the discriminator model and the generator model;

[0075] Step 4: Input the sample image into the trained model, and obtain the corresponding three-dimensional voxel structure based on the output three-dimensional probabilistic voxel structure.

[0076] In step 1, data augmentation employs random mirroring, rotation, and grayscale perturbation to enhance sample diversity and network generalization ability.

[0077] Two-dimensional images are filtered to find those with a signal-to-noise ratio of 1.5-3.0 and a contrast ratio of 0.3-0.7 as input microscopic image samples.

[0078] Image filtering includes the following steps: The algorithm automatically calculates the average intensity of entity pixels in the image ( ), average pixel intensity of aperture ( The signal-to-noise ratio (SNR) and contrast ratio (Cw) are calculated using the parameters mentioned above. The formula for calculating the signal-to-noise ratio (SNR) is as follows:

[0079] ;

[0080] The formula for calculating contrast ratio (Cw) is as follows:

[0081] ;

[0082] Based on the stability and regularity of the calculation results, the signal-to-noise ratio (SNR) [1.5-3.0] and contrast ratio (Cw) [0.3-0.7] were selected as the core indicators of image input quality, and samples that did not meet the indicators were removed.

[0083] In step 2, the parameters used when constructing the algorithm model are as follows:

[0084] Generator: Input noise dimension [80-120], transposed convolution kernel sizes are [3×3×3-5×5×5], [3×3×3-4×4×4], [3×3×3-4×4×4], and the number of output channels are [96-160], [48-80], and [2-4] respectively. The output layer uses... The activation function maps voxel values ​​to a probability distribution of pores / matrix.

[0085] Discriminator: Convolutional kernel size [3×3-5×5], number of input channels [1-2], number of output channels [48-80], [96-160], [1] respectively, output layer uses The activation function outputs the probability of the structure's "realism" (0-1).

[0086] Among these measures, residual blocks are added between the transposed convolutional layers of the generator to alleviate the gradient vanishing problem and improve the accuracy of micro-feature restoration.

[0087] Step 3, discriminator training includes the following steps:

[0088] a. Real sample input: Take the reference 3D features from the training set, input them into the discriminator, and obtain the real sample probability. ;

[0089] b. Sample Input Generation: The generator receives a random noise vector and a 2D image feature vector, and generates pseudo-3D structural features. Input the data into the discriminator to obtain the true probability of the generated sample. ;

[0090] c. Calculate the discriminator loss:

[0091] ;

[0092] Where λ is the weight coefficient of the feature matching loss, L D The discriminator loss measures the discriminator's ability to distinguish between "real structures" and "generated structures," where N is the batch size. This represents the feature output of the intermediate convolutional layer of the discriminator. Let be the probability of the discriminator's accuracy for the i-th true 3D structural feature. Let be the probability of the discriminator's assessment of the authenticity of the false 3D structural features output by the i-th generator. This represents the feature vector extracted from the real 3D structural features by the intermediate convolutional layers. L1 is the feature vector extracted by the intermediate convolutional layer of the discriminator from the pseudo-3D structural features output by the generator; L1 is the loss function, which calculates the difference between the two input features.

[0093] d. Backpropagation and parameter update: With generator parameters fixed, perform backpropagation on... The derivative is calculated, and the Adam optimizer is used to update the parameters of the convolutional and fully connected layers of the discriminator. The learning rate is adjusted according to the scheduling strategy.

[0094] Generator training includes the following steps:

[0095] a. Generating fake samples: The generator receives a random noise vector and a 2D image feature vector again to generate new fake 3D structural features. ;

[0096] b. Discriminator scoring: Input the discriminator to obtain the true probability. ;

[0097] c. Calculate the generator loss:

[0098] ;

[0099] Among them, L G The generator loss is N, where N is the batch size. For the discriminator to generate pseudo-3D structural features again for the i-th generator The probability of authenticity is given by λ, where λ is the weight coefficient of the feature matching loss, L1 is the loss function, and the difference between the two input features is calculated. This represents the feature vector extracted from the real 3D structural features by the intermediate convolutional layers. For intermediate convolutional layers to pseudo-3D structures Feature vectors extracted from features To calculate the structural consistency loss, the porosity deviation between the generated structure and the input 2D image is calculated; the larger the deviation, the greater the loss. We assign weights ([0.2-0.4]) to ensure that the porosity of the generated structure matches that of the input image;

[0100] d. Backpropagation and parameter update: With the discriminator parameters fixed, for Take the derivative, update the generator parameters, and complete one round of training.

