Material heat storage performance prediction method based on multi-source domain knowledge distillation

Abstract

This invention relates to a method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation. First, the crystal structure of a first sample is obtained, and its thermal storage performance parameters are calculated using first-principles calculations to construct a target dataset. Second, the crystal structure and various readily available physicochemical properties of a second sample are searched, and multiple source domain pre-training datasets are constructed by grouping according to different parameters. Next, isomorphic second models are trained using each source domain dataset, with each model predicting one physicochemical property parameter, and its second intermediate layer feature representation is extracted. Then, a first model with the same structure as the first sample is trained to predict the thermal storage performance parameters, and knowledge distillation is introduced during training: the intermediate layer features of each second model are fused using learnable weighted methods to generate reference features. The difference between these reference features and the first intermediate layer features of the first model is calculated, and the model is optimized by combining the error of the predicted values. Finally, the crystal structure of the sample to be predicted is input into the trained first model to obtain the predicted thermal storage performance parameters. This method is suitable for small-sample scenarios with limited thermal storage performance label data and can provide rational guidance for the design of thermal storage materials.

Description

Technical Field

[0001] This invention relates to the field of data-driven modeling and performance prediction technology for thermochemical thermal storage materials, and in particular to a method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation. Background Technology

[0002] Energy storage technology can enhance the absorption capacity of renewable energy, facilitate frequency and peak regulation of traditional coal-fired power plants, and optimize the energy structure, making it an important technology during the energy transition period. Thermochemical thermal energy storage systems absorb high-temperature heat energy through chemical reactions of a medium at high temperatures, storing it as chemical energy and releasing it when necessary. They offer advantages such as extremely long storage times, non-toxic and harmless materials, and ease of long-term transportation. Metal oxides have simple reaction systems and high thermal density, showing great application potential. However, metal oxide materials suffer from various reactivity issues, significantly limiting their commercial application.

[0003] Currently, research on the thermochemical thermal storage performance of metal oxides focuses on material modification. By changing the preparation method of the material, such as doping or loading, the morphology and structure of the material are altered, thereby improving important properties such as reaction kinetics and cycle stability. Traditional methods are mostly experimental verification.

[0004] However, experimental verification requires the preparation of samples in advance, which involves a large amount of engineering work. Therefore, those skilled in the art have proposed a method to calculate the thermal storage performance of a material based on its crystal structure using first-principles calculations. In this way, the crystal structure of the material can be obtained through computer simulation, eliminating the need to prepare physical samples and improving efficiency.

[0005] However, this kind of first-principles calculation requires the calculator to have a great deal of professional knowledge. On the one hand, it places high demands on the calculator, and on the other hand, the calculation is also very complex, requiring experienced professionals to use high-performance computing equipment to complete it over several days. This still slows down the progress of new material research and development.

[0006] To overcome the problems of excessive computational load and long processing time in first-principles calculations, some practitioners have proposed using model prediction methods. This involves first calculating the thermal storage performance of some materials using first-principles calculations, then constructing training samples, and finally using machine learning to predict the thermal storage performance of other materials. For example, the reference (《Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties》. Physical Review Letters, 2018, 120(14).) innovatively proposes a machine learning model for fast end-to-end prediction of crystalline material properties, called Crystal Graph Neural Network (CGCNN). This network requires only a crystal information file as input, transforms the crystal information into a graph through a graph convolutional neural network architecture, and trains the model to predict crystal properties. It has been used for the physicochemical properties of some materials and is currently used for predicting general crystal physical properties such as formation energy, band gap, and elastic modulus. However, this type of method relies on large amounts of target property label data, and suffers from insufficient training data and poor model generalization ability in small-sample scenarios for thermal storage performance prediction.

[0007] To improve the performance of machine learning on small datasets, existing research has proposed transfer learning or knowledge distillation methods to transfer knowledge learned from a source task to a target task. However, general transfer learning and knowledge distillation methods require a strong correlation between the source and target tasks, such as detecting cats and detecting tigers. Otherwise, the target task still needs a sufficiently large sample size before transferring the knowledge to the target task. In the field of crystal structure data, it is difficult to find a single source task that can form a strong correlation with thermal storage performance, making it difficult to apply traditional transfer learning and knowledge distillation methods to predict thermal storage performance.

