A fuel microstructure prediction method based on a generative deep learning model

Abstract

This application relates to the field of computational materials science and technology, and particularly to a fuel microstructure prediction method based on a generative deep learning model, which includes the following steps: obtaining a data set; the data set includes SEM photos under different process conditions; preprocessing the data set, and obtaining a training set based on the preprocessed data set; constructing a generative deep learning model based on the conditional variational autoencoder model; inputting the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model; inputting the process conditions into the trained generative deep learning model to generate a specific fuel microstructure. This application directly generates microstructure images from chemical compositions and process preparation parameters by means of a generative deep learning model, getting rid of professional restrictions, and can directly output images consistent with microscopic detection results, solving problems such as high cost, long time consumption, and complex process of traditional experiments and simulation calculations.

Description

Technical Field

[0001] This application relates to the field of computational materials science technology, and in particular to a method for predicting the fuel microstructure based on a generative deep learning model. Background Art

[0002] As the core of a nuclear reactor, the performance of nuclear fuel is crucial for the performance and safety of the reactor. Previous studies have shown that adding Ti to UO2 can effectively improve the comprehensive performance of the fuel. Factors such as different Ti doping amounts, sintering temperature, and sintering holding time will all affect the microstructure of UO2 fuel. In order to obtain the fuel microstructure under different conditions, a large number of experimental studies need to be carried out by traditional methods, but the experimental conditions of UO2 are relatively harsh, with high costs and long cycles, making it difficult to carry out the fuel microstructure prediction method.

[0003] Currently, predicting the microstructure generally adopts the process of modeling - simulation calculation. Microstructure parameters such as the main phase and grain size are obtained through simulation calculation methods such as first - principles and thermodynamics. In the method of constructing a microstructure model using microstructure parameters, the following disadvantages exist: (1) The simulation calculation process is complex, requiring professional personnel and software for a large amount of calculation. (2) There are obvious differences between the simulation calculation results and the actual results. Therefore, there is an urgent need for a fuel microstructure prediction method to reduce experimental work and get rid of professional limitations. Summary of the Invention

[0004] This application provides a method for predicting the fuel microstructure based on a generative deep learning model. By means of the generative deep learning model, the microstructure image is directly generated from the chemical composition and process preparation parameters, getting rid of professional limitations, and being able to directly output an image consistent with the microscopic detection result, solving the problems of high cost, long time consumption, and complex process in traditional experiments and simulation calculations.

[0005] This application provides a method for predicting the fuel microstructure based on a generative deep learning model, including the following steps: First, obtain a data set; the data set includes SEM photos under different process conditions, and different process conditions are composed of combinations of different chemical compositions, sintering temperatures, and holding times; then, pre - process the data set to obtain the pre - processed data set, and based on the pre - processed data set, obtain a training set; next, based on the conditional variational auto - encoder model, construct a generative deep learning model; and input the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model; finally, input the process conditions into the trained generative deep learning model to generate a specific fuel microstructure.

[0006] In some exemplary embodiments, obtaining a dataset includes: First, for different chemical compositions and preparation process conditions, collecting SEM images of fuels; Then, based on the SEM images of fuels, establishing an SEM photo dataset corresponding to different chemical compositions - sintering temperatures - holding times.

[0007] In some exemplary embodiments, preprocessing the dataset includes: preprocessing the images and experimental data in the dataset respectively; Among them, preprocessing the images includes: measuring the scale bar length on the image, calculating the scale of the image, and scaling the image based on the scale, and performing histogram equalization processing on the image after converting it into a grayscale image to enhance the image contrast and visual effect of the image; Then converting the image into an RGB image, and finally completing the image preprocessing through random cropping and horizontal and vertical flipping operations; preprocessing the experimental data includes: using the mean-variance normalization method to process the variables of sintering temperature, Ti content, and holding time, and calculating the mean and standard deviation of each column and performing data conversion.

[0008] In some exemplary embodiments, based on the conditional variational autoencoder model, constructing a generative deep learning model; and inputting the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model, including: First, based on the conditional variational autoencoder model, increasing the depth of the encoder and decoder, optimizing the conditional variational autoencoder model, and constructing a generative deep learning model; Then, setting training parameters, and performing model training and optimization on the generative deep learning model based on the training set to obtain a trained generative deep learning model.

[0009] In some exemplary embodiments, increasing the depth of the encoder and decoder includes: using an attention mechanism to embed different process conditions and increasing the depth of the encoder and decoder; expanding both the encoder and decoder to 6 units, each unit including a convolutional layer, a pooling layer, and a normalization layer, and connecting the last two units using a residual structure.

