Rare earth tantalate material screening method based on transfer learning and application

By constructing a deep neural network model based on transfer learning, predicting the thermal conductivity of rare earth tantalate materials, the problem of inefficiency in traditional methods is solved, efficient and accurate material screening is achieved, and experimental costs and time are reduced.

CN119028493APending Publication Date: 2024-11-26KUNMING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411079723.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Rare earth tantalate thermal barrier coating materials have many types of elements and complex proportions. Traditional material screening methods are inefficient and lack a material database for thermal conductivity, resulting in inefficient material design.

Method used

Using a transfer learning method, a high-precision deep neural network model is constructed. By collecting the thermal conductivity data of related materials, feature screening and weighting mathematical operations are carried out, neural network models are built to predict the thermal conductivity of rare earth tantalate materials, and to achieve Quick filter.

Benefits of technology

It improves the efficiency and accuracy of material design, reduces experimental costs and time, solves the problem of data sparsity, and significantly improves the accuracy and reliability of predictions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028493A_ABST
    Figure CN119028493A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of tantalate thermal barrier coatings, in particular to a rare earth tantalate material screening method based on transfer learning and application, a high-precision rare earth tantalate material thermal conductivity prediction model is established by utilizing a transfer learning technology, rapid prediction of thermal conductivity of multi-component RETaO4 is achieved, and the screening method is suitable for screening of high-precision rare earth tantalate materials. A multi-component rare earth tantalate material with low thermal conductivity is successfully found from millions of candidate materials, including 0.20% of Sm, 0.10% of Eu, 0.10% of Ho, 0.20% of Y, 0.40% of Yb, 0.10% of Sm, 0.10% of Eu, 0.10% of Y, 0.40% of Yb, 0.10% of Ce and 0.10% of Lu, and experimental verification is carried out. According to the method, a big data set for predicting the thermal conductivity is constructed by utilizing different system data sets with generality, the contradiction between a complex component space and a sparse small sample is solved, and a new thought is provided for the design of a multi-component rare earth tantalate material.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of tantalate thermal barrier coatings, and in particular to a rare earth tantalate material screening method based on transfer learning and an application thereof. Background Art

[0002] Thermal barrier coatings (TBCs) are important protective materials for aircraft engines and gas turbines, and are mainly composed of an alloy substrate, a bonding layer, a thermally grown oxide layer (TGO), and a top ceramic layer. They have the advantages of high temperature corrosion resistance and oxidation resistance, and have become an important method to increase the operating temperature of hot end components and extend their service life in recent years.

[0003] However, rare earth tantalate (RETaO4) thermal barrier coating materials involve many types of elements, and the combination of different types and proportions of elements constitutes a complex and high-dimensional search space. The traditional trial and error method is inefficient, and material screening based on the Slack equation is too simple. The computational component of the first-principles calculation is very expensive. There is an urgent need to develop more efficient and faster methods to accelerate material design.

[0004] In recent years, the development of artificial intelligence and materials informatics has achieved remarkable results, and data-driven machine learning technology is considered to be a feasible method in the field of material design. Machine learning methods can achieve rapid prediction of thermal conductivity, but there is currently no material database for the thermal conductivity of rare earth tantalate (RETaO4), and the types of elements involved in the literature are few and concentrated. The application of transfer learning technology in the field of materials provides a new idea for solving the sparse small sample problem faced by thermal barrier coating ceramic materials. Summary of the invention

[0005] The purpose of the present invention is to provide a rare earth tantalate material screening method and application based on transfer learning, to construct a high-precision deep neural network model, and to use the model to quickly predict in the composition space to achieve rapid screening of the thermal conductivity of rare earth tantalate materials.

