Method, device, equipment, medium and product for predicting thermal barrier coating thermal conductivity

CN123032641APending Publication Date: 2026-11-03SHANGHAI JIAOTONG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611203231.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-10
Publication Date
2026-11-03

AI Technical Summary

Technical Problem

Existing technologies cannot assess the high-temperature thermal conductivity of thermal barrier coatings in real time while in service, and existing prediction methods ignore microstructure geometry information, resulting in poor prediction accuracy.

Method used

By obtaining the mechanical feature vector of the thermal barrier coating, an image generation model and semantic segmentation technology are used to generate a predicted microstructure image. The index values ​​of the defect instance image are determined by combining the target statistical index, and the thermal conductivity is calculated using a regression prediction model.

Benefits of technology

It enables online prediction of thermal conductivity without peeling off the coating. By combining microstructure geometry information, the prediction accuracy is improved and the cost and difficulty are reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123032641A_ABST
    Figure CN123032641A_ABST
Patent Text Reader

Abstract

The application discloses a thermal barrier coating thermal conductivity prediction method, device, equipment, medium and product, and relates to the thermal conductivity prediction field. The method comprises the following steps: acquiring a characteristic vector of a coating to be predicted, the characteristic vector comprising a mechanical characteristic vector, and the mechanical characteristic vector being used for indicating indentation test mechanical parameters of the thermal barrier coating; determining a predicted microstructure image of the coating to be predicted according to the characteristic vector of the coating to be predicted; determining a defect instance image corresponding to a target defect type according to the predicted microstructure image; determining an index value of the defect instance image under a target statistical index according to the defect instance image; and determining a predicted thermal conductivity of the coating to be predicted according to the index value of the defect instance image under the target statistical index. The application realizes online prediction of the thermal conductivity of the thermal barrier coating while ensuring prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of thermal conductivity prediction, and in particular to a method, apparatus, equipment, medium, and product for predicting the thermal conductivity of thermal barrier coatings. Background Technology

[0002] Thermal barrier coatings are widely used for thermal insulation and protection of high-temperature hot-end components such as aero engines and gas turbines, and their service performance directly affects engine efficiency and service life. Thermal conductivity is a core parameter characterizing the thermal insulation performance of thermal barrier coatings, and its value is strongly influenced by factors such as porosity, crack morphology, and interlayer bonding state in the coating's microstructure.

[0003] Currently, the determination of high-temperature thermal conductivity of thermal barrier coatings includes direct measurement methods and prediction methods that establish end-to-end thermal conductivity regression models based on mechanical characteristics as input. The former method requires peeling the coating off the substrate to prepare an independent sample, which is irreversible and cannot achieve in-situ evaluation under in-service conditions. Although the latter method can perform online prediction, it ignores the physical interpretability of microstructure geometric information and has poor prediction accuracy. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, equipment, medium, and product for predicting the thermal conductivity of thermal barrier coatings, which can achieve online prediction of the thermal conductivity of thermal barrier coatings while ensuring prediction accuracy.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting the thermal conductivity of a thermal barrier coating, including: Obtain the property vector of the coating to be predicted. The property vector includes a mechanical feature vector, which is used to indicate the mechanical parameters of the indentation test of the thermal barrier coating. Based on the characteristic vector of the coating to be predicted, the predicted microstructure image of the coating to be predicted is determined; Based on the predicted microstructure image, determine the defect instance image corresponding to the target defect type; Based on the defect instance image, determine the index value of the defect instance image under the target statistical index; Based on the index values ​​of defect instance images under the target statistical index, the predicted thermal conductivity of the coating to be predicted is determined.

[0006] Secondly, this application provides a thermal barrier coating thermal conductivity prediction device, comprising: The acquisition module is used to acquire the characteristic vector of the coating to be predicted. The characteristic vector includes a mechanical feature vector, which is used to indicate the mechanical parameters of the indentation test of the thermal barrier coating. The image generation module is used to determine the predicted microstructure image of the coating to be predicted based on the characteristic vector of the coating to be predicted. The image segmentation module is used to determine the defect instance image corresponding to the target defect type based on the predicted microstructure image; The processing module is used to determine the index value of the defect instance image under the target statistical index based on the defect instance image; The prediction module is used to determine the predicted thermal conductivity of the coating to be predicted based on the index value of the defect instance image under the target statistical index.

[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the thermal barrier coating thermal conductivity prediction method of the first aspect described above.

[0008] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the thermal barrier coating thermal conductivity prediction method of the first aspect described above.

[0009] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the thermal barrier coating thermal conductivity prediction method described in the first aspect.

[0010] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method for predicting the thermal conductivity of thermal barrier coatings. The method involves determining the predicted microstructure image of the coating based on its characteristic vector; determining a defect instance image corresponding to the target defect type based on the predicted microstructure image; determining the index value of the defect instance image under the target statistical index based on the defect instance image; and determining the predicted thermal conductivity of the coating based on the index value of the defect instance image under the target statistical index. The entire prediction process does not require peeling the coating from the substrate; the predicted thermal conductivity of the coating can be obtained solely through its mechanical characteristic vector. This reduces the cost and difficulty of predicting the thermal conductivity of thermal barrier coatings and enables online prediction of the thermal conductivity of thermal barrier coatings. Meanwhile, since the predicted thermal conductivity of the coating is determined based on the target statistical index of the coating, and the target statistical index indicates the geometric information of the microstructure, the physical interpretability of the geometric information of the microstructure is combined when predicting the thermal conductivity, which can ensure the accuracy of the thermal conductivity prediction. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 An application environment diagram for the thermal conductivity prediction method of thermal barrier coatings provided in the embodiments of this application; Figure 2 A schematic flowchart illustrating the method for predicting the thermal conductivity of thermal barrier coatings provided in this application embodiment; Figure 3 A flowchart illustrating the training process of the target diffusion model provided in this application embodiment; Figure 4 A schematic diagram comparing two-point correlation data corresponding to different target regions provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating the relationship between FID value and training steps provided in an embodiment of this application. Figure 6 A comparative diagram of the first object-level geometry descriptor and the second object-level geometry descriptor provided in the embodiments of this application; Figure 7 A schematic diagram comparing real microstructure images and predicted microstructure images provided in the embodiments of this application; Figure 8 This is a schematic diagram comparing the predicted thermal conductivity with the actual thermal conductivity using different methods. Figure 9 A comparative diagram of the mean absolute percentage error (MAPE) for different methods; Figure 10 A schematic diagram of the functional modules of the thermal barrier coating thermal conductivity prediction device provided in the embodiments of this application; Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] The thermal conductivity prediction method for thermal barrier coatings provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the characteristic vector of the coating to be predicted to server 104. After receiving the characteristic vector, server 104 processes the data based on the characteristic vector to obtain the predicted thermal conductivity of the coating. Server 104 can then feed back the obtained predicted thermal conductivity of the coating to terminal 102. Furthermore, in some embodiments, the thermal conductivity prediction method for thermal barrier coatings can also be implemented independently by server 104 or terminal 102. For example, terminal 102 can directly process the characteristic vector of the coating to be predicted, or server 104 can obtain the characteristic vector of the coating to be predicted from the data storage system and process it.