[0010] In some exemplary embodiments, when training the generative deep learning model, the images of each process condition use the encoder separately, and all images share a decoder; combining the input process conditions through a linear layer into an embedding vector of length 10, and splicing the embedding vector and the latent vector of the encoder and then inputting them into the decoder; the loss function during training consists of a reconstruction loss and a KL divergence loss.

[0011] In some exemplary embodiments, the reconstruction loss is represented using cross-entropy loss, and the corresponding program is:

[0012] F.BCE_loss(image, generated_image)

[0013] The KL divergence loss corresponding program is as follows:

[0014] kl_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())

[0015] Among them, mu is the mean of the latent variable, and logvar is the standard deviation of the latent variable.

[0016] In some exemplary embodiments, inputting process conditions into the trained generative deep learning model to generate a specific fuel microstructure includes: First, converting the input compositional process conditions into an input embedding vector through a linear layer, combining the embedding vector generated during training and the latent variable, and using an attention mechanism to generate a new latent variable; then, inputting the new latent variable into the trained decoder to generate the corresponding image.

[0017] In some exemplary embodiments, converting the input compositional process conditions into an input embedding vector through a linear layer, combining the embedding vector generated during training and the latent variable, and using an attention mechanism to generate a new latent variable includes: Denoting the input variable as a i , and the latent variable as z i , where i is the serial number, and z i is the latent variable generated during training, consisting of 64 numbers; the calculation formula for the latent variable is:

[0018]

[0019] Among them, z4 is the latent variable of the test set data with serial number 4, and a4 is the input variable.

[0020] In some exemplary embodiments, inputting the new latent variable into the trained decoder to generate the corresponding image, and the image is used to display the grain shape and size of the microstructure.

[0021] The technical solution provided by the embodiments of the present application has at least the following advantages:

[0022] An embodiment of the present application provides a fuel microstructure prediction method based on a generative deep learning model, including the following steps: First, obtain a data set; the data set includes SEM photos under different process conditions, and different process conditions are composed of combinations of different chemical components, sintering temperatures, and holding times; then, preprocess the data set to obtain a preprocessed data set, and based on the preprocessed data set, obtain a training set; next, based on the conditional variational autoencoder model, construct a generative deep learning model; and input the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model; finally, input the process conditions into the trained generative deep learning model to generate a specific fuel microstructure.

[0023] The fuel microstructure prediction method based on the generative deep learning model provided by the present application can quickly realize the prediction of the fuel microstructure under different chemical components and preparation process conditions by using the generative deep learning method. The generated images can be directly compared with SEM images, truly reflecting the microstructure of the fuel, and greatly reducing the workload of experiments and traditional simulation calculations. Description of the Drawings

[0024] One or more embodiments are illustrated by way of example in the accompanying drawings, and these exemplary illustrations do not limit the embodiments unless otherwise stated. The figures in the drawings do not constitute a scale limitation.

[0025] Figure 1 It is a flowchart of a fuel microstructure prediction method based on a generative deep learning model provided by an embodiment of the present application.

[0026] Figure 2 It is a comparison diagram of the original image and the preprocessed image provided by an embodiment of the present application.

[0027] Figure 3 It is a curve of the loss function changing with the number of training rounds provided by an embodiment of the present application.

[0028] Figure 4 It is a comparison diagram of the original image and the generated image provided by an embodiment of the present application. Detailed Embodiments

[0029] As can be seen from the background technology, traditional methods require a large number of experimental studies to obtain the fuel microstructure under different conditions. However, the experimental conditions for UO2 are relatively harsh, with high costs and long cycles, making it difficult to carry out the fuel microstructure prediction method.

[0030] To solve the above technical problems, an embodiment of the present application provides a fuel microstructure prediction method based on a generative deep learning model. The method includes the following steps: First, obtain a data set; the data set includes SEM photos under different process conditions, and different process conditions are composed of combinations of different chemical components, sintering temperatures, and holding times; then, preprocess the data set to obtain a preprocessed data set, and based on the preprocessed data set, obtain a training set; next, based on the conditional variational autoencoder model, construct a generative deep learning model; and input the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model; finally, input the process conditions into the trained generative deep learning model to generate a specific fuel microstructure. The fuel microstructure prediction method based on the generative deep learning model provided by the present application directly generates a microstructure image from chemical components and process preparation parameters with the help of the generative deep learning model, getting rid of professional limitations, and can directly output an image consistent with the microscopic detection result, solving the problems of high cost, long time consumption, and complex process in traditional experiments and simulation calculations.

[0031] The following will elaborate on each embodiment of the present application in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present application and be able to implement it.