[0006] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:

[0007] A rare earth tantalate material screening method based on transfer learning comprises the following steps:

[0008] S1: Collect the experimental data of thermal conductivity of each rare earth tantalate material at 20-900°C. The chemical formula of the collected data is RETaO4, where RE is a combination of one or more rare earth elements Y, Ce, Nd, Sm, Eu, Gd, Dy, Lu, and Ho, Ta represents tantalum, O represents oxygen, and the element dosage ratio of RE:Ta:O is 1:1:4;

[0009] S2: Collecting experimental data of thermal conductivity of rare earth niobate and rare earth zirconate materials at different temperatures. The chemical formulas of the collected data include but are not limited to A2B2O7, RETaO7, and RETa3O9;

[0010] S3: Taking the element dosage ratio as the weight, various weighted mathematical operations are performed on the element properties corresponding to the chemical formulas of RETaO4, A2B2O7, RETaO7, and RETaO9 to obtain weighted features, which are used as initial features;

[0011] S4: Feature screening was performed on the RETaO4 dataset. The Pearson correlation coefficient between the initial features was calculated, 0.85 was set as the strong correlation threshold, the features were divided into 9 groups, only one feature was retained in each group, redundant features were removed, and a feature pool was constructed. The calculation formula for the correlation coefficient was:

[0012]

[0013] Among them: x and y are the average values ​​of the two features in the sample, x i and i Represents the actual values ​​of the two features in the i-th sample;

[0014] S5: The optimal sub-features are selected through the feature recursive elimination method; the order of feature deletion from the first to the last is: 1. first ionization energy, 2. bulk modulus, 3. melting point, 4. ion radius ratio, 5. mass disorder parameter, 6. radius disorder parameter, and the final retained features are: ionic bond strength, temperature, and Clementi effective nuclear charge number;

[0015] S6: Build a neural network model. The model has a 4-layer structure, with an input dimension of 3 and an output dimension of 1. Use ReLU as the activation function, Adam as the optimizer, and the learning rate is set to 0.01. According to the optimal sub-features, the base model based on the rare earth tantalate RETaO4 dataset and the source domain model based on a large dataset composed of multiple material systems are trained respectively;

[0016] S7: The weights of the first two layers of the trained neural network source domain model are assigned to the first two layers of the base model, and the weights of the last two layers of the base model are initialized using Xavier to achieve the purpose of optimizing the neural network base model;

[0017] S8: For the 12 elements described in S1, select the corresponding number of elements from the 12 elements according to the 5-element combination, 6-element combination, 7-element combination, and 8-element combination, and arrange and combine them, design the component space, calculate the features and generate Excel, use the optimized neural network base model to read Excel, predict the thermal conductivity of all components, and arrange the components in ascending order from low to high according to the predicted thermal conductivity values.

[0018] S9: The low thermal conductivity component screened out by S8 is Sm 0.20 Eu 0.10 Ho 0.10 Y 0.20 Yb 0.40 TaO4、Sm 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 TaO4, experimentally prepare the material and verify the performance.

[0019] Furthermore, the rare earth tantalate data set RE in step S1 includes at least one piece of rare earth tantalate thermal conductivity data of a combination of two to ten elements.

[0020] Furthermore, in step S2, A is a combination of one or more rare earth elements Y, Ce, Nd, Sm, Eu, Gd, Dy, Ho, Er, Tm, Yb, and Lu, B represents a combination of one or more rare earth elements Ta, Nb, and Zr, O represents an oxygen element, and the atomic number of the oxygen element is 4, 7, or 9, so the element dosage ratio of A:B:O is 1:1:4, 1:1:7, 2:2:7, or 1:3:9;

[0021] Furthermore, the element properties in step S3 include but are not limited to: atomic number, atomic mass, period, group, density, melting point, bulk modulus, mass disorder parameter, Pauling electronegativity, AR electronegativity, MB electronegativity, first ionization energy, electron affinity, Clementi effective nuclear charge number, Slater effective nuclear charge number, charge-to-radius ratio, ionic bond strength, ratio of RE position to Ta ion radius, radius disorder parameter and covalent radius.

[0022] Furthermore, in step S4, the Pearson correlation coefficient is used to eliminate redundant features, and the calculation formula is:

[0023]

[0024] Among them: x and y are the average values ​​of the two features in the sample, x i and i Indicates the actual value of two features. According to the correlation threshold of 0.85, the features are divided into 9 groups, and one feature is retained in each group.

