A method for predicting a porous medium structure using machine learning
By using a generative adversarial network model to predict the structure of porous media, the problem of cumbersome and costly microstructure characterization schemes for porous media is solved, achieving rapid and efficient microstructure prediction and providing a reference for the design of efficient preparation process schemes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2024-04-01
- Publication Date
- 2026-07-03
Smart Images
Figure CN118279653B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of microstructure prediction technology, and in particular to a method for predicting porous media structure using machine learning. Background Technology
[0002] Porous media are complex materials composed of a solid framework and pores segmented by the framework. Based on pore structure, they can be further classified into isotropic and anisotropic porous media. The pore structure of a porous media determines its physical (mass transfer rate, permeability, electrical conductivity, adsorption characteristics) and chemical (catalytic activity, stability) properties. Due to the complexity of the geometry of porous media (torsional and random distribution of pore channels), the internal flow and heat transfer phenomena are also very complex. The structure of porous media is greatly affected by the preparation process, and the preparation process is complex. For different applications, extensive experiments are needed to explore the optimal preparation process under different conditions.
[0003] Porous media are typical multi-scale geometric systems. To analyze their structures, current simulation methods for reconstructing porous media mainly include physical methods, numerical methods, and fractal theory. However, the structures obtained using simulation methods differ significantly from the actual porous media. With the development of advanced image acquisition and collection technologies, instruments such as scanning electron microscopes, nuclear magnetic resonance imaging (NMR), and X-ray CT are used to obtain realistic two-dimensional and three-dimensional images of porous media, thereby analyzing their internal microstructure. However, traditional sampling methods can damage the original structure of the porous media. Furthermore, three-dimensional structural characterization is characterized by high equipment requirements, cumbersome measurement and post-processing steps, and high costs. Therefore, it is necessary to develop simpler methods to obtain the two-dimensional or three-dimensional structures of porous media, providing more and more effective reference data for the design of fabrication processes. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting the structure of porous media using machine learning. This method solves the problems of cumbersome operation and high time cost of existing microstructure characterization schemes, and can quickly and efficiently obtain the microstructure prediction results of porous media prepared under unknown target conditions, greatly improving the prediction accuracy and efficiency.
[0005] The technical solution adopted in this invention is as follows:
[0006] A method for predicting porous media structures using machine learning includes:
[0007] Images of porous media prepared under different working conditions and corresponding parameter information are obtained, including process parameters during the preparation process;
[0008] The image is processed based on the phase characteristics of porous media, and the color values of regions with the same phase are unified so that regions with different phases correspond to different color values, thereby obtaining an image dataset.
[0009] Two images C1 and C2 under different working conditions are randomly selected from the image dataset to form an image sample. An image sample and the corresponding parameter information are used to form a training sample. Several training samples are used to form a training set.
[0010] A generative adversarial network model is trained using the training set. It employs an unsupervised image-to-image conversion algorithm. For each input training sample, the output is: a reconstructed image C1. 1-1 The first predicted image C2 is generated from image C2 and used to predict image C1. 2-1 Reconstructing image C2 2-2 The second predicted image C1 is generated from image C1 to predict image C2. 1-2 The model is validated by the output predicted images until it meets the requirements, thus completing the training and obtaining the prediction model.
[0011] The porous medium image and corresponding parameter information prepared under known working conditions are obtained, and then input together with the preset parameter information of the target object into the prediction model to predict the image of the target object.
[0012] The further technical solution is as follows:
[0013] The acquisition of images of porous media prepared under different working conditions includes:
[0014] Porous media are prepared under different process parameters to obtain cross-sections of the porous media, and the cross-sections are photographed to obtain corresponding images.
[0015] Image processing is performed based on the phase characteristics of porous media, including:
[0016] A neural network-based image recognition and processing model is used to identify and classify phase features, and then adjust the color values of the regions corresponding to different phase features to obtain the processed image.
[0017] The phase characteristics include solid phase characteristics, porous phase characteristics, and characteristics of each solid phase component.
[0018] The structure of the generative adversarial network model includes an encoder, a shared encoder-generator module, a generator, and a discriminator. The encoder encodes the input original images C1 and C2 and their corresponding parameter information. The shared encoder-generator module extracts the encoded image feature information and parameter feature information, and outputs them to the generator and discriminator. The generator receives random noise from the feature information as input and outputs a structured image to the discriminator. The discriminator determines whether the structured image is the input original images C1 and C2, i.e., whether it is real data, and outputs a scalar value to represent the authenticity of the structured image.
