A Dynamic Material Property Prediction Method and System Based on Incremental Learning and Knowledge Distillation

Abstract

This invention discloses a dynamic material performance prediction method and system based on incremental learning and knowledge distillation, belonging to the field of materials science and engineering technology. The method generates embedded representations for materials through a dynamic material encoder. For known materials, embeddings are obtained by looking up the embedding table; for newly added materials, new embeddings are generated through a neural network, and the embedding table is updated. The system employs an incremental learning mechanism, which can progressively update the model when new materials are added, ensuring the prediction accuracy of new materials while maintaining the learning results from existing data. Through a knowledge distillation mechanism, the performance predictions of similar materials are used as teacher signals. The predictions of the teacher model and the student model are weighted and fused by combining physical property similarity and structural similarity, further improving the accuracy of new material performance prediction.

Description

Technical Field

[0001] This invention relates to the fields of materials science and engineering technology, and in particular to a dynamic material performance prediction method and system based on incremental learning and knowledge distillation. Specifically, it relates to a technical solution that combines graph neural networks, Mahalanobis distance, incremental learning, dynamic material coding, and knowledge distillation. Background Technology

[0002] With the continuous emergence of new materials, especially in the fields of chemical engineering and materials science, the design and optimization process of materials is becoming increasingly complex. Traditional materials design typically relies on empirical formulas or experimental data analysis, which is not only costly but also time-consuming, often failing to meet the needs of rapid development of new materials. Therefore, how to efficiently and accurately predict the properties of new materials, especially in the early stages of materials development, has become a major challenge in the field of materials science.

[0003] In materials science, material properties are typically influenced by a variety of factors, including composition, microstructure, and processing techniques. Over the past few years, with the rapid development of big data technology and machine learning, data-driven methods have become crucial tools for materials design and optimization. In particular, deep learning algorithms have demonstrated exceptional performance in handling complex, high-dimensional data, achieving significant progress in materials property prediction, materials discovery, and formulation optimization. For example, deep neural network-based models can learn the complex relationships between different physical properties from historical materials data, thereby predicting the properties and behavior of new materials.

[0004] However, traditional machine learning methods typically suffer from the following problems:

[0005] The cold start problem for new materials: Most machine learning models rely on large amounts of training data for effective training, but for new materials, historical data is lacking, making direct and effective prediction impossible. How to quickly learn the properties of new materials and integrate them into existing models is a significant challenge for current technology.

[0006] The problem of model generalization ability: The performance of materials is affected by a variety of factors, and existing models are often trained on static data, which makes them perform poorly when facing new materials or predicting performance under different conditions. In particular, the generalization ability of existing models is often insufficient when facing unknown or very special materials.

[0007] Knowledge distillation and transfer learning: To improve the accuracy of new material predictions, many methods have attempted to incorporate the concepts of knowledge distillation and transfer learning. These methods can leverage existing material data and models to enhance the predictive power for new materials. However, how to correctly utilize knowledge distillation techniques while ensuring efficient model training remains a challenging research problem.

[0008] Data sparsity and model update issues: With the continuous introduction of new materials, material databases and knowledge graphs will become increasingly massive. This places higher demands on traditional machine learning models: the models need to be continuously and incrementally updated to adapt to the addition of new materials and ensure the accuracy and real-time performance of model predictions.

[0009] To address the above issues, incremental learning and knowledge distillation methods have attracted increasing attention from researchers in recent years. Incremental learning allows models to progressively absorb new data without retraining, which is particularly important for dynamically changing materials databases. Through incremental learning, materials property prediction models can continuously learn the properties of new materials and incorporate them into their existing knowledge framework, thus enabling the model to remain updated and gradually improve its predictive capabilities without losing existing knowledge.

[0010] Meanwhile, knowledge distillation technology can enhance the adaptability of student models to new data by transferring existing knowledge from teacher models to student models. Especially when faced with new materials, the training of student models by the teacher model can effectively avoid the "catastrophic forgetting" problem that may occur in traditional training processes, that is, the introduction of new materials will not cause the model to forget the original knowledge.