[0016] In one exemplary embodiment, such as Figure 2 As shown, a method for predicting the thermal conductivity of a thermal barrier coating is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are taken: S201 to S208.

[0017] S201. Obtain the characteristic vector of the coating to be predicted.

[0018] Among them, the characteristic vector includes the mechanical characteristic vector, which is used to indicate the mechanical parameters of the indentation test of the thermal barrier coating.

[0019] For example, the polished cross section of the coating to be predicted can be tested using an Anton Paar MCT3 instrument with a micron-scale indentation. Following the ASTM C1327-03 standard, a Vickers indenter is used with parameters set as follows: maximum load 4.9 N, loading rate 10 N / min, and peak load holding time 10 s. Ten indentations are evenly distributed, with the indentation spacing not less than four times the diagonal length of the indentation. Finally, the hardness, elastic modulus, plastic depth, and elastic work percentage corresponding to each indentation are obtained. Then, based on the micron-indentation load-displacement curve of each point, determine the plastic depth and maximum indentation depth of that point, and calculate the ratio of the plastic depth to the maximum indentation depth corresponding to that point; take the average of the ratios of the plastic depth to the maximum indentation depth corresponding to 10 points to obtain the mean value of the ratio of plastic depth to maximum indentation depth. Then calculate the mean hardness, mean elastic modulus, mean ratio of hardness to elastic modulus, mean percentage of elastic work, coefficient of variation of hardness, and coefficient of variation of elastic modulus for the 10 points. Based on the mechanical parameters of the indentation test, a feature vector is constructed and standardized by z-score to obtain the mechanical feature vector.

[0020] S202. Based on the characteristic vector of the coating to be predicted, determine the predicted microstructure image of the coating to be predicted.

[0021] For example, a pre-trained image generation model can be used to determine the predicted microstructure image based on the feature vector. The pre-trained image generation model can be an existing image generation model trained based on the feature vector and the real microstructure image corresponding to the feature vector. The training process is a conventional technique in this field and will not be described in detail here.

[0022] For example, multiple image generation can be performed on the feature vector of the coating to be predicted based on a pre-trained image generation model to obtain multiple predicted microstructure images of the coating to be predicted.

[0023] S203. Based on the predicted microstructure image, determine the defect instance image corresponding to the target defect type.

[0024] For example, the target defect type may include three categories: crack, equiaxed porosity, and unmelted region; the defect instance image may include crack instance image, equiaxed porosity instance image, and unmelted region instance image.

[0025] It is understandable that multiple defect instance images may be identified in the predicted microstructure image.

[0026] For example, based on existing semantic segmentation methods for the microstructure of thermal barrier coatings, the corresponding defect instance image of the target defect type can be determined from the predicted microstructure image; there are no restrictions on the specific semantic segmentation method, as long as it can determine the corresponding defect instance image of the target defect type from the microstructure image.

[0027] For example, the semantic segmentation method for the microstructure of thermal barrier coatings can be the semantic segmentation method disclosed in patent publication number CN117876304A and patent publication number CN121582564A. The semantic segmentation method will not be described in detail here.

[0028] It should be noted that when there are multiple predicted microstructure images, it is possible to determine the defect instance image corresponding to the target defect type in all predicted microstructure images.

[0029] S204. Based on the defect instance image, determine the index value of the defect instance image under the target statistical index.

[0030] For example, the target statistical index can be a mathematical statistical index of the descriptive parameters corresponding to the defect instance.

[0031] It can be understood that the description parameters are the geometric information description of the defect instance, which may include the orientation, connectivity, branching characteristics, skeleton length, density and width statistics of the crack instance; the equivalent diameter and nearest neighbor spacing statistics of the equiaxed pore instance; the roundness and aspect ratio of the unmelted region instance; the description parameters corresponding to the defect instance of different defect types may be different, and there are no restrictions on the description parameters.

[0032] For any given descriptive parameter, all values ​​of that descriptive parameter in all defect instance images can be recorded, and the mean, standard deviation, maximum, minimum, median, etc. of all values ​​can be used as target statistical indicators.

[0033] For example, the target statistical index can be set manually based on experience. For instance, the target statistical index may include "the mean of all crack lengths" or "the maximum of all crack lengths". It can be understood that the target statistical index can indicate the physical geometric information of the defect instance image.

[0034] S205. Based on the index values ​​of the defect instance image under the target statistical index, determine the predicted thermal conductivity of the coating to be predicted.

[0035] For example, a pre-trained target thermal conductivity prediction model can be used to predict the thermal conductivity of the coating to be predicted based on the index value of the defect instance image under the target statistical index, and thus obtain the predicted thermal conductivity.

[0036] For example, the target thermal conductivity prediction model can be obtained by training an initial prediction model based on the index value of the actual thermal barrier coating under the target statistical index and the actual thermal conductivity of the actual thermal barrier coating.

[0037] For example, the initial prediction model can be a regression prediction model or a neural network model.

[0038] For example, the regression prediction model can be an existing gradient boosting regression prediction model or other regression prediction models, without limitation.

[0039] For example, the training process of a target thermal conductivity prediction model may include: Multiple thermal conductivity training samples were obtained, each thermal conductivity training sample including multiple real microstructure images of the thermal barrier coating and real thermal conductivity; For multiple real microstructure images, steps similar to those described in S203 to S204 can be performed, except that multiple predicted microstructure images are replaced with multiple real microstructure images to obtain the target statistical indicators corresponding to the multiple real microstructure images. The first feature vector is constructed based on the target statistical indicators corresponding to multiple real microstructure images; Using the first feature vector as the model input and the actual thermal conductivity as the model output, the initial regression prediction model is trained to obtain the target thermal conductivity prediction model.

[0040] It is understandable that the specific training process for regression prediction models is a conventional technique in this field, and will not be elaborated here.