[0025] Furthermore, the feature recursive elimination method in step S5 specifically includes: first, using all initial features to train the model; second, evaluating the impact of each feature on the model's predictive ability, this step depends on the feature importance score provided by the model; then, based on these scores, identifying and removing the least important features; finally, repeating the above process until the predetermined number of features is reached or the model performance is no longer significantly improved. The order of feature deletion from the first to the last time is: 1. first ionization energy, 2. bulk modulus, 3. melting point, 4. ionic radius ratio, 5. mass disorder parameter, 6. radius disorder parameter, and the features finally retained are: ionic bond strength, temperature, Clementi effective nuclear charge number;.

[0026] Furthermore, in step S6, the neural network base model reads Excel to load data and uses it as input features, and uses MinMaxScaler to normalize the data, and saves the normalized model for subsequent use. The neural network model includes a four-layer structure: the first fully connected layer, the input dimension is 3, and the output dimension is 128; the second fully connected layer, the input dimension is 128, and the output dimension is 64; the third fully connected layer, the input dimension is 64, and the output dimension is 32; the fourth fully connected layer, the input dimension is 32, and the output dimension is 1. At the same time, each layer uses ReLU as the activation function. In the model training step, the model is trained using the mean square error loss function and the Adam optimizer, and the training cycle is set to 1000 times. In each training cycle, the model parameters are updated by the back propagation algorithm to minimize the mean square error loss function. In transfer learning, the neural network source domain model needs to maintain the same structure as the base model, so the same method is used to train the neural network source domain model.

[0027] Furthermore, the optimization of the neural network in step S7 specifically includes: loading the weights of the source domain model, initializing the weights of the first two layers of the base model to the weights of the source domain model; the specific details are: assigning the weight data of the first fully connected layer of the source domain model to the first fully connected layer of the base model; assigning the weight data of the second fully connected layer of the source domain model to the second fully connected layer of the base model; reinitializing the weights of the last two layers of the base model using the Xavier initialization method; reinitializing the weights of the third and fourth fully connected layers of the base model, and setting the bias to zero.

[0028] Furthermore, constructing the component space in step S8 specifically includes: selecting a corresponding number of elements from 12 elements according to a 5-element combination, a 6-element combination, a 7-element combination, and an 8-element combination for the RE position in the RETaO4 system, and performing arrangement and combination, the doping range of each element is 0-1, and the component space is designed according to a gradient of 0.1. The generated chemical formula combinations are used to construct the component space to be screened, and the optimized neural network based model is used to predict it, and the components are arranged from low to high according to the predicted values.

[0029] Furthermore, in step S9, the materials screened out are: Sm 0.20 Eu 0.10 Ho 0.10 Y 0.20 Yb 0.40 TaO4、Sm 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 TaO4, the thermal conductivity of the obtained samples was tested and compared with the predicted value.

[0030] On the other hand, the present invention proposes the application of the above method in the rapid prediction and screening of thermal conductivity of rare earth tantalates.

[0031] Beneficial effects of the present invention:

[0032] Through transfer learning technology, a high-precision deep neural network model is constructed to achieve rapid prediction of the thermal conductivity of rare earth tantalate materials in the composition space. Traditional methods such as trial and error and methods based on the Slack equation are inefficient and simple and crude, while first-principles calculations are expensive. In contrast, the present invention introduces machine learning technology to provide clear directional guidance for experiments by constructing a thermal conductivity prediction model, effectively reducing the time and cost of the experiment. In response to the problem of insufficient data in traditional thermal conductivity models, the present invention uses transfer learning to optimize and improve them, with better prediction performance.