[0011] Knowledge graphs are an important tool for knowledge representation and storage in materials science. They connect information such as materials, formulations, and properties through graph structures, forming multi-dimensional materials databases. Using techniques such as Graph Neural Networks (GNNs), knowledge graphs can effectively represent the relationships between different materials as graph structures, further improving prediction accuracy and model generalization ability. Graph neural networks can capture complex nonlinear relationships between nodes in graph structures; therefore, combining graph neural networks and knowledge graphs in materials property prediction can significantly improve the ability to model the similarities and differences between materials.

[0012] However, existing knowledge graph and incremental learning methods still have some limitations. In the construction of large-scale knowledge graphs and incremental learning, how to efficiently process data, reduce computational overhead, and simultaneously ensure model stability and accuracy remains a challenging problem. Furthermore, how to rationally utilize the knowledge of known materials during incremental learning while effectively integrating the characteristics of new materials into the model is an important research direction in current incremental learning and knowledge distillation techniques.

[0013] To overcome the above problems, a dynamic material property prediction method and system based on incremental learning and knowledge distillation has been proposed. Summary of the Invention

[0014] This invention addresses the shortcomings of existing technologies by providing a dynamic material performance prediction method and system based on incremental learning and knowledge distillation. It aims to solve problems such as cold start for new materials, insufficient model generalization ability, and low update efficiency in existing material performance prediction methods. This method utilizes graph neural networks and knowledge graphs, combined with incremental learning and knowledge distillation techniques, to effectively handle the cold start problem for new materials in material performance prediction. Furthermore, by combining teacher and student models, the predictive ability of the model is continuously optimized. Through this method, material performance prediction can not only utilize historical data for efficient prediction but also rapidly incorporate knowledge about new materials into existing models through incremental learning and knowledge distillation, thereby continuously improving the model's predictive accuracy and generalization ability.

[0015] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0016] In a first aspect, the present invention provides a dynamic material property prediction method based on incremental learning and knowledge distillation, comprising the following steps:

[0017] S1: Acquire input data and normalize it; use an incremental learning framework to combine the characteristics of the new material with existing training data for training;

[0018] S2: Obtain the attribute data of the new material, use the dynamic material encoder to generate an embedding vector for the new material, merge the embedding vector with the embedding vector of the existing material, and update the embedding table in the dynamic material encoder.

[0019] S3: Calculate similarity metrics based on the property data of the new material to find known materials that are similar to the new material;

[0020] S4: Calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; rank the known materials and the new material based on the Mahalanobis distance, and select the known material most similar to the new material as the candidate.

[0021] S5: Use graph neural networks to embed the structural information of materials to generate structural feature vectors; calculate the structural similarity between the new material and known materials; based on the calculated structural similarity, select known materials with high similarity as teacher signals to assist in the performance prediction of the new material;

[0022] S6: Obtain the basic performance prediction results of the new material and use them as the prediction of the student model; combine the basic performance prediction results of the new material and the weighted average of the teacher signal, and perform dynamic fusion based on similarity confidence to obtain the enhanced performance prediction results;

[0023] S7: By dynamically adjusting the weighting ratio of the teacher signal and the student model, and combining it with the confidence of the maximum similarity, the final performance prediction result is obtained.

[0024] Furthermore, in S1, the input data is normalized to obtain a mean of 0 and a standard deviation of 1.

[0025] Furthermore, the input data obtained in S1 is used to map the time series data into a representation in a high-dimensional feature space through an embedding layer.

[0026] Furthermore, in S2, a dynamic material encoder is used to generate an embedding vector for the new material, including: using the dynamic material encoder to map the physical property data of the new material into a low-dimensional embedding space to generate an embedding vector for the new material.

[0027] Furthermore, the generation of the embedding vector is expressed by the following formula:

[0028] ;

[0029] The newly generated embedding vectors are merged with the embedding vectors of the known materials to obtain the updated embedding table:

[0030] ;

[0031] In the above formula, Represents the embedding vector of the new material. These are the physical property data of the new material. It is the mapping function of the dynamic material encoder; This is the updated material embedding table. It is the embedding vector of the known material.