[0041] For example, grid search combined with five-fold cross-validation can also be used to optimize the hyperparameters of the initial regression prediction model across multiple thermal conductivity training samples.

[0042] For example, the initial regression prediction model can specifically be an XGBoost regression prediction model.

[0043] For example, a second feature vector can be constructed based on the statistical indices of the description parameters corresponding to various target description parameters. The second feature vector is then input into the XGBoost regression prediction model to obtain the predicted thermal conductivity of the coating to be predicted.

[0044] This application provides a method for predicting the thermal conductivity of thermal barrier coatings. The method involves determining the predicted microstructure image of the coating based on its characteristic vector; determining a defect instance image corresponding to the target defect type based on the predicted microstructure image; determining the index value of the defect instance image under the target statistical index based on the defect instance image; and determining the predicted thermal conductivity of the coating based on the index value of the defect instance image under the target statistical index. The entire prediction process does not require peeling the coating from the substrate; the predicted thermal conductivity of the coating can be obtained solely through its mechanical characteristic vector. This reduces the cost and difficulty of predicting the thermal conductivity of thermal barrier coatings and enables online prediction of the thermal conductivity of thermal barrier coatings. Meanwhile, since the predicted thermal conductivity of the coating is determined based on the target statistical index of the coating, and the target statistical index indicates the defect geometric information of the microstructure, the physical interpretability of the microstructure geometric information is combined when predicting thermal conductivity, which can ensure the accuracy of the thermal conductivity prediction.

[0045] In some possible implementations, the target statistical indicator is selected from all statistical indicators using a screening strategy based on the importance of random forest features.

[0046] It is understandable that if there are 50 descriptive parameters corresponding to a defect instance, and each descriptive parameter corresponds to 5 statistical indicators (such as mean, standard deviation, maximum value, minimum value, and median), when all statistical indicators are determined as target statistical indicators, the dimension of the aforementioned first feature vector is 50 × 5 = 250 dimensions, that is, the input dimension of the initial regression prediction model is 250 dimensions. At this time, if the initial regression prediction model is directly trained in a small sample scenario, the model will easily remember the random noise and accidental correlations in the thermal conductivity training set, resulting in severe overfitting. This leads to high accuracy in thermal conductivity prediction under the thermal conductivity training set, but a sharp drop in accuracy for new coatings to be predicted. Therefore, in the case of small sample size, this embodiment can use a random forest feature importance selection strategy to determine the importance score of all statistical indicators by using each first feature vector and the corresponding thermal conductivity label (i.e., the corresponding true thermal conductivity). The top k (which can be 10) statistical indicators that meet the importance threshold (i.e. are greater than the preset importance threshold) are taken as target indicators. Then, the aforementioned first feature vector is k-dimensional, which means that the input dimension of the initial regression prediction model is k-dimensional, which can avoid model overfitting in small sample size scenarios from the root.

[0047] To improve the accuracy of generating predicted microstructure images, in some possible embodiments, multiple predicted microstructure images of the coating to be predicted are determined based on the characteristic vector of the coating to be predicted, including: Based on the characteristic vector of the coating to be predicted, multiple predicted microstructure images of the coating to be predicted are determined based on the target diffusion model; the target diffusion model is used to take the characteristic vector as a conditional input and output the predicted microstructure image corresponding to the mechanical feature vector. Among them, such as Figure 3 As shown, the training process of the target diffusion model may include the following steps S301 to S302.

[0048] S301. Obtain the training sample set.

[0049] The training sample set includes multiple training samples, each of which includes a characteristic vector of the thermal barrier coating and a real microstructure image.

[0050] For example, the construction of a mechanical feature vector includes: Obtain micron-level indentation test data at multiple points on the thermal barrier coating; Based on micron-level indentation test data from multiple locations, the mechanical parameters of the indentation test are determined. These mechanical parameters include the mean hardness, mean elastic modulus, mean ratio of hardness to elastic modulus, mean ratio of plastic depth to maximum indentation depth, mean percentage of elastic work, coefficient of variation of hardness, and coefficient of variation of elastic modulus. Based on the mechanical parameters of the indentation test, a feature vector is constructed and z-score normalized to obtain the mechanical feature vector.

[0051] For example, atmospheric plasma spraying (APS) can be used to prepare 50 thermal barrier coating samples using commercially available 8 wt.% yttrium-stabilized zirconium oxide (8YSZ) powder as the spraying material and Hastelloy-X high-temperature alloy as the substrate. During the preparation process, the following spraying process parameters were systematically changed to cover a wide range of microstructure distributions: spraying distance (70–105 mm), powder feed rate (14.0–14.6 g / min), current (380–420 A), and argon main gas flow rate (90–125 SLPM); the resulting coating thickness ranged from 280 to 320 μm. For example, the polished cross-section of each thermal barrier coating sample can be tested using an Anton Paar MCT3 instrument with a micron-scale indentation array. Following the ASTM C1327-03 standard, a Vickers indenter is used with parameters set as follows: maximum load 4.9 N, loading rate 10 N / min, and peak load holding time 10 s. Ten indentation points are evenly distributed on each thermal barrier coating sample, with the indentation spacing not less than four times the diagonal length of the indentation. Finally, the hardness, elastic modulus, plastic depth, and elastic work percentage corresponding to each point are obtained. Then, based on the micron-indentation load-displacement curve of each point, determine the plastic depth and maximum indentation depth of that point, and calculate the ratio of the plastic depth to the maximum indentation depth corresponding to that point; take the average of the ratios of the plastic depth to the maximum indentation depth corresponding to 10 points to obtain the mean value of the ratio of plastic depth to maximum indentation depth. Then calculate the mean hardness, mean elastic modulus, mean ratio of hardness to elastic modulus, mean percentage of elastic work, coefficient of variation of hardness, and coefficient of variation of elastic modulus for the 10 points. Based on the mechanical parameters of the indentation test, a feature vector is constructed and standardized by z-score to obtain the mechanical feature vector.

[0052] For example, a TESCAN RISE-MAGNA scanning electron microscope can be used to characterize the polished cross-section of the same thermal barrier coating sample in BSE mode, with an accelerating voltage of 20 kV and a working distance of 10 mm. Ten non-overlapping images at random locations are acquired on each sample cross-section, with the original image resolution being 3072×3072 pixels. Next, image preprocessing was performed: the original image was randomly cropped into 512×512 pixel blocks, requiring that the proportion of structural pixels in the cropped area be no less than 30% (grayscale threshold I>20). If this requirement was not met, center cropping was used; the grayscale values ​​were linearly normalized to [-1,1]; a total of 500 real microstructure images were obtained from 50 thermal barrier coating samples.