[0008] In summary, those skilled in the art urgently desire to develop a method that can improve the accuracy of predicting the thermal storage performance of materials using a small number of limited samples without strongly correlated source tasks, thereby solving the problem that obtaining samples through first-principles calculations is too time-consuming. Summary of the Invention

[0009] The purpose of this invention is to provide a method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation. This method uses a CIF crystal structure file as a unified crystal structure input. It trains isomorphic second models using the physicochemical properties of multiple source domains related to the crystal structure and readily available. During the training of the first model used to predict thermal storage performance parameters, second intermediate layer feature representations are extracted from the same intermediate layer of each second model. These representations are then fused using learnable weighted methods, and distance constraints are established between these fusion methods and the corresponding intermediate layer feature representation of the first model. Therefore, even without strongly correlated source tasks, the prediction accuracy of the thermal storage performance of materials can be improved using a small number of samples, thus solving the problem of excessively time-consuming sample calculations using first-principles calculations.

[0010] The objective of this invention can be achieved through the following technical solutions: A method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation includes: Step S1: Obtain the crystal structure of the first sample, calculate the thermal storage performance parameters based on first principles, and construct the crystal structure-thermal storage performance dataset by combining the crystal structure and thermal storage performance parameters into the first sample. Step S2: Search to obtain the crystal structure and its corresponding physicochemical properties of the second sample, take the crystal structure and physicochemical properties as the second sample, and group the second sample according to the physicochemical properties to construct multiple crystal structure-source domain property pre-training datasets for different physicochemical properties. Step S3: Train the corresponding second model using the second samples in the pre-training dataset of each crystal structure-source domain property. Each second model corresponds one-to-one with each physicochemical property parameter. The input is the crystal structure, the output is the corresponding physicochemical property parameter, and the intermediate layer output is the feature representation of the second intermediate layer. Step S4: Use the first sample to train a first model with a structure that is completely identical to the second model, wherein the input of the first model is the crystal structure, the output is the thermal storage performance parameters, and the intermediate layer output is the first intermediate layer feature representation; Step S5: Input the crystal structure of the sample to be predicted into the trained first model, and obtain the thermal storage performance parameters output by the first model as the prediction result.

[0011] The thermal storage performance parameters include oxygen vacancy formation energy, Debye temperature, and specific heat capacity. There are multiple second models, each corresponding to a specific thermal storage performance parameter.

[0012] The carrier of the crystal structure is a CIF crystal structure file; Both the first and second models are graph neural network models, which obtain the crystal's structural information based on the parsing of CIF crystal structure files. The process of obtaining the crystal's structural information includes: Parse CIF crystal structure files to extract crystal structure data. By treating atoms in a crystal as nodes in a graph and chemical bonds as edges, a crystal graph can be constructed. For each node in the crystal diagram, update the information of each node with its own information, the information of adjacent nodes connected by chemical bonds, and the edge information between adjacent nodes; After multiple updates, the information of each node is averaged and pooled, and the information of all nodes is then spliced ​​together to obtain the crystal's structural information.

[0013] The physicochemical properties include molar mass, molar volume, density, average atomic molar mass, average atomic molar volume, average atomic van der Waals radius, formation energy, band gap, average electronegativity of oxygen atom's nearest neighbor atoms, elastic modulus, average metal-oxygen bond length, average metal-oxygen bond angle, average electron affinity, average atomic first ionization energy, O2p band center, and d band center.

[0014] The second sample in step S2 is obtained by searching an open-source database.