[0032] Furthermore, in S2, the property data of the new material is obtained, including its chemical composition and physical properties.

[0033] Furthermore, in S5, based on the structural feature vectors generated by graph neural network embedding, cosine similarity is used to measure the structural similarity between the new material and known materials.

[0034] Furthermore, the method also includes incremental updates of the knowledge graph, specifically: adding attribute data of the new material to the knowledge graph, calculating the similarity between it and known materials; updating the adjacency matrix of the knowledge graph using the similarity information; and updating the adjacency matrix in the graph convolution operation using the relational data of the new material and its similar materials, ensuring that the material similarity is reflected in the learning of the graph neural network.

[0035] Secondly, the present invention also provides a dynamic material property prediction system based on incremental learning and knowledge distillation, comprising:

[0036] The data acquisition module is used to acquire input data and normalize it; an incremental learning framework is used to combine the characteristics of the new material with existing training data for training.

[0037] The dynamic material encoder module is used to acquire the property data of new materials, generate embedding vectors for new materials through the dynamic material encoder, merge the embedding vectors with the embedding vectors of existing materials, and update the embedding table in the dynamic material encoder.

[0038] The similarity calculation module calculates a similarity metric based on the property data of the new material and searches for known materials that are similar to the new material.

[0039] The physical property vector calculation module is used to calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; based on the Mahalanobis distance, the known materials and the new material are sorted, and the known material most similar to the new material is selected as the candidate.

[0040] The structural feature vector generation module uses a graph neural network to embed the structural information of the material to generate a structural feature vector; it calculates the structural similarity between the new material and known materials; based on the calculated structural similarity, it selects known materials with high similarity as teacher signals to assist in the performance prediction of the new material.

[0041] The performance prediction module is used to obtain the basic performance prediction results of the new material and use them as the prediction of the student model. The basic performance prediction results of the new material and the weighted average of the teacher signal are combined and dynamically fused based on similarity confidence to obtain the enhanced performance prediction results.

[0042] The dynamic fusion module is used to dynamically adjust the weighting ratio of teacher signals and student models, and fuse them with the confidence score of the maximum similarity to obtain the final performance prediction result.

[0043] Furthermore, it also includes an incremental update module, which adds the attribute data of new materials to the knowledge graph, calculates the similarity between the new materials and known materials, updates the adjacency matrix of the knowledge graph using the similarity information, and updates the adjacency matrix in the graph convolution operation using the relational data of the new materials and their similar materials, ensuring that the material similarity is reflected in the learning of the graph neural network.

[0044] In summary, compared with the prior art, the beneficial effects of the above technical solution are:

[0045] (1) The present invention adopts an incremental learning method, which can quickly adapt to the addition of new materials with a small number of updates without completely retraining the model. This incremental update not only significantly reduces training time and computational cost, but also ensures that the model always maintains its adaptability to new materials, avoiding redundant computation of frequent full training;

[0046] (2) The dynamic material encoder in this invention can automatically generate an embedding vector for a new material and add it to the existing embedding table when the new material is added. Unlike traditional methods, this invention not only supports cold start for new materials, but also dynamically updates and optimizes the material embedding, enabling the model to more accurately understand the characteristics of new materials and make reasonable performance predictions. In this way, the model can continuously learn and adapt to new information in the material library, improving the accuracy and reliability of predictions;

[0047] (3) This invention introduces knowledge distillation technology, which enhances the predictive ability of the model by utilizing the similarity of known materials during the prediction of new materials. Traditional methods rely on the student model to predict alone, while this invention introduces the prediction information of the teacher model and combines it with the prediction results of similar materials to dynamically adjust the prediction weights and optimize the prediction performance of new materials. The application of knowledge distillation enables this invention to effectively overcome the problem of data scarcity when processing new materials, transfer knowledge from similar materials, and significantly improve the accuracy and robustness of the prediction results;