[0053] For example, each real microstructure image can be paired with its corresponding mechanical feature vector to obtain training samples, and thus a training sample set.

[0054] S302. Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the original data to train the initial DDPM network to obtain the target diffusion model.

[0055] For example, the initial DDPM network can be an existing conditional DDPM model. During training, the feature vector is used as the model condition information, and the real microstructure image is used as the raw data for model training to obtain the target diffusion model. It can be understood that the raw data refers to the target ground truth during model training.

[0056] It should be noted that the specific network structure of the conditional DDPM model and the training of the conditional DDPM model are conventional techniques in this field and will not be elaborated here.

[0057] It is understandable that after the initial DDPM is trained, it can be validated using a validation sample set. Once the initial DDPM meets the validation criteria, it is determined to be the target diffusion model. It should be noted that the training sample set and validation sample set are different depending on the coating sample they correspond to. All real microstructure images and mechanical feature vectors corresponding to the same coating sample are only included in one of the training sample set or the validation sample set, to avoid the data corresponding to the same coating sample appearing in both the training sample set and the validation sample set at the same time.

[0058] This application trains the initial DDPM network with a training sample set including mechanical feature vectors and real microstructure images to obtain a target diffusion network. This allows the target diffusion network to learn the complex mapping relationship between mechanical feature vectors and microstructure images, thereby enabling it to quickly generate predicted microstructure images based on mechanical feature vectors. Microstructure images can be quickly obtained without SEM cross-sectional sample preparation, thus improving the efficiency of microstructure image acquisition. In some possible implementations, the initial DDPM network includes a backbone network, which includes multiple downsampling units and multiple upsampling units; The initial DDPM network is used to add noise to the input real microstructure image at a time step, resulting in a noisy microstructure image. The input mechanical feature vector is mapped to a mechanical embedding vector, and the mechanical embedding vector is linearly projected to obtain multiple first single-channel feature maps. The scale of the multiple first single-channel feature maps corresponds to the output feature scale of multiple downsampling units and the input feature scale of multiple upsampling units. The first single-channel feature maps are used to perform feature concatenation with the output features of the corresponding downsampling units or with the input features of the corresponding upsampling units. The noisy microstructure image, the time step information corresponding to the noisy microstructure image, and the multiple first single-channel feature maps are input into the backbone network for feature processing to obtain the predicted noise corresponding to the noisy microstructure image. Based on the predicted noise corresponding to the noisy microstructure image and the real noise corresponding to the noisy microstructure image, the noise prediction loss is determined; the initial DDPM network is updated with the goal of minimizing the noise prediction loss.

[0059] For example, the backbone network of the initial DDPM network can be a U-Net network, including an initial convolutional unit, six downsampling units, a bottleneck layer, six upsampling units, and an output convolutional unit connected in sequence. Each upsampling unit is connected to a downsampling path feature at the corresponding scale via a skip connection.

[0060] It should be noted that the specific network structure of the DDPM network backbone is existing technology and will not be elaborated here.

[0061] For example, the forward process of the initial DDPM network can be defined as a Markov chain that progressively adds Gaussian noise to the data: In the formula, This represents the noisy microstructure image at time step t-1. Noisy microstructure image up to time step t The conditional probability distribution of the forward process; Indicates a Gaussian distribution; Let be the noise variance scheduling parameter at time step t, and let linear scheduling be used; Represents the identity matrix; T can take the value 1000; , ; The noisy microstructure image at time step t. It can be represented as: In the formula, Represents a true microscopic structure image; Indicates Gaussian noise; This represents the cumulative product of the signal retention coefficients from step 1 to step t, which represents the total retention ratio of the original signal after t steps. Let represent the noise variance scheduling parameter at time step s, using linear scheduling, where s is a cumulative variable from 1 to t.

[0062] For example, the input mechanical feature vector can be mapped to a mechanical embedding vector by a first multilayer perceptron; the first multilayer perceptron can be a two-layer MLP, including an input layer (both input and output dimensions are 7 (the dimensions of the aforementioned mechanical feature vector)), a linear layer (7-dimensional input and 256-dimensional output), SiLU activation, a linear layer (256-dimensional input and 256-dimensional output), and SiLU activation.

[0063] For example, the mechanical embedding vector can be linearly projected by multiple first linear projection layers to obtain multiple first single-channel feature maps. The multiple first linear projection layers can be mutually independent fully connected layers without activation functions. It can be understood that when performing linear projection, the parameters of the fully connected layers are learned independently according to the target resolution (the output feature scale of the corresponding downsampling unit and the input feature scale of the corresponding upsampling unit for the first single-channel feature map), so that the conditional information (i.e., the mechanical feature vector) can participate in the calculation of the feature map at different spatial scales with matching spatial resolution.

[0064] It is understandable that the number of the first single-channel feature map is the same as the number of upsampling units and the number of downsampling units (both are 6).

[0065] It can be understood that after the first single-channel feature map is concatenated with the output feature of the corresponding downsampling unit or with the input feature of the corresponding upsampling unit, it will be used as the input feature of the next unit in the network. In this way, the first single-channel feature map can be injected into the feature maps of each scale of the downsampling path and the upsampling path of the backbone network, which can realize the multi-scale transmission and control of conditional information (i.e., mechanical feature vector) in each layer of the network.

[0066] For example, the mean square error between the predicted noise and the real noise can be calculated based on the real noise corresponding to the noisy microstructure image and the predicted noise corresponding to the noisy microstructure image, thus obtaining the noise prediction loss.

[0067] For example, the network parameters of the initial DDPM network can be updated in reverse with the goal of minimizing the noise prediction loss until the network converges or reaches a preset number of iterations to obtain the target diffusion model.

[0068] This embodiment trains the initial DDPM network with a training sample set including mechanical feature vectors and real microstructure images to obtain a target diffusion network. This allows the target diffusion network to learn the complex mapping relationship between mechanical feature vectors and microstructure images, thereby enabling it to quickly generate predicted microstructure images based on mechanical feature vectors. Microstructure images can be quickly obtained without SEM cross-sectional sample preparation, thus improving the efficiency of microstructure image acquisition. Simultaneously, the mechanical feature vector characterizing the mechanical properties of the thermal barrier coating is mapped into a mechanical embedding vector through a first multilayer perceptron. Linear projection yields multiple first single-channel feature maps corresponding to the output feature scales of multiple downsampling units and the input feature scales of multiple upsampling units. These first single-channel feature maps are then used for feature concatenation with the output features of the corresponding downsampling units or with the input features of the corresponding upsampling units. This enables multi-scale transmission and control of the mechanical feature vector of the thermal barrier coating across the backbone network layers. Consequently, the multiphase microstructure features at both the coarse-scale (overall distribution of unmelted regions) and fine-scale (crack width, pore size) levels are effectively constrained, achieving accurate reproduction of the complex multiphase statistical characteristics of the thermal barrier coating and thus improving the generation accuracy of the microstructure image.