[0101] In step 3, the standard for a well-trained model is that the discriminator cannot distinguish the difference between the generator's sample and the input sample.

[0102] Secondly, the present invention also provides the application of the above method in the pore structure characterization and analysis of composite materials with hierarchical pore structures, comprising the following steps:

[0103] Step i: Calculate the porosity and pore size distribution of the three-dimensional voxel structure obtained by the above method;

[0104] Step ii: Perform material property analysis based on the three-dimensional voxel structure and the porosity and pore size distribution results obtained in step i.

[0105] Step i involves calculating the porosity of the three-dimensional voxel structure, including the following steps:

[0106] After binarizing the generated three-dimensional voxel structure, the voxel proportions are statistically analyzed to obtain the porosity results.

[0107] Specifically, a voxel statistical algorithm is used: through The function binarizes the three-dimensional probabilistic voxel structure, marking voxels with a probability greater than a set threshold (threshold range [0.4-0.6]) as pores, otherwise as matrix; and counts the total number of pore voxels in the three-dimensional structure. ) and the total number of primes ( Calculate the porosity P using the following formula:

[0108] .

[0109] Calculating the pore size distribution of a three-dimensional voxel structure involves the following steps:

[0110] Based on the binarized three-dimensional voxel structure, an integrated analysis of aperture distribution is achieved by employing a process of connected aperture extraction, equivalent diameter calculation, and distribution statistics.

[0111] Connecting hole extraction: A 3D connected region labeling algorithm is used, which defines a voxel as connected if they are adjacent along the x / y / z axes. All independent connecting holes in the 3D structure are extracted, and isolated small holes with a volume smaller than a set threshold are excluded to avoid noise interference.

[0112] Equivalent diameter calculation: For each independent connected hole, calculate its volume, based on the spherical volume formula. The equivalent sphere radius is converted to the equivalent sphere diameter. The equivalent diameter is used as the characteristic pore size. Distribution statistics: The equivalent diameter is divided into [8-15] intervals, and the interval range is adjusted according to the actual pore size range. The proportion of the number of pores or the proportion of the pore volume in each interval is statistically analyzed to generate a pore size distribution histogram and data table. The calculation process is fully automated and requires no manual intervention. The results can be directly used for the analysis of microscopic properties such as material permeability and mechanical properties.

[0113] The following steps are also included between step i and step ii:

[0114] Step i-1: Perform simulation format compatibility conversion;

[0115] Step i-2: Verify the multi-dimensional model.

[0116] Step i-1 involves automating the format conversion of the generated 3D binary voxel structure. The aim is to integrate the generated 3D structure with porosity data, perform simulation-compatible format conversion, and output standardized characterization results that can be directly used for material property simulation. It supports standard formats from mainstream simulation platforms and can directly interface with workflows such as finite element analysis and molecular dynamics simulations without the need for additional format conversion tools. Compared with commonly used format conversion tools, the model retains a higher level of detail (see Figures 5(a) and 5(b)).

[0117] Specifically, step i-1 includes the following steps:

[0118] Step 1: First, isolate small pores by PoreSpy filtering to exclude pores with a volume smaller than [5-20] voxels, and extract the pore connectivity network to preserve details;

[0119] Step II: Using the improved Marching Cubes algorithm with a step size of 1 to preserve details to the greatest extent, the voxel structure is converted into a triangular mesh. Then, the STL format is exported to support simulation and automatically embeds parameters such as porosity and average pore size for direct use by simulation software.