[0015] The training process in step S4 includes: Step S4-1: Input the crystal structure in the first sample into each trained second model to obtain the second intermediate layer feature representation of each second model; Step S4-2: Input the crystal structure in the first sample into the first model, which is completely consistent with the structure of the second model, to obtain the first intermediate layer feature representation and output thermal storage performance parameters of the first model; Step S4-3: Based on the feature representations of each second intermediate layer and the learnable second weight sequence, the reference intermediate feature representation is calculated by weighted summation. Step S4-4: Based on the difference between the reference intermediate feature representation and the first intermediate layer feature representation, and the difference between the thermal storage performance parameters output by the first model and the thermal storage performance parameters in the first sample, the comprehensive loss is calculated. Step S4-5: Modify the model parameters of the second model and the learnable second weight sequence based on the comprehensive loss.

[0016] The overall loss is: in: To account for the overall losses, The latent feature loss represents the difference between the reference intermediate feature representation and the first intermediate layer feature representation. For the average error loss, To reduce the weight of the fusion process, Features of the first intermediate layer For reference, intermediate feature representation, For distance measurement function, N The number of samples in the first sample. The first model is based on the thermal storage performance parameters output for the i-th first sample. y i Let be the thermal storage performance parameters in the i-th first sample.

[0017] The distance metric function is Euclidean distance or cosine distance.

[0018] A material thermal storage performance prediction device based on multi-source domain knowledge distillation includes a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the method described above.

[0019] A storage medium having a program stored thereon, which, when executed, implements the method described above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. Using CIF crystal structure files as a unified crystal structure input, isomorphic second models are trained using multiple source domain physicochemical properties that are related to the crystal structure and are relatively easy to obtain. During the training of the first model used to predict thermal storage performance parameters, the feature representations of the second intermediate layer are extracted from the same intermediate layer of each second model. After learning-friendly weighted fusion, distance constraints are established between these representations and the corresponding intermediate layer feature representation of the first model. This allows for improved prediction accuracy of the thermal storage performance of materials using a small number of samples, even without strongly correlated source tasks, thus solving the problem of excessively time-consuming sample calculations using first-principles calculations.

[0021] 2. The input carrier for crystal structures is standardized as CIF files, a standard and universal crystal structure data format in materials science. This standardization of input improves the method's versatility and reproducibility. Furthermore, both the first and second models are specifically defined as graph neural network models. The process of parsing CIF files into crystal diagrams and performing information propagation and aggregation is described, fully leveraging the inherent advantages of graph neural networks in characterizing graph data such as crystal structures, effectively capturing the chemical bonds and interactions between atoms.

[0022] 3. It significantly reduces the cost and barrier to building large-scale source domain pre-training datasets, and can easily utilize the massive crystal structure and computational property data in public materials databases such as Materials Project and OQMD.

[0023] 4. A learnable and adaptive knowledge transfer mechanism is implemented. By weighted fusion of the second intermediate layer features of each pre-trained second model in the same intermediate layer, a reference intermediate feature is generated. This reference intermediate feature is then used to supervise the first intermediate layer features of the corresponding layer of the first model. This allows the model to automatically learn during training how to select and combine the most useful knowledge for the current thermal storage performance prediction task from multiple source domain tasks, rather than simply averaging or randomly selecting, thereby improving the efficiency and effectiveness of knowledge distillation.

[0024] 5. On the one hand, ensure that the model's predictions of thermal storage performance parameters are as accurate as possible. On the other hand, constrain the intermediate feature representations of the model to approach the reference features that integrate multi-source knowledge, so as to effectively utilize multi-source data. Furthermore, the importance of multiple source domain tasks obtained by automatic learning can rationally guide the performance regulation of thermal storage materials, providing direction for the development of high-performance thermal storage materials. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the technical route of the present invention; Figure 2 A comparison chart of the predicted and actual Debye temperatures for the material; Figure 3 A comparison chart showing the predicted and actual values ​​of oxygen vacancy formation energy in materials; Figure 4 A comparison chart showing the predicted and actual values ​​of the specific heat capacity (logarithm to base 10) of the material at 1000K; Figure 5 This is a schematic diagram of the main steps of the method of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] A method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation, such as Figure 1 and Figure 5 As shown, it includes: Step S1: Obtain the crystal structure of the first sample, calculate the thermal storage performance parameters based on first principles, and construct the crystal structure-thermal storage performance dataset by combining the crystal structure and thermal storage performance parameters into the first sample. In this embodiment, the carrier of the crystal structure is a CIF crystal structure file, and accurate thermal storage performance parameters can be obtained through first-principles calculations.