[0033] Transfer learning technology enables the model to maintain high predictive performance even when the amount of data is insufficient. By pre-training the model in related fields or on tasks with sufficient data, and then fine-tuning these models for the prediction of thermal conductivity of rare earth tantalates, the accuracy and reliability of the prediction are significantly improved. In step S6, the neural network base model is trained based on the rare earth tantalate RETaO4 data set, and the source domain model is trained based on a large data set of multi-material systems. These data sets include data on related materials such as rare earth niobates and rare earth zirconates, which expands the scope of the data set and enhances the generalization ability of the model. In step S7, by optimizing the parameters of the neural network base model and combining the parameters of the source domain model, the model can better adapt to the characteristics of rare earth tantalate materials and improve the accuracy of the prediction.

[0034] In materials science, the problem of sparse small samples is a common challenge. The present invention solves the problem of data sparsity by introducing transfer learning technology, using data from other material systems for pre-training, and then applying it to the prediction of thermal conductivity of rare earth tantalate materials. In step S2, thermal conductivity data of related materials such as rare earth niobates and rare earth zirconates are collected. The chemical formulas of these data include but are not limited to A2B2O7, RETaO7, and RETaO9 types, which expands the coverage of the data set and increases the diversity and richness of the data. In step S3, the elemental properties of rare earth tantalates and other materials are weighted mathematically calculated by weighted average feature calculation to generate more comprehensive initial features. These elemental properties include atomic number, atomic mass, period, family, density, melting point, bulk modulus, Pauling electronegativity, etc., which provide richer input information for the model. In step S4, redundant features are eliminated through feature screening to ensure the independence and effectiveness of the input features and improve the predictive ability of the model.

[0035] Through machine learning technology, the model can provide clear directional guidance for the experiment, reduce the blindness and trial and error process of the experiment, and improve the efficiency and success rate of the experiment. In step S8, the component space is constructed, and the optimized neural network based model is used for prediction. The predicted values ​​are arranged from low to high to quickly screen out the expected low thermal conductivity materials. In this step, the RE position in the RETaO4 system is constructed according to the dose ratio of each element 0-1, with a step length of 0.1, to construct the component space to be screened, ensuring the comprehensiveness and systematicness of the screening process. In step S9, the screened materials are experimentally prepared and their performance is verified to ensure the accuracy and practicality of the model prediction results. Specific experimental verification includes the screening of screened materials such as Sm 0.20 Eu 0.1 0Ho 0.10 Y 0.20 Yb 0.40 TaO4、Sm 0.10 Eu 0.10 Y0.20 Yb 0.40 Ce 0.10 Lu 0.10 The thermal conductivity of TaO4 was tested to verify the reliability and accuracy of the model prediction.

[0036] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0038] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0039] Figure 2 is the feature recursive elimination error map;

[0040] Figure 3 This is a schematic diagram of the performance of the deep neural network model after transfer learning;

[0041] Figure 4 Sm 0.20 Eu 0.10 Ho 0.10 Y 0.20 Yb 0.40 Schematic diagram comparing the actual value and predicted value of TaO4 material;

[0042] Figure 5 To calculate Sm using the Slack model 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 Comparison chart between the thermal conductivity of TaO4, the thermal conductivity predicted by this transfer learning model, and the actual thermal conductivity of the material. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Example 1

[0045] like Figure 1 As shown, the rare earth tantalate material screening method based on transfer learning described in this embodiment includes the following steps:

[0046] S1: Collect the base model data set: the experimental data of thermal conductivity of each rare earth tantalate material at the corresponding temperature. The chemical formula of the collected data is RETaO4.

[0047] S2: Collect source domain model data set: experimental data of thermal conductivity of other rare earth thermal barrier coating materials at corresponding temperatures, including rare earth niobates, rare earth zirconates, etc. The data are all from published papers.