[0019] The generator's loss function aims to minimize the discriminator's ability to correctly identify generated data, while the discriminator's loss function aims to maximize its ability to distinguish between real and generated data.
[0020] During the training of the generative adversarial network model, the generator's output is constrained by pre-defined physical knowledge and experience, including:
[0021] The difference between the two sets of parameter information corresponding to the input original image is calculated, and a non-linear function is used to calculate based on the difference. The number and position of pixels in the image generation process are limited based on the calculation result.
[0022] Both the generator and the discriminator are network structures that combine convolutional neural networks and fully connected networks; the loss functions of both the generator and the discriminator are binary cross-entropy functions.
[0023] The process parameters include calcination temperature and / or component ratio, wherein the component ratio includes raw material component ratio and finished product component ratio.
[0024] The beneficial effects of this invention are as follows:
[0025] This invention uses unsupervised image-to-image conversion methods from machine learning, based on partial experimental observation data, to predict the microstructure of porous media under various unknown preparation conditions. Compared with traditional structural analysis methods, this invention greatly saves experimental and time costs, and provides efficient and accurate reference data for designing preparation process schemes corresponding to the microstructure of porous media in different application scenarios.
[0026] This invention establishes a prediction model using an unsupervised image-to-image conversion algorithm and employs generative adversarial networks to improve the training and prediction accuracy of the algorithm. Furthermore, by comparing the phase features in the output image with the corresponding input parameter information, physical constraints are added to the algorithm, further enhancing the accuracy of the model's predictions.
[0027] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the prediction model training process according to an embodiment of the present invention. Detailed Implementation
[0029] The following describes specific embodiments of the present invention.
[0030] This embodiment of a method for predicting porous media structures using machine learning includes:
[0031] S1. Obtain images of porous media prepared under different working conditions and corresponding parameter information, including process parameters during the preparation process.
[0032] Specifically, porous media can be prepared by methods such as dry pressing, casting, phase inversion, and freeze drying. A flat cross-section can be obtained by mechanical or ion polishing, and then the cross-section can be photographed to obtain images.
[0033] The process parameters include calcination temperature and / or component ratios, whereby the component ratios include raw material component ratios and finished product component ratios. Examples include pore-forming agent ratios, slurry component ratios, and post-forming component ratios.
[0034] S2. Based on the phase characteristics of porous media, the image is processed to unify the color values of regions with the same phase, so that regions with different phases correspond to different color values, and the image dataset is obtained.
[0035] The phase characteristics include solid phase characteristics, porous phase characteristics, and characteristics of each solid phase component.
[0036] Specifically, images that meet the requirements can be generated directly during the image acquisition stage. For example, during image capture, the voltage and current can be adjusted by combining scanning electron microscopy to fully distinguish different solid components.
[0037] Specifically, a neural network-based image recognition processing model can be used to identify and classify phase features, and then adjust the color values of the regions corresponding to different phase features to obtain the processed image.
[0038] S3. Select any two images C1 and C2 from the image dataset under different working conditions to form an image sample. Use one image sample and the corresponding parameter information to form a training sample. Use several training samples to form a training set.
[0039] A generative adversarial network model is trained using the training set. It employs an unsupervised image-to-image conversion algorithm. For each input training sample, the output is: a reconstructed image C1. 1-1 The first predicted image C2 is generated from image C2 and used to predict image C1. 2-1 Reconstructing image C2 2-2 The second predicted image C1 is generated from image C1 to predict image C2. 1-2 The model is validated by the output predicted images until it meets the requirements, thus completing the training and obtaining the prediction model.
[0040] See Figure 1 The structure of the generative adversarial network model includes encoders (E1, E2), encoder-generator shared (Z), generators (G1, G2) and discriminators (D1, D2).
[0041] The encoders (E1, E2) are used to encode the input raw images C1, C2 and the corresponding parameter information (Synthesis condition 1, Synthesis condition 2);
[0042] The encoder and generator share (Z) the output of the encoder (E1, E2), extract the encoded image feature information and parameter information, and output them to the generator (G1, G2) and discriminator (D1, D2). The generator (G1, G2) receives a random noise from the feature information as input and outputs a structured image (C1). 1-1 C2 2-1 C2 2-2 C1 1-2 The input data (D1, D2) is given to the discriminator; the discriminator distinguishes whether its input data, i.e., the structural image, is the original input image C1, C2, i.e., whether it is real data. If it is, it outputs a scalar value T, indicating that its input is real data; if not, it outputs a scalar value F, indicating that its input is new data generated by the generator.