[0028] The thermal storage performance parameters include oxygen vacancy formation energy, Debye temperature, and specific heat capacity. There are multiple second models, each corresponding to one of the thermal storage performance parameters.

[0029] Step S2: Search to obtain the crystal structure and its corresponding physicochemical properties of the second sample, take the crystal structure and physicochemical properties as the second sample, and group the second sample according to the physicochemical properties to construct multiple crystal structure-source domain property pre-training datasets for different physicochemical properties. Both the first and second models are graph neural network models, which obtain the crystal's structural information based on the parsing of CIF crystal structure files. The process of obtaining the crystal's structural information includes: Parse CIF crystal structure files to extract crystal structure data. By treating atoms in a crystal as nodes in a graph and chemical bonds as edges, a crystal graph can be constructed. For each node in the crystal diagram, update the information of each node with its own information, the information of adjacent nodes connected by chemical bonds, and the edge information between adjacent nodes; After multiple updates, the information of each node is averaged and pooled, and the information of all nodes is then spliced ​​together to obtain the crystal's structural information.

[0030] In addition, in this embodiment, the physicochemical properties include molar mass, molar volume, density, average atomic molar mass, average atomic molar volume, average atomic van der Waals radius, formation energy, band gap, average electronegativity of oxygen atom's nearest neighbor atoms, elastic modulus, average metal-oxygen bond length, average metal-oxygen bond angle, average electron affinity, average atomic first ionization energy, O2p band center and d band center. In other embodiments, it may also be a part of all the above physicochemical properties.

[0031] In this embodiment, the second sample is obtained based on an open-source database search. Specifically, by calling an API, the general rules for crystal data of materials that meet these conditions are downloaded. Taking the Materials Project as an example, the general rules for crystal data include the crystal's structural information and some calculated properties, including but not limited to formation energy, band gap energy, and elastic modulus. This data is then processed in batches to ultimately obtain the crystallographic information file of the crystal, as well as some crystal properties available in the database.

[0032] After acquiring the crystallographic information files for all crystal structures, the downloaded files were batch-analyzed using Pymatgen, an open-source materials science library based on Python, to obtain low-cost properties that can be calculated solely from the crystal structure. These properties include, but are not limited to: average atomic mass, average molar volume, average metal-oxygen bond length, average elemental electronegativity, average elemental electron affinity, and average elemental ionization energy. Subsequently, these property data were combined with the crystallographic information files to form a second sample dataset of "crystal structure-source domain properties." This dataset was used to train a second model, which is a source domain model.

[0033] Step S3: Train the corresponding second model using the second samples in the pre-training dataset of each crystal structure-source domain property. Each second model corresponds one-to-one with each physicochemical property parameter. The input is the crystal structure, the output is the corresponding physicochemical property parameter, and the intermediate layer output is the feature representation of the second intermediate layer. Specifically, the features of the second intermediate layer serve as latent features of the source domain model and can be utilized by the first model.

[0034] The first model, based on the crystal's structural information, can obtain the feature representation of the second intermediate layer and the predicted values ​​of its physicochemical properties. Specifically, in the training of the second model, the loss function is the mean absolute error.