[0120] Compared with existing Fiji tools, the simulation format compatibility conversion method of this invention has a high degree of channel connectivity restoration, requires no manual intervention throughout the process, avoids the errors introduced by manually operating third-party format conversion tools, and does not require additional format conversion or manual parameter input when importing into software such as Fluent, ABAQUS, COMSOL, and LAMMPS. It can directly carry out finite element analysis, multiphysics simulation, and molecular dynamics simulation.

[0121] In step i-2, the multi-dimensional model verification includes comparing the porosity and pore size distribution results calculated by the model with the measured values ​​to verify the consistency between the three-dimensional structure generated by the model and the actual structure of the sample.

[0122] Comparing the pore size distribution results calculated by the model with the measured values ​​includes: visually comparing the pore size distribution results generated by the model with the statistical data measured by BET real samples to ensure that the distribution of the main pore size characteristics is consistent.

[0123] Comparing the porosity distribution results calculated by the model with the measured values ​​includes comparing the porosity calculation results generated by the model with the values ​​measured by the water absorption method, with the deviation controlled within [3%-5%].

[0124] In addition, step i-2 verifies the crystal phase growth law of the three-dimensional structure generated by the model to check whether the crystal phase growth law of the three-dimensional structure generated by the model is consistent with the crystal phase growth law of the real structure of the sample.

[0125] Taking a three-dimensional isotropic material as an example, the verification process is as follows: extract the X, Y, and Z view images and calculate the porosity in each direction (deviation ≤ [2%-4%]); input the extracted view images into the original algorithm again to calculate the pore size distribution and verify the consistency of the three-dimensional structure in different dimensions.

[0126] The method of the present invention will be described in detail below with reference to specific embodiments.

[0127] Example 1:

[0128] Step 1: Preprocess the 2D images to generate a training sample set.

[0129] Images are filtered to find those with a signal-to-noise ratio of 3.0 and a contrast ratio of 0.4 as input samples;

[0130] Input data: Double-beam electron microscope images of alumina, where black areas represent pores and white areas represent the alumina matrix;

[0131] Preprocessing steps: OTSU threshold segmentation is used to obtain a binarized image. The batch function is called to randomly crop the binarized image into 64×64 sub-images, generating 1000 samples. These samples are then divided into training and validation sets in an 8:2 ratio. The training set is augmented to 1600 samples and organized into training data with a batch size of 32.

[0132] Step 2: Construct the algorithm model

[0133] Network parameter settings:

[0134] Generator: Input noise dimension 100, transposed convolution kernel size is 4×4×4, 3×3×3, 3×3×3 respectively, output channel number is 128, 64, 2 respectively;

[0135] Discriminator: Convolution kernel size 3×3, input channel number 1, output channel number 64, 128, 1 respectively;

[0136] Optimizer: Adam optimizer, β1=0.5, β2=0.999;

[0137] Loss function: Binary cross-entropy loss.

[0138] Step 3: Train the model

[0139] Initialize the generator and discriminator parameters, and fix the random seed to ensure reproducibility;

[0140] The training process is iterated for 800 rounds, and the similarity of the generated structures is tested on the validation set after each round of training.

[0141] Training is stopped when the Jaccard coefficient on the validation set stabilizes above 0.85, and the trained model weights are saved.

[0142] Step 4: Input the sample image into the trained model to obtain the corresponding three-dimensional voxel structure.

[0143] Load the pre-trained generator model, inputting a 4×4×4 random noise vector; the generator outputs a 64×64×64 three-dimensional probability voxel structure, which is then processed... Binarization yields a three-dimensional binary structure, which in turn yields the corresponding three-dimensional voxel structure.

[0144] Application Example 1:

[0145] Step i: Calculate the porosity and pore size distribution of the three-dimensional voxel structure obtained in Example 1.