[0069] In some possible embodiments, the characteristic vector also includes a process parameter vector, which is used to indicate the preparation process parameters of the thermal barrier coating; The initial DDPM network is also used to map the input process parameter vector into a process embedding vector. The process embedding vector is linearly projected to obtain multiple second single-channel feature maps. The scale of the multiple second single-channel feature maps corresponds to the output feature scale of multiple downsampling units and the input feature scale of multiple upsampling units. The second single-channel feature maps are used to perform feature concatenation with the output features of the corresponding downsampling units or with the input features of the corresponding upsampling units. The noisy microstructure image, the time step information corresponding to the noisy microstructure image, and the multiple second single-channel feature maps are input into the backbone network for feature processing to obtain the prediction noise corresponding to the noisy microstructure image.

[0070] For example, the construction of the process parameter vector includes: Obtain the preparation process parameters of the thermal barrier coating, including spraying distance, powder feeding rate, current, main gas flow rate, and auxiliary gas flow rate; Based on the preparation process parameters, a feature vector is constructed and z-score normalized to obtain the process parameter vector.

[0071] For example, the input process parameter vector can also be mapped to a process embedding vector using a second multilayer perceptron; the second multilayer perceptron can be a two-layer MLP, including an input layer (both input and output dimensions are 5 (the dimensions of the aforementioned process feature vector)), a linear layer (5-dimensional input, 128-dimensional output), SiLU activation, and a linear layer (128-dimensional input, 256-dimensional output). For example, the mechanical embedding vector can also be linearly projected by multiple second linear projection layers to obtain multiple second single-channel feature maps. The multiple second linear projection layers can be mutually independent fully connected layers without activation functions. It can be understood that when performing linear projection, the parameters of the fully connected layers are learned independently according to the target resolution (the output feature scale of the corresponding downsampling unit and the input feature scale of the corresponding upsampling unit for the second single-channel feature map), so that the conditional information (i.e., the process parameter vector) can participate in the calculation of the feature map at different spatial scales with matching spatial resolution.

[0072] It is understandable that the number of the second single-channel feature map is the same as the number of upsampling units and the number of downsampling units (both are 6).

[0073] It is understandable that injecting the second single-channel feature map into the feature maps of each scale of the downsampling path and upsampling path of the backbone network can realize the multi-scale transmission and control of conditional information (i.e. process parameter vector) in each layer of the network.

[0074] It should be noted that when training the initial DDPM network at this time, one of the conditional modes (mechanical feature vector or process parameter vector) is randomly selected as the network input for each training batch, the input of the other mode is set to empty, and the gradients of the corresponding multilayer perceptron and linear projection layer are frozen, while the backbone network is updated under all conditional modes.

[0075] For example, the first batch is trained with 5 mechanical feature vectors, and the second batch is trained with 5 process parameter vectors. During the first batch of training, the second multilayer perceptron is frozen, and the first multilayer perceptron, the first linear projection layer, and the backbone network are updated. During the second batch of training, the first multilayer perceptron is frozen, and the second multilayer perceptron, the second linear projection layer, and the backbone network are updated.

[0076] In this way, the target diffusion network can learn the complex mapping relationship between the process parameter vector and the microstructure image, thereby quickly generating the predicted microstructure image based on the process parameter vector. When the thermal barrier coating to be predicted does not have a corresponding mechanical feature vector, the predicted microstructure image can be obtained directly through the process parameter vector corresponding to the preparation process parameters of the thermal barrier coating.

[0077] In some possible embodiments, based on the training sample set, using feature vectors as conditional information and real microstructure images as raw data, the initial DDPM network is trained to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the model to be verified. Obtain the mechanical feature vectors and real microstructure images of the thermal barrier coating to be verified; Based on the mechanical characteristic vector of the thermal barrier coating to be verified, and based on the model to be verified, the predicted microstructure image of the thermal barrier coating to be verified is determined. Based on the actual microstructure image of the thermal barrier coating to be verified, the first two-point correlation data of the thermal barrier coating to be verified in the target structural region are determined based on the two-point correlation function. Based on the predicted microstructure image of the thermal barrier coating to be verified, the second two-point correlation data of the thermal barrier coating to be verified in the target structural region is determined based on the two-point correlation function. When the first two points of relevant data and the second two points of relevant data meet the first preset condition, the model to be verified is determined as the target diffusion model.

[0078] For example, the determination of the mechanical feature vector and the real microstructure image of the thermal barrier coating to be verified is similar to the construction method of the aforementioned training samples, and will not be repeated here.

[0079] For example, the target structural region may include cracked regions, equiaxed porous regions, and unmelted regions.

[0080] It is understandable that the two-point correlation function describes the probability that two randomly selected pixels with a distance of r in a microstructure image belong to a specific phase (target region). It is a classic method for quantitatively characterizing the spatial statistical features of multiphase materials.

[0081] For example, for any target region, a probability function (i.e., two-point correlation data) that varies with the distance r between the two points can be obtained based on the two-point correlation function: In the formula, Indicates an indicator function; This means that if pixel x belongs to the target region, the value is 1; otherwise, the value is 0. The value is 1 if a pixel is at a distance r from pixel x and belongs to the target region; otherwise, the value is 0. This indicates that the statistical average is taken over all locations in the entire microstructure image.

[0082] For example, efficient computation can be achieved through Fast Fourier Transform, and a one-dimensional value can be obtained by radial averaging. A hierarchical aggregation strategy was adopted. First, the mean value was taken within the same thermal barrier coating sample to obtain the mean curve of that sample. Then, the curves of all test samples were aggregated across samples to obtain the confidence interval envelope diagram of "mean ± standard deviation".

[0083] It is understandable that the mean curve and confidence interval envelope plot are two-point correlation data.

[0084] It is understandable that the obtained two-point correlation data can characterize the size, volume fraction and spatial distribution of defect phases such as pores and cracks, and are key statistics for evaluating the physical usability of generated microstructures.

[0085] For example, the first preset condition can be that, for any target region, the similarity between the second two-point correlation data corresponding to the predicted microstructure image and the first two-point correlation data corresponding to the real microstructure image meets a preset similarity threshold.