[0032] The following content is all specific examples of the specific implementation process provided to elaborate in detail the technical solution to be protected by the present application. However, the present application can also adopt other implementation manners different from this description. Those skilled in the art can, under the guidance of the idea of the present application, adopt different technical solutions to implement the content within the scope of the present application. Therefore, the present application is not strictly limited by the following specific implementation cases.

[0033] Refer to Figure 1 An embodiment of the present application provides a fuel microstructure prediction method based on a generative deep learning model, including the following steps:

[0034] Step S1, obtain a data set; the data set includes SEM photos under different process conditions, and different process conditions are composed of combinations of different chemical components, sintering temperatures, and holding times.

[0035] Step S2, preprocess the data set to obtain a preprocessed data set, and based on the preprocessed data set, obtain a training set.

[0036] Step S3, based on the conditional variational autoencoder model, construct a generative deep learning model; and input the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model.

[0037] Step S4: Input the process conditions into the trained generative deep learning model to generate the microstructure of a specific fuel.

[0038] In some embodiments, obtaining the dataset in step S1 includes: collecting SEM images of fuels for different chemical compositions and preparation process conditions; and establishing a dataset of SEM photos corresponding to different chemical compositions - sintering temperatures - holding times based on the SEM images of the fuels.

[0039] Specifically, in step S1, a dataset required for model training is obtained through experiments and literature collection. The data type of the training dataset is Ti content - sintering temperature - holding time - corresponding SEM photos.

[0040] In some embodiments, preprocessing the dataset in step S2 includes: preprocessing the images and experimental data in the dataset respectively; among them, preprocessing the images includes: measuring the scale bar length on the image, calculating the scale of the image, scaling the image based on the scale, and performing histogram equalization processing on the image after converting it to a grayscale image to enhance the image contrast and visual effect; then converting the image to an RGB image, and finally completing the image preprocessing through random cropping and horizontal and vertical flipping operations; preprocessing the experimental data includes: processing the variables of sintering temperature, Ti content, and holding time using the mean - variance normalization method, calculating the mean and standard deviation of each column and performing data conversion.

[0041] In some embodiments, in step S3, a generative deep learning model is constructed based on the conditional variational auto - encoder model; and the data of the training set is input into the generative deep learning model for model training to obtain a trained generative deep learning model, including: first, based on the conditional variational auto - encoder model, increasing the depth of the encoder and decoder to optimize the conditional variational auto - encoder model and construct a generative deep learning model; then, setting training parameters, and performing model training and optimization on the generative deep learning model based on the training set to obtain a trained generative deep learning model.

[0042] Specifically, this application optimizes and improves the conditional variational auto - encoder (CVAE) model for the scenario of predicting the fuel microstructure, embeds the composition process conditions with an attention mechanism, and increases the depth of the Encoder - Decoder. Set training parameters and perform model training and optimization based on the enhanced dataset.

[0043] Finally, input the data of the specified doped Ti content, sintering temperature, and sintering holding time into the model, and the model can generate the microstructure of a specific fuel.

[0044] By using the generative deep learning method, this application can quickly realize the prediction of the fuel microstructure under different chemical compositions and preparation process conditions. The generated images can be directly compared with SEM images, truly reflecting the fuel microstructure, and greatly reducing the workload of experiments and traditional simulation calculations.

[0045] Specifically, in the process of establishing the dataset, for different chemical compositions and preparation process conditions, SEM images of the fuel are collected to establish a dataset of chemical composition - sintering temperature - holding time - corresponding SEM photos. The data sources include actual experimental results and fuel microstructure images in papers. In this embodiment, a total of 12 SEM images corresponding to process conditions are collected. The specific Ti content, sintering temperature, and sintering holding time in the process conditions are shown in Table 1. Among them, No. 1, 4, 5, and 11 are from the literature, and the rest are supplemented by experiments.

[0046] Table 1 Temperature, Ti content, and time dataset

[0047]

[0048] In the process of data preprocessing, first, image preprocessing work is carried out. The scale bar length on the image is measured using ImageJ software, and the scale of the image is calculated. Among them, the scale of No. 4 is the smallest and does not need to be scaled. Other images are scaled according to the 4-bit standard. The specific measurement results and reduction ratios are shown in Table 2.

[0049] Table 2 Actual ratio and reduction ratio of images

[0050]

[0051] Use opencv to read the scaled-down images, first convert the read images to grayscale images, and use the cv2.equalizeHist function to perform histogram equalization processing on the images to increase the image contrast and visual effect of the enhanced image, and then convert the images to RGB images. Before training, image enhancement is carried out. First, use the RandomCrop method to randomly crop the images to a size of 256×256, and then randomly flip horizontally and vertically (with a probability of 0.2). Such operations can increase the data quantity. After the images are preprocessed as Figure 2 shown, Figure 2 in which, the left side is the original image, and the right side is the preprocessed image.