[0146] Porosity calculation: Statistical analysis of the number of pore volumetric elements in a three-dimensional structure, and the total number of prime elements. Calculate porosity using the formula; Output results:

[0147] Aperture distribution calculation: Connectivity was used to mark the connected pores in the 3D binary structure. A volume threshold of 10 voxels was set, and isolated pores smaller than this threshold were excluded. A total of [200-300] independent connected pores were extracted. Equivalent diameter calculation: For each connected pore, the number of voxels was counted. According to actual volume Converted to actual volume, and then through Calculate the equivalent diameter; generate a histogram of aperture distribution and a data statistics table, which includes the number, volume, equivalent diameter and interval of each connected aperture.

[0148] Step i-1: Simulation format compatibility conversion

[0149] For the 3D binary voxel structures generated by the program, an automated format conversion is achieved. First, isolated small pores are filtered using PoreSpy to exclude pores with a volume smaller than [5-20] voxels. The pore connectivity network is extracted to preserve details. Then, an improved Marching Cubes algorithm with a step size of 1 is used to convert the voxel structure into a triangular mesh. Subsequently, a format that can be directly supported by mainstream simulation software is exported, and parameters such as porosity and average pore size are automatically embedded for direct use by simulation software. Compared with the existing Fiji tool, this conversion method has a high degree of pore connectivity restoration, requires no manual intervention throughout the process, avoids the errors introduced by manually operating third-party format conversion tools, and requires no additional format conversion or manual parameter input when importing into software such as Fluent, ABAQUS, COMSOL, and LAMMPS. Finite element analysis, multiphysics simulation, and molecular dynamics simulation can be performed directly.

[0150] Step i-2: Multi-dimensional model verification

[0151] The pore size distribution results generated by the model were visualized and compared with the statistical data measured from real BET samples. The results demonstrate that the pore size distribution of the 3D simulation and BET samples have good similarity (see...). Figure 6 This 3D simulation method can effectively characterize the pore size distribution characteristics of porous materials.

[0152] Step ii: Perform material property analysis

[0153] Material properties were analyzed based on the three-dimensional voxel structure, porosity, and pore size distribution.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for producing a three-dimensional structure of a porous material by image data processing, characterized by, The method comprises the following steps: Step 1: screening two-dimensional images to obtain images meeting the signal-to-noise ratio and contrast as input micrograph samples, and pre-processing the input micrograph to generate a training sample set; The pre-processing of the input micrograph comprises: binarization segmentation and data enhancement of the input micrograph; Step 2: combining generator and discriminator network parameter optimization to build a model of the algorithm; Step 3: training the model, including discriminator model training and generator model training; Step 4: inputting the sample image into the trained model to obtain the corresponding three-dimensional voxel structure based on the output three-dimensional probability voxel structure; In step 3, the discriminator training comprises the following steps: a. Real sample input: take the reference three-dimensional features of the training set, input the discriminator, and get the authenticity probability of the real sample ; b. Generating sample input: the generator receives a random noise vector and a two-dimensional image feature vector, generates a false three-dimensional structure feature , input discriminator, get the authenticity probability of the generated sample ; c. Calculate the discriminator loss: wherein λ is a weight coefficient of the feature matching loss, L D is a discriminator loss, N is a batch size, is a probability of authenticity of the i-th real three-dimensional structure feature by the discriminator, is a probability of authenticity of the i-th generated fake three-dimensional structure feature by the discriminator, is a feature vector extracted by the intermediate convolutional layer from the real three-dimensional structure feature, is a feature vector extracted by the intermediate convolutional layer of the discriminator from the generated fake three-dimensional structure feature; L 1 is a loss function, which calculates the difference between two input features; d. Backpropagation and parameter update: fixing the generator parameters, we take the derivative of the loss with respect to the discriminator parameters, and update the parameters of the discriminator's convolutional and fully connected layers using the Adam optimizer with learning rate adjusted according to the schedule. d. Backpropagation and parameter update: fixing the generator parameters, we take the derivative of the loss with respect to the discriminator parameters, and update the parameters of the discriminator's convolutional and fully connected layers using the Adam optimizer with learning rate adjusted The generator training comprises the following steps: (1) Generating fake samples: the generator receives again a random noise vector and a two-dimensional image feature vector, generating a new fake three-dimensional structure feature ; (2) Discriminator score: the input into the discriminator to obtain the authenticity probability of the discriminator output ; (3) Calculate the generator loss: in, L G The generator loss is N, where N is the batch size. For the discriminator to generate pseudo-3D structural features again for the i-th generator The probability of authenticity, where λ is the weight coefficient of the feature matching loss. L 1 represents the loss function, which calculates the difference between two input features. This represents the feature vector extracted from the real 3D structural features by the intermediate convolutional layers. For intermediate convolutional layers to pseudo-3D structures Feature vectors extracted from features To calculate the structural consistency loss, the porosity deviation between the generated structure and the input 2D image is calculated; the larger the deviation, the greater the loss. We use weights to ensure that the porosity of the generated structure matches the input image; (4) Backpropagation and parameter update: Fix the parameters of the discriminator, take the derivative, update the parameters of the generator, and complete a round of training. solving the derivative, updating the parameters of the generator, and completing a round of training.