[0048] (4) This invention calculates the similarity between materials and then weights and fuses the prediction results of the student model and the teacher model. Compared with traditional single prediction models, this invention adopts a dynamic fusion mechanism based on similarity, which can adaptively adjust the prediction contributions of the two based on the relationship between the new material and similar materials. This method enables the model to more flexibly weigh multiple information sources when facing different materials, thereby providing more accurate performance prediction results;

[0049] (5) This invention enhances the reasoning ability of the model by constructing a knowledge graph and combining it with the similarity of the physical properties of materials. The knowledge graph not only helps to establish the relationship between materials, but also enhances the model's understanding of the implicit connections between materials through graph neural networks. Compared with traditional methods that only use single material information for prediction, this invention utilizes the graph structure to better mine the similarity between materials, thereby improving the reasoning and generalization ability of the prediction model. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the dynamic material property prediction method based on incremental learning and knowledge distillation of the present invention;

[0051] Figure 2 A dynamic material property prediction interface based on incremental learning and knowledge distillation;

[0052] Figure 3 This is a training and evaluation error diagram of the present invention. Detailed Implementation

[0053] The principles and features of the present invention are described below with reference to all the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0054] This invention discloses a dynamic material property prediction method and system based on incremental learning and knowledge distillation.

[0055] Reference Figures 1-3 This invention discloses a dynamic material property prediction method based on incremental learning and knowledge distillation, comprising the following steps:

[0056] S1: Acquire input data and normalize it; use an incremental learning framework to combine the characteristics of the new material with existing training data for training;

[0057] S2: Obtain the attribute data of the new material, use the dynamic material encoder to generate an embedding vector for the new material, merge the embedding vector generated for the new material with the embedding vector of the existing material, and update the embedding table in the dynamic material encoder.

[0058] S3: Calculate similarity metrics based on the property data of the new material to find known materials that are similar to the new material;

[0059] S4: Calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; rank the known materials and the new material based on the Mahalanobis distance, and select the known material most similar to the new material as the candidate.

[0060] S5: Use graph neural networks to embed the structural information of materials to generate structural feature vectors; calculate the structural similarity between the new material and known materials; based on the calculated structural similarity, select known materials with high similarity as teacher signals to assist in the performance prediction of the new material;

[0061] S6: Obtain the basic performance prediction results of the new material and use them as the prediction of the student model; combine the basic performance prediction results of the new material and the weighted average of the teacher signal, and perform dynamic fusion based on similarity confidence to obtain the enhanced performance prediction results;

[0062] S7: By dynamically adjusting the weighting ratio of the teacher signal and the student model, and combining it with the confidence of the maximum similarity, the final performance prediction result is obtained.

[0063] This invention dynamically predicts the performance and cost of new materials during the material design and optimization process. This method continuously optimizes the model and improves prediction accuracy as new materials are introduced, making it applicable to fields such as performance prediction, formulation optimization, and material discovery, and has broad application prospects.

[0064] This invention is particularly applicable to industries such as chemical engineering, materials science, and environmental engineering. It can effectively predict the performance of new materials in practical applications, provide strong support for materials design, reduce R&D costs, and improve the efficiency of material performance discovery and optimization.

[0065] The following is a detailed explanation of each of the above steps:

[0066] S1: Acquire input data and normalize it; use an incremental learning framework to combine the characteristics of the new material with existing training data for training.

[0067] Specifically, the data is normalized to have a mean of 0 and a standard deviation of 1, facilitating subsequent processing and model training. An incremental learning framework is used to continuously combine the characteristics of the new material with existing training data to improve the model's generalization ability.

[0068] The input data acquired is the original material formulation data. The purpose is to transform this raw material formulation data into a format that the neural network can process. Material formulation data typically consists of the proportions, properties, and other features of different materials; therefore, it needs to be mapped into a high-dimensional feature space through an embedding layer to facilitate subsequent model processing. The formula is expressed as:

[0069] ;

[0070] in, Represents material formulation data Embedded representation; It is the weight matrix of the embedding layer; It is the bias term of the embedding layer.