[0048] S3: Taking the element dose ratio as the weight, perform weighted mathematical operations on the features corresponding to the RE, Ta, and O elements to obtain the weighted average feature, and use the weighted average feature as the initial feature. The specific features are as follows:

[0049]

[0050]

[0051] S4: By calculating the Pearson correlation coefficient between the initial features, the redundant initial features with correlation > 0.85 are eliminated to construct the feature pool. The Pearson correlation coefficient calculation formula is as follows:

[0052]

[0053] Among them: x and y are the average values ​​of the two features in the sample, x i and i Indicates the actual value of the two features. For these two features, the larger the absolute value of the correlation coefficient, the stronger the correlation between the two features. Usually, only one of the features needs to be retained. According to the correlation threshold of 0.85, the features are divided into 9 groups, and one feature is retained in each group. After removing redundant features, the feature grouping and retention are as follows:

[0054]

[0055] S5: The optimal sub-features are selected through recursive feature elimination. The specific method is as follows: first, the model is trained using all the initial features; second, the impact of each feature on the model's predictive ability is evaluated, which usually depends on the feature importance score provided by the model; then, based on these scores, the least important features are identified and removed; finally, the above process is repeated until the predetermined number of features is reached or the model performance is no longer significantly improved. Finally, through recursive feature elimination, the remaining features are ionic bond strength, temperature, and Clementi effective nuclear charge number.

[0056] S6: According to the optimal sub-features, a neural network base model based on rare earth tantalate RETaO4 and a neural network source domain model based on a large data set including rare earth niobate, rare earth zirconate, etc. are trained. The specific method is: the data set is randomly divided into a training set and a test set at a ratio of 8:2, and a four-layer neural network is built. The first layer has an input dimension of 3 and an output dimension of 128, the second layer has an input dimension of 128 and an output dimension of 64, the third layer has an input dimension of 64 and an output dimension of 32, and the fourth layer has an input dimension of 32 and an output dimension of 1. ReLU is used as the activation function to train two models.

[0057] S7: Use the trained neural network source domain model to optimize the parameters of the neural network base model. The specific method is: select the source domain model, transfer the weights, import the weights of the first two layers of the pre-trained model into the model, and freeze the weights of these layers to prevent them from being updated during subsequent training. Initialize the subsequent layers and use the random initialization method for the weights of the last two layers of the neural network to ensure that these layers can learn the specific features of the new task from scratch, and retrain the neural network base model.

[0058] S8: For the 12 elements described in step S1, a composition space is constructed, and materials that meet the requirements of low thermal conductivity are quickly screened out in the composition space. The specific method is: select the corresponding number of elements from the 12 elements according to the 5-element combination, 6-element combination, 7-element combination, and 8-element combination, and arrange and combine them. The doping range of each element is 0-1, and the composition space is designed according to the gradient of 0.1, generating nearly 300,000 chemical formula combinations.

[0059] S9: Experimentally prepare the low thermal conductivity material selected in step S8 and verify its performance. The specific preparation method is: weigh the raw material powder of corresponding grams, put it into a planetary ball mill and mix it for 24 hours, dry it in a drying oven, put it into a 300x sieve to sieve out the fine powder, and press the dried powder into a disc with a diameter of 15mm and a thickness of 1mm, and then place it in a high-temperature sintering furnace for sintering. After keeping it at 1650℃ for 8 hours, the required dense block tantalate ceramic is obtained. Finally, the thermal conductivity of the obtained sample is tested, and the material Sm 0.20 Eu 0.10 Ho 0.10 Y 0.20 Yb 0.40 TaO4、Sm 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 The thermal conductivity of TaO4 at 900℃ is 1.45W / (m·k) and 1.47W / (m·k), which is close to the predicted results of 1.50W / (m·k) and 1.49W / (m·k).

[0060] At the same time, this embodiment uses the first principle calculation program VASP based on density functional theory (DFT) to calculate the parameters A, M-, θD, δ, γ, T, n required by the Slack model and bring them into the Slack thermal conductivity model:

[0061]

[0062] Calculate Sm 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 Thermal conductivity of TaO4 material. The experimental results, transfer learning model prediction results, and Slack model calculation results are shown in the following table:

[0063]

[0064]

[0065] Among them, the root mean square error between the predicted value and the experimental value of the transfer learning model described in this patent is: 0.054, and the root mean square error between the calculated value and the experimental value of the traditional calculation method Slack model is: 0.535. The accuracy of this method is better than that of the traditional thermal conductivity calculation method.