[0086] It is understandable that when the first two related data points and the second two related data points meet the first preset condition, it can be considered that the generated predicted microstructure image is statistically equivalent to the real microstructure image. The mechanical and thermal performance results obtained by substituting the predicted microstructure image into simulation tools such as finite element method are consistent with the results corresponding to the real microstructure image. Therefore, the initial DDPM network training is completed.

[0087] It is understandable that when the data related to the first two points and the data related to the second two points do not meet the first preset condition, network training needs to continue until the first preset condition is met.

[0088] For example, under the mechanical characteristics of five verification samples (thermal barrier coatings to be verified), 10 predicted microstructure images were generated for each sample using a target diffusion model. Two-point correlation functions between the actual and generated images were calculated for the crack region, equiaxed porosity region, and unmelted region, respectively. The first and second two-point correlation data are shown below. Figure 4 As shown.

[0089] refer to Figure 4 It can be seen that the initial value of the two-point correlation function of the equiaxed porous phase is about 0.06, which rapidly decays to a low plateau at r ≈ 50 pixels. The actual and generated curves and the confidence interval envelope almost completely overlap across the entire range. The two-point correlation function value of the crack phase decays more slowly, with an initial value of about 0.055 to 0.060, and tends to plateau at r ≈ 100 pixels. The actual and generated curves match well. The two-point correlation function value of the unmelted region phase decays the most rapidly. Both curves converge to near zero outside r = 150 pixels, but the generated curve decays slightly faster than the actual curve in the r = 50 to 150 pixel range. The overlap of the confidence interval envelope is lower than that of the other two phases, reflecting the complex morphology and large variability among samples of this phase. Overall, the three phases... The mean curve and confidence interval envelope are in good agreement with the real image, verifying the physical usability of the generated microstructure at the spatial statistical properties level under the mechanical feature condition mode.

[0090] In some possible embodiments, based on the training sample set, using feature vectors as conditional information and real microstructure images as raw data, the initial DDPM network is trained to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the model to be verified. Obtain the mechanical feature vectors and real microstructure images of the thermal barrier coating to be verified; Based on the mechanical characteristic vector of the thermal barrier coating to be verified, and based on the model to be verified, the predicted microstructure image of the thermal barrier coating to be verified is determined. Calculate the FID value between the actual microstructure image and the predicted microstructure image of the thermal barrier coating to be verified; When the FID value meets the preset threshold, the model to be verified is determined as the target diffusion model.

[0091] For example, the preset threshold can be 105, and satisfying the preset threshold means that the FID value is less than 105.

[0092] It is understandable that when the FID value does not meet the preset threshold, network training needs to continue until the preset threshold is met.

[0093] Continuing with the above 5 validation samples as an example, the relationship between the obtained FID value and the number of training steps is as follows: Figure 5 As shown.

[0094] refer to Figure 5 As can be seen, the FID value was 301.6 in the early stage of model training, and dropped to 103.7 after 80,000 training steps, gradually approaching the baseline of the distribution within the real data (the FID value between real microstructure images is about 50.5). The continuous narrowing of the FID value of the predicted microimage from the real baseline indicates that the model continuously improves its ability to capture the statistical features of real microstructures as the training process progresses, and the mechanical feature vector can effectively constrain the overall distribution characteristics of the generated image.

[0095] In some possible embodiments, based on the training sample set, using feature vectors as conditional information and real microstructure images as raw data, the initial DDPM network is trained to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the model to be verified. Obtain the mechanical feature vectors and real microstructure images of the thermal barrier coating to be verified; Based on the mechanical characteristic vector of the thermal barrier coating to be verified, and based on the model to be verified, the predicted microstructure image of the thermal barrier coating to be verified is determined. Extract the first object-level geometric descriptor from the real microstructure image of the thermal barrier coating to be verified, and the second object-level geometric descriptor from the predicted microstructure image. The object-level geometric descriptor may include crack orientation and skeleton length / density, equivalent diameter and nearest neighbor distance of equiaxed pores, and circularity and aspect ratio of unmelted regions. When the first object-level geometric descriptor and the second object-level geometric descriptor meet the second preset condition, the model to be verified is determined as the target diffusion model.

[0096] For example, the second preset condition can be that the first object-level geometry descriptor and the second object-level geometry descriptor are of the same or similar type: For crack orientation, the second preset condition includes the same orientation; for skeleton length / density, the second preset condition includes the length difference being less than a preset length difference threshold and the density difference being less than a preset density difference threshold; for equiaxed pore equivalent diameter, the second preset condition includes the equivalent diameter difference being less than a preset equivalent diameter difference threshold; for equiaxed pore nearest neighbor distance, the second preset condition includes the nearest neighbor distance difference being less than a preset nearest neighbor distance difference threshold; for unmelted region roundness, the second preset condition includes the roundness difference being less than a preset roundness difference threshold; for unmelted region aspect ratio, the second preset condition includes the aspect ratio difference being less than a preset aspect ratio difference threshold.

[0097] It is understandable that when the first object-level geometric descriptor and the second object-level geometric descriptor do not meet the second preset condition, network training needs to continue until the second preset condition is met.

[0098] Continuing with the above five verification samples as examples, the obtained first object-level geometry descriptor and second object-level geometry descriptor are as follows: Figure 6 As shown.

[0099] refer to Figure 6 It can be seen that, at the geometric level, the crack orientation distribution exhibits a nearly uniform distribution from -90° to +90° in both the real and generated images; the equivalent diameter distribution of equiaxed pores, the nearest neighbor distance distribution, and the circularity distribution of unmelted regions are highly consistent with the real images in terms of peak position, distribution width, and overall morphology. These quantitative results demonstrate that the predicted microstructure images not only match the real microstructure images in terms of phase volume fraction but also possess physical usability in terms of the geometric morphological statistical characteristics of each defect phase.