[0043] The generator's loss function aims to minimize the discriminator's ability to correctly identify generated data, while the discriminator's loss function aims to maximize its ability to distinguish between real and generated data.
[0044] In this embodiment, during the training process of the generative adversarial network model, the generator's output is limited by preset physical knowledge and experience, including:
[0045] The difference between the two sets of parameter information corresponding to the input original image is calculated. A non-linear function is used to calculate based on the difference. The number and position of pixels in the image generation process are limited based on the calculation result, thereby improving the accuracy of the generator output.
[0046] For example, the pre-set physical knowledge and experience is that if the calcination temperature is increased by 10°C during preparation, a certain solid phase component of the final sample will increase or decrease, and the number and position of the pixels representing the corresponding component in the generated image will change accordingly.
[0047] Therefore, this embodiment adds physical constraints to the algorithm based on the preparation process and physical information. By using partial differential equations to calculate and restrict image generation, the accuracy of the generator output is improved, thereby improving the prediction accuracy.
[0048] Specifically, in each training step of the generative adversarial network (GAN) model, the generator's parameters are first fixed to improve the discriminator's judgment ability. This is achieved by providing the discriminator with real data and fake data generated by the generator. Then, the discriminator's parameters are optimized to correctly classify real and fake data by adjusting the learning rate, loss function, and deepening the network structure. Next, the discriminator's parameters are fixed again to improve the generator's generation ability, making the generated fake data even more difficult for the discriminator to distinguish. This is achieved by optimizing the generator's parameters so that the discriminator incorrectly classifies fake data as real data. These steps are repeated continuously during training, each attempting to improve both the generator and discriminator, until a certain balance is reached where the generator can generate fake data that is very close to real data, while the discriminator struggles to distinguish between real and fake data.
[0049] In this embodiment, both the generator and the discriminator are network structures that combine convolutional neural networks and fully connected networks; the loss functions of both the generator and the discriminator are binary cross-entropy functions.
[0050] S4. Obtain the porous medium image Ck and corresponding parameter information prepared under known working conditions, and input it together with the preset parameter information of the target object into the prediction model to predict the image Cu of the target object. It can be understood that when the trained prediction model is applied, the model's output is the image output by the generator.
[0051] Based on this, this embodiment uses machine learning, based on some experimental data, to predict the structure of the target porous medium under unknown preparation conditions using the obtained prediction image Cu. It can predict the structure under various preparation conditions, meeting the needs of many applications of porous media.
[0052] The following experimental example illustrates the details and effectiveness of the solution in this embodiment.
[0053] This experimental example uses the dry pressing method to prepare solid oxide fuel cell anodes as the research object. The structure of the target solid oxide fuel cell is predicted using the scheme described in the above embodiments, including the following steps:
[0054] (1) Preparation of porous structure: The anode material was NiO / 8% yttrium oxide stabilized zirconium oxide (8YSZ) powder. Five groups of powders with equivalent Ni to 8YSZ volume ratios of 3:7, 4:6, 5:5, 6:4, and 7:3 were obtained. The two powders in each group were thoroughly mixed using a ball mill, and then a suitable amount of powder was pressed into thin sheets using a dry powder press. The thin sheets were calcined in a high-temperature box furnace to form a dense anode substrate. The anode substrate was then reduced at high temperature using hydrogen. Through the volume shrinkage of the NiO reduction process, a porous structure was formed, and five groups of samples were obtained.
[0055] (2) Image acquisition and preprocessing: The cross-sections of the five groups of samples were polished. Specifically, one or more of the following methods could be used: mechanical polishing, ultrasonic polishing, ion polishing, and chemical polishing, to obtain a flat cross-section. The cross-sectional structure was observed using a scanning electron microscope. During the image acquisition process, the acceleration voltage and current of the scanning electron microscope were adjusted to distinguish the Ni and 8YSZ phases. The image should display all cross-sectional features as much as possible. Based on the images acquired by the scanning electron microscope, the images were processed using a convolutional neural network to complete phase segmentation and single-value processing. Ni, 8YSZ, and the pore phase were identified and distinguished, and labeled as monochromatic values.