[0035] Step S4: Use the first sample to train a first model with a structure that is completely consistent with the second model. The input of the first model is the crystal structure, the output is the thermal storage performance parameters, and the output of the intermediate layer is the feature representation of the first intermediate layer. The training process in step S4 includes: Step S4-1: Input the crystal structure from the first sample into each of the trained second models. In the process, the second intermediate layer feature representations of each second model are obtained, and the engineering can also obtain the latent feature set. ,in, For the first second model The output of the second intermediate layer feature representation is also the first second model. Potential characteristics, For the second second model The output of the second intermediate layer feature representation is also the second second model. Potential characteristics, For the first N A second model The output of the second intermediate layer feature representation, which is also the first N A second model Potential characteristics; Step S4-2: Input the crystal structure from the first sample into the first model whose structure is completely identical to that of the second model. In this process, the first intermediate layer feature representation of the first model is obtained. Z and output thermal storage performance parameters; Step S4-3: Based on the feature representations of each second intermediate layer and the learnable second weight sequence The reference intermediate feature representation is calculated by weighted summation. : in: As a reference intermediate feature representation, it represents the comprehensive potential features of the source domain. M For the number of second models, For the first j A second model The output of the second intermediate layer feature representation, which is also the first j A second model Potential characteristics, For the first j A second model The weight.

[0036] Step S4-4: Based on the difference between the reference intermediate feature representation and the first intermediate layer feature representation, and the difference between the thermal storage performance parameters output by the first model and the thermal storage performance parameters in the first sample, the comprehensive loss is calculated. The overall loss is as follows: in: To account for the overall losses, The latent feature loss represents the difference between the reference intermediate feature representation and the first intermediate layer feature representation. For the average error loss, To reduce the weight of the fusion process, Features of the first intermediate layer For reference, intermediate feature representation, The distance metric function can be either Euclidean distance or cosine distance. N The number of samples in the first sample. The first model is based on the thermal storage performance parameters output for the i-th first sample. y i Let be the thermal storage performance parameters in the i-th first sample.

[0037] Step S4-5: Modify the model parameters of the second model and the learnable second weight sequence based on the comprehensive loss.

[0038] Step S5: Input the crystal structure of the sample to be predicted into the trained first model, and obtain the thermal storage performance parameters output by the first model as the prediction result.

[0039] Using Debye temperature, oxygen vacancy formation energy, and specific heat capacity at 1000 K (using the common logarithm) as target thermal storage performance, the model was trained. Ultimately, the first model's performance in predicting Debye temperature was as follows: the mean absolute error (MAE) for the training set was 25.162, and for the test set it was 45.896; the regression coefficient R² for the predicted / true values ​​on the training set was 0.926, and the regression coefficient R² for the predicted / true values ​​on the test set was... 2 The value is 0.807; the comparison between the predicted Debye temperature and the actual value is shown in the figure below. Figure 2 As shown. The first model's performance in predicting oxygen vacancy formation energy is as follows: the mean absolute error (MAE) for the training set is 0.141, and the MAE for the test set is 0.412; the regression coefficient R between the predicted and actual values ​​in the training set is... 2 The regression coefficient Ri of the model's predicted values / true values ​​in the test set is 0.975. 2 The value is 0.845; the comparison between the oxygen vacancy formation energy predicted by the model and the actual value is shown in the figure below. Figure 3 As shown. The first model's performance in predicting the specific heat capacity (using the common logarithm) at 1000K is as follows: the mean absolute error (MAE) for the training set is 0.010, and the mean absolute error (MAE) for the test set is 0.075; the regression coefficient R between the predicted and actual values ​​in the training set is... 2 The regression coefficient Ri of the model's predicted values / true values ​​on the test set is 0.984. 2 The value is 0.858; the comparison between the specific heat capacity (using the common logarithm) predicted by the model at 1000K and the actual value is shown in the figure below. Figure 4 As shown.

[0040] In this embodiment of the invention, a standardized crystal structure-thermal storage performance database is obtained through first-principles calculations; a large number of metal oxide crystal structures and easily accessible physicochemical properties are obtained from open-source material databases to construct a crystal structure-source domain property dataset; a graph neural network machine learning model is trained on the crystal structure-source domain property dataset to obtain source domain models; multiple source domain models are used to perform multi-source crystal physical property knowledge distillation to train a high-performance, highly generalizable thermal storage performance prediction model; structural files are obtained from the open-source material database and input into the model to output predictions of thermal storage properties, which significantly reduces the range of candidate materials and avoids the high manpower and material consumption of experimental trial and error methods.