[0100] In some possible embodiments, based on the training sample set, using feature vectors as conditional information and real microstructure images as raw data, the initial DDPM network is trained to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the model to be verified. Obtain the mechanical feature vectors and real microstructure images of the thermal barrier coating to be verified; Based on the mechanical characteristic vector of the thermal barrier coating to be verified, and based on the model to be verified, the predicted microstructure image of the thermal barrier coating to be verified is determined. Based on the actual microstructure image of the thermal barrier coating to be verified, the first two-point correlation data of the thermal barrier coating to be verified in the target structural region is determined based on the two-point correlation function; based on the predicted microstructure image of the thermal barrier coating to be verified, the second two-point correlation data of the thermal barrier coating to be verified in the target structural region is determined based on the two-point correlation function. Calculate the FID value between the actual microstructure image and the predicted microstructure image of the thermal barrier coating to be verified; Extract the first object-level geometric descriptor from the real microstructure image of the thermal barrier coating to be verified, and the second object-level geometric descriptor from the predicted microstructure image. The object-level geometric descriptor may include crack orientation and skeleton length / density, equivalent diameter and nearest neighbor distance of equiaxed pores, and circularity and aspect ratio of unmelted regions. When the first two-point related data and the second two-point related data meet the first preset condition, the FID value meets the preset threshold, and the first object-level geometric descriptor and the second object-level geometric descriptor meet the second preset condition, the model to be verified is determined as the target diffusion model.

[0101] In this way, when the first preset condition, the FID threshold, and the second preset condition are all met, the model to be verified can be determined as the target diffusion model, thereby maximizing the guarantee that the generation accuracy of the target diffusion model meets the requirements of practical applications.

[0102] It is understandable that any two of the first preset condition, the FID threshold, and the second preset condition can be satisfied to determine the model to be verified as the target diffusion model, without any restrictions.

[0103] For example, consistency comparison and verification of generation capabilities under different conditions (i.e., mechanical feature vectors and process parameter vectors) can also be performed.

[0104] In some possible embodiments, based on the training sample set, using feature vectors as conditional information and real microstructure images as raw data, the initial DDPM network is trained to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network, thus obtaining the trained initial DDPM network. The target shadow parameters are used as the network parameters of the initial DDPM network after training to obtain the target diffusion model; The target shadow parameter is determined based on the network parameters during the training process using the exponential moving average method, from the initial DDPM network training to the completion of training.

[0105] For example, the exponential moving average coefficient of the exponential moving average method can be 0.999.

[0106] It is understandable that during the training phase, instead of using the exponential moving average method to directly update the network parameters of the initial DDPM network, the shadow parameters are calculated synchronously after the initial DDPM network completes a gradient update. The shadow parameters are only passively calculated and updated and do not participate in the training process. When the initial DDPM network is trained, the shadow parameters are also calculated and updated, and the target shadow parameters are obtained. At this time, the initial DDPM network completely discards its current network parameters and loads the target shadow parameters as network parameters to obtain the target diffusion model.

[0107] It should be noted that the exponential moving average method is a conventional technique in this field and will not be elaborated upon here.

[0108] It is understandable that in actual implementation, when performing verification of two-point correlation data / FID value / object-level geometric descriptor, the model to be verified that meets the first preset condition, the FID threshold, or the second preset condition is the initial DDPM network that has been trained.

[0109] Thus, during the training phase of the target diffusion model, the original network parameters can be used for gradient updates to ensure the accuracy of the training gradients. Meanwhile, in the subsequent inference phase, the low volatility of the target shadow parameters calculated by the exponential moving average method can be utilized to improve the generation accuracy of predicted microstructure images in small sample scenarios.

[0110] It is understandable that, during the inference phase of the target diffusion model, Gaussian noise... Start by inputting the given conditions. The following iterative denoising process is performed based on the DDPM reverse denoising process: Obtaining denoised images step by step (i.e., predicting microstructure images), which are then inversely normalized for visualization and downstream analysis.

[0111] It is understandable that during the training phase, when the feature vector also includes the process parameter vector, the target diffusion model obtained by the aforementioned model training method can also gradually obtain the predicted microstructure image by iteratively denoising according to the process parameter vector corresponding to the coating to be predicted and the reverse denoising process of DDPM. It should be noted that during the inference phase, the mechanical feature vector and the process parameter vector are not the joint input of the target diffusion model, and the conditional input of the target diffusion model can only be one of the two.

[0112] Figure 7 This is a schematic diagram comparing the actual microstructure image and the predicted microstructure image provided in the embodiments of this application.

[0113] To illustrate the significant advancements of the method presented in this application, five independent test samples were selected, covering the entire range of thermal conductivity in the dataset (0.7–1.5 W·m). -1 ·K -1 The predicted thermal conductivity of five test samples was obtained using different methods (including direct mechanical feature regression (path 1), prediction method based on real microstructure images (path 2), and the method of this application (path 3)). These predicted thermal conductivity values ​​were then compared with the actual thermal conductivity measured using a Netzsch LFA 457 laser flare analyzer, and the corresponding mean absolute percentage error (MAPE) was calculated. The final test results can be referenced. Figure 8 and Figure 9 .

[0114] Figure 8 This diagram illustrates the comparison between predicted and actual thermal conductivity using different methods. Figure 9 This is a comparative diagram of the mean absolute percentage error (MAPE) for different methods.

[0115] refer to Figure 8 It can be seen that the data points corresponding to path 2 (real image) and path 3 (generated image) are closely clustered near the diagonal (the predicted thermal conductivity is equal to the real thermal conductivity), indicating that the thermal conductivity predicted based on the real microstructure image and the generated microstructure image are highly consistent with the real thermal conductivity. In contrast, the data points corresponding to path 1 (direct prediction) are significantly more dispersed relative to the diagonal, and the deviation is more significant, indicating that the method of omitting the intermediate layer of microstructure and directly predicting the thermal conductivity by regression of mechanical feature vectors is difficult to accurately capture the complex mapping relationship between the mechanical features of the coating and the thermal conductivity.

[0116] refer to Figure 9 It can be seen that the mean absolute percentage error (MAPE) of path 3 is 10.36%, which is close to 9.06% of path 2, but much lower than 29.05% of path 1.

[0117] Based on the same inventive concept, this application also provides a thermal barrier coating thermal conductivity prediction device for implementing the above-described thermal barrier coating thermal conductivity prediction method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more thermal barrier coating thermal conductivity prediction device embodiments provided below can be found in the above-described limitations of the thermal barrier coating thermal conductivity prediction method, and will not be repeated here.

[0118] In one exemplary embodiment, such as Figure 10 As shown, a thermal barrier coating thermal conductivity prediction device is provided, comprising: The acquisition module 1001 is used to acquire the characteristic vector of the coating to be predicted. The characteristic vector includes a mechanical feature vector, which is used to indicate the mechanical parameters of the indentation test of the thermal barrier coating. Image generation module 1002 is used to determine the predicted microstructure image of the coating to be predicted based on the characteristic vector of the coating to be predicted. The image segmentation module 1003 is used to determine the defect instance image corresponding to the target defect type based on the predicted microstructure image; The processing module 1004 is used to determine the index value of the defect instance image under the target statistical index based on the defect instance image; The prediction module 1005 is used to determine the predicted thermal conductivity of the coating to be predicted based on the index value of the defect instance image under the target statistical index.