2. The method of claim 1, wherein, In step 2, the parameters used when building the model of the algorithm are as follows: Generator: input noise dimension 80-120, transpose convolution kernel size 3x3x3-5x5x5, 3x3x3-4x4x4, 3x3x3-4x4x4, output channel number 96-160, 48-80, 2-4, respectively, output layer uses activation function, which maps voxel values to a probability distribution of pores / matrix; Discriminator: convolution kernel size 3x3-5x5, input channel number 1-2, output channel number 48-80, 96-160, 1 in turn, output layer adopts activation function, output structure "authenticity" probability.

3. The method of claim 1, wherein, In step 1, the two-dimensional images meeting the signal-to-noise ratio of 1.5-3.0 and the contrast of 0.3-0.7 are found as input micrograph samples.

4. The method of claim 3, wherein, The screening of the two-dimensional image comprises the following steps: the algorithm automatically calculates the average value of the intensity of the solid pixels in the image , the average value of the intensity of the pore pixels , and calculates the signal-to-noise ratio SNR and the contrast Cw using the above parameters; The signal-to-noise ratio SNR calculation formula is as follows: The contrast Cw calculation formula is as follows: 。 5. Use of the method according to any one of claims 1 to 4 for the characterization and analysis of the pore structure of composite materials having a hierarchical pore structure, characterized in that, The method comprises the following steps: Step i: calculating the porosity and pore size distribution of the three-dimensional voxel structure obtained by the method of any one of claims 1-4; Step ii: analyzing the material performance according to the three-dimensional voxel structure and the porosity and pore size distribution obtained in step i.

6. Use according to claim 5, characterized in that, In step i, the calculation of the porosity of the three-dimensional voxel structure comprises the following steps: After binarization of the generated three-dimensional voxel structure, the voxel proportion is counted to obtain the porosity result; The calculation of the pore size distribution of the three-dimensional voxel structure comprises the following steps: Based on the binarized three-dimensional voxel structure, the integrated analysis of the pore size distribution is realized by using the connected hole extraction equivalent diameter calculation-distribution statistics process.

7. Use according to claim 5, characterized in that, Between step i and step ii, the following steps are further included: Step i-1: performing simulation format compatible conversion; Step i-2: multi-dimensional model verification.

8. Use according to claim 7, characterized in that, In step i-1, the simulation format compatible conversion comprises the following steps: Step I: first filter isolated small pores, exclude pores with a volume less than 5-20 voxels, and extract the pore channel connection network to retain details; Step II: use an improved Marching Cubes algorithm to convert the voxel structure into a triangular mesh, and then export it in STL format.

9. Use according to claim 7, characterized in that, In step i-2, the multi-dimensional model verification comprises: comparing the porosity and pore size distribution results calculated by the model with the measured values respectively, to verify the consistency of the three-dimensional structure generated by the model with the real structure of the sample.

Images