[0071] In this embodiment, the formulation data of all materials is first obtained from a database or dataset, including the components and their proportions in each formulation. Let the material formulation data be... ,in This represents the total number of formula data. For the first Each formulation has a feature vector. Each formulation vector contains all the material components of that formulation and their corresponding proportions. The formulation data is obtained from experiments, theories, or simulations.

[0072] Data normalization is a crucial step in machine learning and deep learning, significantly improving model training speed and stability. By standardizing the input data, the influence of different units of measurement is eliminated, resulting in data with zero mean and unit variance. The formula is as follows:

[0073] ;

[0074] in, It is the input data, namely the material formulation data; It is the mean of the input data; It is the standard deviation of the input data; It is the normalized data.

[0075] Incremental learning is a method that allows a model to progressively adapt to new data. In this step, new material data is combined with existing training data to train the model. This method ensures that the model can be dynamically updated when receiving new material data and maintains high generalization ability. The formula is expressed as:

[0076] ;

[0077] in, It is the loss function for incremental training, where It is a dataset; It is the first The loss function for each data point represents the model's prediction error; It is a regularization term used to prevent the model from overfitting; It is the regularization coefficient, which controls the influence of the regularization term.

[0078] S2: Obtain the attribute data of the new material, use the dynamic material encoder to generate an embedding vector for the new material, merge the generated embedding vector of the new material with the embedding vector of the existing material, and update the embedding table in the dynamic material encoder.

[0079] Specifically, physical property data corresponding to the formulation is obtained from materials databases or experimental measurement data. Let the physical property data be... ,in It is the first A vector of physical property data for each formulation. The property data of new materials typically includes their chemical composition, physical properties, mechanical properties, thermal properties, and electrical properties.

[0080] In this step, the properties of the new material are input into a dynamic material encoder, which generates an embedding vector of the new material via a neural network for subsequent performance prediction. The formula for generating the embedding vector is expressed as:

[0081] ;

[0082] in, This represents the embedding vector generated by the new material; This represents the property data of the new material; It is the mapping function of the dynamic material encoder.

[0083] The newly generated embedding vectors are merged with the embedding vectors of the known materials to obtain the updated embedding table:

[0084] ;

[0085] in, This is the updated material embedding table; It is a set of embedding vectors of existing materials.

[0086] In this process, new material embedding vectors are merged with existing material embedding vectors to form a new material embedding table. This embedding table expands continuously as new materials are added, thereby enhancing the model's adaptability to new materials.

[0087] S3: Calculate similarity metrics based on the property data of the new material to find known materials that are similar to the new material.

[0088] Specifically, in the prediction process of new materials, the teacher model is first used, which obtains the prediction signal by using the similarity of known materials. By calculating the similarity between the new material and existing materials, the most similar material is selected, and the prediction results of these materials guide the prediction of the new material. The formula is expressed as:

[0089] ;

[0090] in, This represents the prediction results of the teacher model; It is the number of similar materials selected; It is the first The weights of similar materials are calculated based on similarity. It is the first Prediction results for similar materials.

[0091] By calculating the similarity between the new material and existing materials, a weighted average is used to fuse the outputs of the teacher model (predictions based on similar materials) and the student model (predictions based on the current formulation). This method effectively utilizes information from similar materials to improve the prediction accuracy of the new material. The formula is expressed as:

[0092] ;

[0093] in, This is the final performance prediction result; These are the prediction results from the student model; It is the fusion coefficient, which represents the weighted ratio of the student model and the teacher model, and is usually calculated based on the similarity confidence score.

[0094] S4: Calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; sort the known materials and the new material based on the Mahalanobis distance, and select the known material that is most similar to the new material as the candidate.

[0095] Specifically, the physical property data of the new material and known materials are first extracted, and then the Mahalanobis distance between the materials is calculated based on this data. This Mahalanobis distance measures the similarity between the new material and known materials in the property space, and the formula is as follows:

[0096] ;

[0097] in, For attribute similarity; For the physical property data of new materials; For known physical property data of materials; This is the Mahalanobis distance.