[0056] (3) Establish an unsupervised image-to-image conversion algorithm. Randomly select two images from five groups of sample images as image pairs, for a total of 20 image pairs. The corresponding parameter information is the volume ratio of Ni and 8YSZ. Use the image pairs and the corresponding volume ratio information to train and verify the algorithm to obtain the prediction model. During the training phase, for each image pair consisting of two images, the input is four images.
[0057] (4) Prediction using a prediction model: Set the input parameter information to a Ni and 8YSZ volume ratio of 3.5:6.5, use an image with a Ni and 8YSZ volume ratio of 3:7 as the input image, and obtain a prediction image corresponding to the Ni and 8YSZ volume ratio of 3.5:6.5 working condition, thus completing the prediction of the Ni and 8YSZ volume ratio of 3.5:6.5 working condition image.
[0058] Similarly, by modifying the input parameter information and setting the volume ratio of Ni to 8YSZ to 4.5:5.5, and using an image with a Ni to 8YSZ volume ratio of 4:6 as the image input, the corresponding predicted image is obtained. This completes the prediction of the image under the Ni to 8YSZ volume ratio of 4.5:5.5. By analogy, the prediction of images under various unknown target conditions can be completed, thereby realizing the prediction of porous media structures prepared under unknown target conditions.
[0059] This invention uses unsupervised image-to-image conversion methods from machine learning, based on partial experimental observation data, to predict the microstructure of porous media under different preparation conditions, which greatly saves experimental and time costs and provides high-quality porous media for different applications.
[0060] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting porous media structure using machine learning, characterized in that, include: Images of porous media prepared under different working conditions and corresponding parameter information are obtained, including process parameters during the preparation process; The image is processed based on the phase characteristics of porous media, and the color values of regions with the same phase are unified so that regions with different phases correspond to different color values, thereby obtaining an image dataset. Two images C1 and C2 under different working conditions are randomly selected from the image dataset to form an image sample. An image sample and the corresponding parameter information are used to form a training sample. Several training samples are used to form a training set. A generative adversarial network model is trained using the training set. It employs an unsupervised image-to-image conversion algorithm. For each input training sample, the output is: a reconstructed image C1. 1-1 The first predicted image C2 is generated from image C2 and used to predict image C1. 2-1 Reconstructing image C2 2-2 The second predicted image C1 is generated from image C1 to predict image C2. 1-2 The model is validated by the output predicted images until it meets the requirements, thus completing the training and obtaining the prediction model. The porous medium prepared under known working conditions and its corresponding parameter information are obtained, and then input together with the preset parameter information of the target object into the prediction model to predict the image of the target object. The structure of the generative adversarial network model includes an encoder, a shared encoder-generator module, a generator, and a discriminator. The encoder encodes the input original images C1 and C2 and their corresponding parameter information. The shared encoder-generator module extracts the encoded image feature information and parameter feature information, and outputs them to the generator and discriminator. The generator receives random noise from the feature information as input and outputs a structured image to the discriminator. The discriminator determines whether the structured image is the input original images C1 and C2, i.e., whether it is real data, and outputs a scalar value to represent the authenticity of the structured image. The generator's loss function aims to minimize the discriminator's ability to correctly identify generated data, while the discriminator's loss function aims to maximize its ability to distinguish between real data and generated data. During the training of the generative adversarial network model, the generator's output is constrained by pre-defined physical knowledge and experience, including: The difference between the two sets of parameter information corresponding to the input original image is calculated, and a non-linear function is used to calculate based on the difference. The number and position of pixels in the image generation process are limited based on the calculation result.
2. The method for predicting porous media structures using machine learning according to claim 1, characterized in that, The acquisition of images of porous media prepared under different working conditions includes: Porous media are prepared under different process parameters to obtain cross-sections of the porous media, and the cross-sections are photographed to obtain corresponding images.
3. The method for predicting porous media structures using machine learning according to claim 1, characterized in that, Image processing is performed based on the phase characteristics of porous media, including: A neural network-based image recognition and processing model is used to identify and classify phase features, and then adjust the color values of the regions corresponding to different phase features to obtain the processed image. The phase characteristics include solid phase characteristics, porous phase characteristics, and characteristics of each solid phase component.
4. The method for predicting porous media structures using machine learning according to claim 1, characterized in that, Both the generator and the discriminator are network structures that combine convolutional neural networks and fully connected networks; the loss functions of both the generator and the discriminator are binary cross-entropy functions.
5. The method for predicting porous media structures using machine learning according to claim 1, characterized in that, The process parameters include calcination temperature and / or component ratio, wherein the component ratio includes raw material component ratio and finished product component ratio.