[0052] Then, preprocessing of the experimental data is carried out. Since the ranges of the three variables, namely temperature, Ti content, and time, are relatively large, in order to eliminate the influence brought by their ranges, mean-variance normalization is used to process the three variables of temperature, Ti content, and time, that is, the mean and standard deviation of each column are calculated, each data is divided by the standard deviation after subtracting the mean of its column, and finally the data is shown in Table 3.

[0053] Table 3 Results of data normalization processing

[0054]

[0055] Numbers 4 and 11 are selected from the data as the test set, and the remaining data is used as the training set for training.

[0056] In some embodiments, the depths of the encoder and decoder are increased, including: adopting an attention mechanism to embed different process conditions and increasing the depths of the encoder and decoder; expanding both the encoder and decoder to 6 units, each unit including a convolutional layer, a pooling layer, and a normalization layer, and connecting the last two units with a residual structure.

[0057] In some embodiments, when training the generative deep learning model, the images of each process condition use the encoder separately, and all images share a decoder; the input process conditions are combined into an embedding vector of length 10 through a linear layer, and the embedding vector and the latent vector of the encoder are concatenated and then input into the decoder.

[0058] Specifically, in the process of constructing and training the generative deep learning model, the embodiments of the present application select the CVAE model as the basic model and optimize it for the fuel application scenario. The specific optimization contents are as follows:

[0059] (1) Increase the depth of the encoder-decoder, expand it to 6 units, each unit includes a convolutional layer, a pooling layer, and a normalization layer, and connect the last two units with a residual structure.

[0060] (2) The images of each condition use the encoder separately and share a decoder.

[0061] (3) Combine the input condition chemical composition, sintering temperature, and holding time through a linear layer into an embedding vector of length 10.

[0062] (4) Concatenate the embedding vector and the latent vector of the encoder and then input them into the decoder.

[0063] During training, the loss function consists of two parts: the reconstruction loss and the Kullback - Leibler divergence loss. The reconstruction loss uses the cross - entropy loss, and the corresponding program is:

[0064] F.BCE_loss(image, generated_image)

[0065] The corresponding program for the Kullback - Leibler divergence loss is:

[0066] kl_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())

[0067] where mu is the mean of the latent variable and logvar is the standard deviation of the latent variable.

[0068] For the setting of other parameters, the total number of epochs is 1000, the batch size is set to 4, the optimizer is SGD, BCE + KL is used as the loss function, the training learning rate decreases according to the cosine function, and it is trained for 1000 rounds. The parameters of the decoder, the input conditional linear layer, the latent variable, and the generated image are saved every 10 rounds. The loss function during the training process is as Figure 3 shown.

[0069] In some embodiments, in step S4, the process conditions are input into the trained generative deep learning model to generate a specific fuel microstructure, including: First, the input composition process conditions are converted into an input embedding vector through a linear layer. Combining the generated embedding vector and the latent variable during training, a new latent variable is generated using the attention mechanism; then, the new latent variable is input into the trained decoder to generate the corresponding image.

[0070] The specific process is as follows:

[0071] (1). Denote the input variable as a i , the latent variable as z i , i as the serial number, a i is the data in Table 3, and z i is the latent variable generated during training, which consists of 64 numbers. Taking the 4th in the test set as an example, calculate according to the formula of the latent variable:

[0072]

[0073] (2). Input the new latent variable into the trained decoder to generate the corresponding image.

[0074] Figure 4 The original image and the generated image are shown, and the main characteristics of the microstructure - the grain shape and size - can be seen.

[0075] According to the above technical solution, the embodiment of the present application provides a fuel microstructure prediction method based on a generative deep learning model, including the following steps: First, obtain a data set; the data set includes SEM photos under different process conditions, and different process conditions are composed of combinations of different chemical compositions, sintering temperatures, and holding times; then, preprocess the data set to obtain a preprocessed data set, and based on the preprocessed data set, obtain a training set; next, based on the conditional variational autoencoder model, construct a generative deep learning model; and input the data of the training set into the generative deep learning model for model training to obtain a trained generative deep learning model; finally, input the process conditions into the trained generative deep learning model to generate a specific fuel microstructure.

[0076] The fuel microstructure prediction method based on the generative deep learning model provided by the present application can quickly realize the prediction of the fuel microstructure under different chemical compositions and preparation process conditions by using the generative deep learning method. The generated image can be directly compared with the SEM image, truly reflecting the microstructure of the fuel, and greatly reducing the workload of experiments and traditional simulation calculations.