[0066] In summary, the present invention uses transfer learning technology to establish a high-precision thermal conductivity prediction model for rare earth tantalate materials, realizes the rapid prediction of the thermal conductivity of multi-component RETaO4, and solves the problem of small data sets faced by machine learning. Compared with the traditional experimental trial and error method or the first-principles calculation method to obtain thermal conductivity, the thermal conductivity prediction method based on transfer learning adopted by the present invention is more convenient, fast and cost-effective.

[0067] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A rare earth tantalate material screening method based on transfer learning, It is characterized in that The following steps are involved: S1: Collect the thermal conductivity experimental data of each rare earth tantalate material at 20-900℃. The chemical formula of the collected data is RETaO 4 , wherein RE is a combination of one or more rare earth elements Y, Ce, Nd, Sm, Eu, Gd, Dy, Ho, Er, Tm, Yb, Lu, Ta represents tantalum, O represents oxygen, and the element dosage ratio of RE:Ta:O is 1:1:4; S2: Collect the experimental data of thermal conductivity of rare earth niobate and rare earth zirconate materials at different temperatures. The chemical formula of the collected data includes but is not limited to A 2 B 2 O 7 、RETaO 7 ,RETa 3 O 9 type; S3: Taking the element dose ratio as weight, the RETaO 4 and A 2 B 2 O 7 、RETaO 7 ,RETa 3 O 9 Perform various weighted mathematical operations on the element properties corresponding to the chemical formula to obtain weighted features, which are used as initial features; S4: Targeting RETaO 4 The dataset is used for feature screening. By calculating the Pearson correlation coefficient between the initial features, 0.85 is set as the strong correlation threshold, the features are divided into 9 groups, only one feature is retained in each group, redundant features are removed, and a feature pool is constructed. The calculation formula for the correlation coefficient is: Among them: x and y are the average values ​​of the two features in the sample, x i and i Represents the actual values ​​of the two features in the i-th sample; S5: The optimal sub-features are selected through the feature recursive elimination method; the order of feature deletion from the first to the last is:

1. first ionization energy, 2. bulk modulus, 3. melting point, 4. ion radius ratio, 5. mass disorder parameter, 6. radius disorder parameter, and the final retained features are: ionic bond strength, temperature, and Clementi effective nuclear charge number; S6: Build a neural network model with a 4-layer structure, where the input dimension is 3 and the output dimension is 1. Use ReLU as the activation function, Adam as the optimizer, and the learning rate is set to 0.

01. Train the rare earth tantalate RETaO based on the optimal sub-features. 4 The base model of the data set and the source domain model of the large data set composed of multiple material systems; S7: The weights of the first two layers of the trained neural network source domain model are assigned to the first two layers of the base model, and the weights of the last two layers of the base model are initialized using Xavier to achieve the purpose of optimizing the neural network base model; S8: For the 12 elements in S1, select the corresponding number of elements from the 12 elements according to the 5-element combination, 6-element combination, 7-element combination, and 8-element combination, and arrange and combine them, design the component space, calculate the features and generate Excel, use the optimized neural network base model to read Excel, predict the thermal conductivity of all components, and arrange the components in ascending order from low to high according to the predicted thermal conductivity values; S9: The low thermal conductivity component selected by S8 is Sm 0.20 Eu 0.10 Ho 0.10 Y 0.20 Yb 0.40 TaO 4 、Sm 0.10 Eu 0.10 Y 0.20 Yb 0.40 Ce 0.10 Lu 0.10 TaO 4 , experimentally prepare the materials and verify the performance.

2. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The rare earth tantalate data set RE in step S1 includes at least one piece of rare earth tantalate thermal conductivity data of a combination of two to ten elements; In step S2, A is a combination of one or more rare earth elements Y, Ce, Nd, Sm, Eu, Gd, Dy, Ho, Er, Tm, Yb, and Lu, B represents a combination of one or more rare earth elements Ta, Nb, and Zr, O represents an oxygen element, and the atomic number of the oxygen element is 4, 7, or 9, so the element dosage ratio of A:B:O is 1:1:4, 1:1:7, 2:2:7, or 1:3:

9.

3. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The element properties in step S3 include but are not limited to: atomic number, atomic mass, period, group, density, melting point, bulk modulus, mass disorder parameter, Pauling electronegativity, AR electronegativity, MB electronegativity, first ionization energy, electron affinity, Clementi effective nuclear charge number, Slater effective nuclear charge number, charge-to-radius ratio, ionic bond strength, ratio of RE position to Ta ion radius, radius disorder parameter and covalent radius.

4. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The correlation calculation method in step S4 is as follows: Among them: x and y are the average values ​​of the two features in the sample, x i and i Represents the actual value of the two features; for these two features, the larger the absolute value of the correlation coefficient, the stronger the correlation between the two features, and usually only one of the features needs to be retained; according to the correlation threshold of 0.85, the features are divided into 9 groups, and one feature is retained in each group.

5. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The method for recursively eliminating features in step S5 includes: first, using all initial features to train the model; second, evaluating the impact of each feature on the model's predictive ability, which step usually relies on the feature importance score provided by the model; then, based on these scores, identifying and removing the least important features each time; finally, repeating the above process until a predetermined number of features is reached or the model performance is no longer significantly improved; the order of feature deletion from the first to the last time is:

1. first ionization energy, 2. bulk modulus, 3. melting point, 4. ionic radius ratio, 5. mass disorder parameter, 6. radius disorder parameter, and the features finally retained are: ionic bond strength, temperature, and Clementi effective nuclear charge number.

6. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: In step S6, the neural network base model reads Excel to load data and uses it as input features, and uses MinMaxScaler to normalize the data, and saves the normalized model for subsequent use. The neural network model includes a four-layer structure: the first fully connected layer, the input dimension is 3, and the output dimension is 128; the second fully connected layer, the input dimension is 128, and the output dimension is 64; the third fully connected layer, the input dimension is 64, and the output dimension is 32; the fourth fully connected layer, the input dimension is 32, and the output dimension is 1, and each layer uses ReLU as the activation function; In the model training step, the mean square error loss function and Adam optimizer are used to train the model. The training cycle is set to 1000 times. In each training cycle, the model parameters are updated through the back propagation algorithm to minimize the mean square error loss function. The same approach is used in transfer learning to train the neural network source domain model.

7. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The transfer learning method in step S7 is: loading the weights of the source domain model, and initializing the first two layers of the base model to the weights of the source domain model; the specific details are: assigning the weight data of the first fully connected layer of the source domain model to the first fully connected layer of the base model; assigning the weight data of the second fully connected layer of the source domain model to the second fully connected layer of the base model; reinitializing the weights of the last two layers of the base model using the Xavier initialization method; reinitializing the weights of the third and fourth fully connected layers of the base model, and setting the bias to zero.

8. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: The specific method of constructing the component space in step S8 is: 4 The RE positions in the system select the corresponding number of elements from 12 elements according to the 5-element combination, 6-element combination, 7-element combination, and 8-element combination, and arrange them in combination. The doping range of each element is 0-1, and the component space is designed according to the gradient of 0.

1. The generated chemical formula combinations construct the component space to be screened, and the thermal conductivity is predicted using the optimized neural network based model, and the components are arranged from low to high according to the predicted values.

9. The rare earth tantalate material screening method based on transfer learning as claimed in claim 1, Features: In the step S9, the low thermal conductivity material screened out in step S8 is experimentally prepared and the performance is verified; the specific preparation method is: weigh the raw material powder of corresponding grams, put it into a planetary ball mill for mixing for 24 hours, dry it in a drying oven, put it into a 300x sieve to sieve out fine powder, and press the dried powder into a disc with a diameter of 15mm and a thickness of 1mm, and then place it in a high-temperature sintering furnace for sintering. After being kept at 1650°C for 8 hours, the required dense block tantalate ceramic is obtained; finally, the thermal conductivity of the obtained sample is tested.

10. Application of the method according to any one of claims 1 to 9 in the rapid prediction and screening of thermal conductivity of rare earth tantalates.