[0041] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation, characterized in that, include: Step S1: Obtain the crystal structure of the first sample, calculate the thermal storage performance parameters based on first principles, and construct the crystal structure-thermal storage performance dataset by combining the crystal structure and thermal storage performance parameters into the first sample. Step S2: Search to obtain the crystal structure and its corresponding physicochemical properties of the second sample, take the crystal structure and physicochemical properties as the second sample, and group the second sample according to the physicochemical properties to construct multiple crystal structure-source domain property pre-training datasets for different physicochemical properties. Step S3: Train the corresponding second model using the second samples in the pre-training dataset of each crystal structure-source domain property. Each second model corresponds one-to-one with each physicochemical property parameter. The input is the crystal structure, the output is the corresponding physicochemical property parameter, and the intermediate layer output is the feature representation of the second intermediate layer. Step S4: Use the first sample to train a first model with a structure that is completely identical to the second model, wherein the input of the first model is the crystal structure, the output is the thermal storage performance parameters, and the intermediate layer output is the first intermediate layer feature representation; Step S5: Input the crystal structure of the sample to be predicted into the trained first model, and obtain the thermal storage performance parameters output by the first model as the prediction result.

2. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 1, characterized in that, The thermal storage performance parameters include oxygen vacancy formation energy, Debye temperature, and specific heat capacity. There are multiple second models, each corresponding to a specific thermal storage performance parameter.

3. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 1, characterized in that, The carrier of the crystal structure is a CIF crystal structure file; Both the first and second models are graph neural network models, which obtain the crystal's structural information based on the parsing of CIF crystal structure files. The process of obtaining the crystal's structural information includes: Parse CIF crystal structure files to extract crystal structure data. By treating atoms in a crystal as nodes in a graph and chemical bonds as edges, a crystal graph can be constructed. For each node in the crystal diagram, update the information of each node with its own information, the information of adjacent nodes connected by chemical bonds, and the edge information between adjacent nodes; After multiple updates, the information of each node is averaged and pooled, and the information of all nodes is then spliced ​​together to obtain the crystal's structural information.

4. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 1, characterized in that, The physicochemical properties include molar mass, molar volume, density, average atomic molar mass, average atomic molar volume, average atomic van der Waals radius, formation energy, band gap, average electronegativity of oxygen atom's nearest neighbor atoms, elastic modulus, average metal-oxygen bond length, average metal-oxygen bond angle, average electron affinity, average atomic first ionization energy, O2p band center, and d band center.

5. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 1, characterized in that, The second sample in step S2 is obtained by searching an open-source database.

6. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 1, characterized in that, The training process in step S4 includes: Step S4-1: Input the crystal structure in the first sample into each trained second model to obtain the second intermediate layer feature representation of each second model; Step S4-2: Input the crystal structure in the first sample into the first model, which is completely consistent with the structure of the second model, to obtain the first intermediate layer feature representation and output thermal storage performance parameters of the first model; Step S4-3: Based on the feature representations of each second intermediate layer and the learnable second weight sequence, the reference intermediate feature representation is calculated by weighted summation. Step S4-4: Based on the difference between the reference intermediate feature representation and the first intermediate layer feature representation, and the difference between the thermal storage performance parameters output by the first model and the thermal storage performance parameters in the first sample, the comprehensive loss is calculated; Step S4-5: Modify the model parameters of the second model and the learnable second weight sequence based on the comprehensive loss.

7. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 6, characterized in that, The overall loss is: in: To account for the overall losses, The latent feature loss represents the difference between the reference intermediate feature representation and the first intermediate layer feature representation. For the average error loss, To reduce the weight of the fusion process, Features of the first intermediate layer For reference, intermediate feature representation, For distance measurement function, N The number of samples in the first sample. The first model is based on the thermal storage performance parameters output for the i-th first sample. y i Let be the thermal storage performance parameters in the i-th first sample.

8. The method for predicting the thermal storage performance of materials based on multi-source domain knowledge distillation according to claim 7, characterized in that, The distance metric function is Euclidean distance or cosine distance.

9. A material thermal storage performance prediction device based on multi-source domain knowledge distillation, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-8.

Images