[0077] Those of ordinary skill in the art can understand that the above embodiments are specific embodiments for implementing the present application. In actual applications, various changes can be made in form and details without departing from the spirit and scope of the present application. Any person skilled in the art can make their own changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application should be subject to the scope defined by the claims.

Claims

1. A fuel microstructure prediction method based on a generative deep learning model, characterized in that: The following steps are involved: Acquire a data set; the data set includes SEM photos under different process conditions, where the different process conditions are composed of different combinations of chemical composition, sintering temperature, and holding time; the chemical composition is 0% to 30% Ti content, the sintering temperature is 1650° C. to 1750° C., and the holding time is 1 h to 4 h; Preprocessing the data set to obtain a preprocessed data set, and obtaining a training set based on the preprocessed data set; Build a generative deep learning model based on the conditional variational autoencoder model; and inputting the data of the training set into the generative deep learning model to perform model training to obtain a trained generative deep learning model; Input process conditions into the trained generative deep learning model to generate a specific fuel microstructure; In the process of building and training a generative deep learning model, the conditional variational autoencoder model is optimized for fuel application scenarios. The optimization process includes: The encoder and decoder are both expanded to 6 units to increase the depth of the encoder and decoder; each unit includes a convolution layer, a pooling layer, and a normalization layer, and the last two units are connected with a residual structure; Each image of each process condition uses a separate encoder, and all images share a decoder; The input chemical composition, sintering temperature and holding time are combined into an embedding vector of length 10 through a linear layer; The embedding vector and the encoder’s latent vector are concatenated and then fed into the decoder. The process conditions are input into a trained generative deep learning model to generate a specific fuel microstructure, including: The input component process conditions are converted into input embedding vectors through a linear layer, and the embedding vectors and latent variables generated by training are combined to generate new latent variables using the attention mechanism; Input the new latent variable into the trained decoder to generate the corresponding image; The input component process conditions are converted into input embedding vectors through a linear layer. The embedding vectors and latent variables generated by training are combined to generate new latent variables using the attention mechanism, including: Let the input variable be a i , the latent variable is recorded as z i , where i is the sequence number, z i The latent variable generated for training consists of 64 numbers; the calculation formula of the latent variable is: Among them, z4 is the latent variable of the test set data with sequence number 4, and a4 is the input variable.

2. The fuel microstructure prediction method based on a generative deep learning model according to claim 1, characterized in that: Get the dataset, including: Collect SEM images of fuels with different chemical compositions and preparation process conditions; Based on the SEM images of the fuel, a SEM photo dataset corresponding to different chemical compositions, sintering temperatures and holding times was established.

3. The fuel microstructure prediction method based on a generative deep learning model according to claim 1, characterized in that: The data set is preprocessed, including: Preprocessing the images and experimental data in the data set respectively; Preprocessing the image includes: measuring the length of the ruler on the image, calculating the scale of the image, scaling the image based on the scale, converting it into a grayscale image, and performing histogram equalization on the image to enhance the image contrast and image visual effect; then converting the image into an RGB image, and finally completing the image preprocessing by random cropping and horizontal and vertical flipping operations; The experimental data were preprocessed, including: using the mean-variance normalization method to process the sintering temperature, Ti content and holding time variables, calculating the mean and standard deviation of each column and performing data conversion.

4. The fuel microstructure prediction method based on a generative deep learning model according to claim 1, characterized in that: Build a generative deep learning model based on the conditional variational autoencoder model; The data of the training set is input into the generative deep learning model to perform model training to obtain a trained generative deep learning model, including: Based on the conditional variational autoencoder model, increase the depth of the encoder and decoder, optimize the conditional variational autoencoder model, and build a generative deep learning model; The training parameters are set, and the generative deep learning model is trained and optimized based on the training set to obtain a trained generative deep learning model.

5. The fuel microstructure prediction method based on a generative deep learning model according to claim 1, characterized in that: When training the generative deep learning model, the loss function during training consists of reconstruction loss and KL divergence loss.

6. The fuel microstructure prediction method based on a generative deep learning model according to claim 5, characterized in that: The reconstruction loss is expressed using cross entropy loss, and the corresponding procedure is: F.BCE_loss(image, generated_image) The corresponding procedure for KL divergence loss is: kl_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp()) Among them, mu is the mean of the latent variable and logvar is the standard deviation of the latent variable.

7. The fuel microstructure prediction method based on a generative deep learning model according to claim 1, characterized in that: The new latent variables are input into the trained decoder to generate corresponding images, which are used to show the grain shape and size of the microstructure.

Images