[0119] The specific implementation methods and beneficial effects of this device embodiment can be found in the foregoing method embodiments, and will not be repeated here.

[0120] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 11 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements the aforementioned method for predicting the thermal conductivity of thermal barrier coatings.

[0121] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0123] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0124] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0125] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0126] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0127] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0129] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the thermal conductivity of a thermal barrier coating, characterized in that, The method for predicting the thermal conductivity of the thermal barrier coating includes: Obtain the characteristic vector of the coating to be predicted, the characteristic vector including the mechanical feature vector, which is used to indicate the indentation test mechanical parameters of the thermal barrier coating; Based on the characteristic vector of the coating to be predicted, the predicted microstructure image of the coating to be predicted is determined. Based on the predicted microstructure image, determine the defect instance image corresponding to the target defect type; Based on the defect instance image, determine the index value of the defect instance image under the target statistical index; Based on the index values ​​of defect instance images under the target statistical index, the predicted thermal conductivity of the coating to be predicted is determined.

2. The method for predicting the thermal conductivity of thermal barrier coatings according to claim 1, characterized in that, The step of determining multiple predicted microstructure images of the coating to be predicted based on the characteristic vector of the coating to be predicted includes: Based on the characteristic vector of the coating to be predicted, and using the target diffusion model, multiple predicted microstructure images of the coating to be predicted are determined; the target diffusion model is used to take the characteristic vector as a conditional input and output the predicted microstructure image corresponding to the mechanical feature vector. The training process of the target diffusion model includes: Obtain a training sample set; the training sample set includes multiple training samples, each training sample including a characteristic vector of the thermal barrier coating and a real microstructure image; Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the target diffusion model. The initial DDPM network includes a backbone network, which includes multiple downsampling units and multiple upsampling units; The initial DDPM network is used to add noise to the input real microstructure image at a time step, resulting in a noisy microstructure image. The input mechanical feature vector is mapped to a mechanical embedding vector based on a first multilayer perceptron. The mechanical embedding vector is then linearly projected to obtain multiple first single-channel feature maps. The scale of the multiple first single-channel feature maps corresponds to the output feature scale of multiple downsampling units and the input feature scale of multiple upsampling units. The first single-channel feature maps are used for feature concatenation with the output features of the corresponding downsampling units or with the input features of the corresponding upsampling units. The noisy microstructure image, the time step information corresponding to the noisy microstructure image, and the multiple first single-channel feature maps are input into the backbone network for feature processing to obtain the predicted noise corresponding to the noisy microstructure image. Based on the predicted noise corresponding to the noisy microstructure image and the real noise corresponding to the noisy microstructure image, the noise prediction loss is determined; the initial DDPM network is updated with the goal of minimizing the noise prediction loss.

3. The method for predicting the thermal conductivity of thermal barrier coatings according to claim 2, characterized in that, The construction of the mechanical feature vector includes: Obtain micron-level indentation test data at multiple points on the thermal barrier coating; Based on micron-level indentation test data from multiple locations, the mechanical parameters of the indentation test are determined. These mechanical parameters include the average hardness, average elastic modulus, average hardness to elastic modulus ratio, average plastic depth to maximum indentation depth ratio, average elastic work percentage, hardness coefficient of variation, and elastic modulus coefficient of variation. Based on the mechanical parameters of the indentation test, a feature vector is constructed and z-score normalized to obtain the mechanical feature vector.

4. The method for predicting the thermal conductivity of thermal barrier coatings according to claim 2, characterized in that, The method involves training an initial DDPM network based on a training sample set, using feature vectors as conditional information and real microstructure images as raw data to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network and obtain the model to be verified. Obtain the mechanical feature vectors and real microstructure images of the thermal barrier coating to be verified; Based on the mechanical characteristic vector of the thermal barrier coating to be verified, and based on the model to be verified, the predicted microstructure image of the thermal barrier coating to be verified is determined. Based on the actual microstructure image of the thermal barrier coating to be verified, the first and second point correlation data of the thermal barrier coating to be verified in the target structural region are determined based on the two-point correlation function. Based on the predicted microstructure image of the thermal barrier coating to be verified, the second two-point correlation data of the thermal barrier coating to be verified in the target structural region is determined based on the two-point correlation function. When the first two points of relevant data and the second two points of relevant data meet the first preset condition, the model to be verified is determined as the target diffusion model.

5. The method for predicting the thermal conductivity of thermal barrier coatings according to claim 2, characterized in that, The method involves training an initial DDPM network based on a training sample set, using feature vectors as conditional information and real microstructure images as raw data to obtain a target diffusion model, including: Based on the training sample set, the feature vector is used as conditional information and the real microstructure image is used as the raw data to train the initial DDPM network, thus obtaining the trained initial DDPM network. The target shadow parameters are used as the network parameters of the initial DDPM network after training to obtain the target diffusion model; The target shadow parameter is determined based on the network parameters during the training process, using the exponential moving average method, during the training of the initial DDPM network until training is completed.

6. The method for predicting the thermal conductivity of thermal barrier coatings according to claim 1, characterized in that, The target statistical index is selected from all statistical indexes using a screening strategy based on the importance of random forest features.

7. A device for predicting the thermal conductivity of a thermal barrier coating, characterized in that, The thermal barrier coating thermal conductivity prediction device includes: The acquisition module is used to acquire the characteristic vector of the coating to be predicted. The characteristic vector includes a mechanical feature vector, which is used to indicate the indentation test mechanical parameters of the thermal barrier coating. The image generation module is used to determine the predicted microstructure image of the coating to be predicted based on the characteristic vector of the coating to be predicted. The image segmentation module is used to determine the defect instance image corresponding to the target defect type based on the predicted microstructure image; The processing module is used to determine the index value of the defect instance image under the target statistical index based on the defect instance image; The prediction module is used to determine the predicted thermal conductivity of the coating to be predicted based on the index value of the defect instance image under the target statistical index.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the thermal barrier coating thermal conductivity prediction method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the thermal conductivity prediction method for thermal barrier coatings as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the thermal conductivity prediction method for thermal barrier coatings as described in any one of claims 1-6.

Citation Information

Patent Citations

  • Thermal spraying coating structure analysis method based on dual-mode image and machine learning

    CN117876304A

  • Thermal barrier coating microstructure segmentation method, system, equipment, medium and product

    CN121582564A