[0098] The similarity between the new material and known materials calculated using Mahalanobis distance is used to select similar materials in formulation prediction. Based on the calculated Mahalanobis distance, the similarity value between the new material and known materials is used to select the known material most similar to the new material and use it as a candidate material for formulation optimization. Materials with higher similarity are assigned higher weights.

[0099] S5: Use a graph neural network to embed the structural information of the material to generate a structural feature vector; calculate the structural similarity between the new material and known materials; based on the calculated structural similarity, select known materials with high similarity as teacher signals to assist in the performance prediction of the new material.

[0100] In the process of predicting the performance of new materials, knowledge distillation technology is applied to extract teacher signals from the prediction signals of similar materials. The prediction results of similar materials serve as teacher signals, guiding the student model (i.e., the direct prediction of the new material) to learn better. This process is achieved by calculating the similarity between similar materials and the new material.

[0101] Includes the following sub-steps:

[0102] S501: Use a graph neural network (GNN) to encode the material's structural information and generate the material's structural feature vector.

[0103] Here, the structural information of the materials refers to the graph structure information of the material similarity knowledge graph (similarity relationship graph). In this step, a graph neural network (GNN) is used to aggregate the similarity information between materials. Through the graph neural network, relational information can be effectively extracted from the material similarity graph. This allows the prediction results for new materials to take into account their similarity to other materials, thereby improving prediction accuracy.

[0104] The system utilizes a graph neural network to encode the structural features of materials, generating a structural feature vector for each material. The graph neural network can capture the graph structure information between materials, thus providing an effective representation of structural features, as expressed in the following formula:

[0105] ;

[0106] in, The structural feature vector of the material; Let be the adjacency matrix of the materials, representing the graph structure between the materials; Let be the nodal feature matrix of the material, representing the basic properties of the material.

[0107] S502: Calculate the structural similarity between the new material and known materials, and evaluate it based on the cosine similarity of structural features.

[0108] The structural similarity between new materials and known materials can be evaluated using cosine similarity based on the structural feature vectors generated by graph neural networks, as shown in the following formula:

[0109] ;

[0110] in, For structural similarity; This represents the structural feature vector of the new material; The structural feature vector of the known material.

[0111] S6: Obtain the basic performance prediction results of the new material and use them as the prediction of the student model; combine the basic performance prediction results of the new material and the weighted average of the teacher signal, and perform dynamic fusion based on similarity confidence to obtain the enhanced performance prediction results.

[0112] Includes the following steps:

[0113] S601: Obtain the basic prediction results of the new material and use them as the prediction output of the student model.

[0114] In this step, the model uses a student network to make basic predictions about the new material, obtaining the prediction results, expressed by the following formula:

[0115] ;

[0116] in: The prediction results for the student model; This refers to the formulation characteristics of the new material.

[0117] S602: Obtain performance prediction results of similar materials to serve as teacher signals to guide the optimization of student models.

[0118] The system generates teacher signals using a teacher model based on similar materials, and combines these signals with the prediction results from the student model. A weighted average is then used to calculate the enhanced prediction results for the new materials, as shown in the following formula:

[0119] ;

[0120] in, This is the weighted average prediction result of the teacher model; For the first The weights of similar materials are calculated based on their similarity to the new material; For the first Prediction results for similar materials.

[0121] S7: By dynamically adjusting the weighting ratio of the teacher signal and the student model, and combining it with the confidence of the maximum similarity, the final performance prediction result is obtained.

[0122] Specifically, the system uses a knowledge graph to query known materials similar to the new material, and uses their prediction results as teacher signals to assist in optimizing the prediction of the new material. The formula is as follows:

[0123] ;

[0124] in, The enhanced prediction results; The fusion coefficient is dynamically adjusted based on the similarity confidence level. The basic prediction results for the student model; This represents the weighted prediction results of the teacher model.

[0125] S8: Incremental update of the knowledge graph, specifically: adding the attribute data of the new material to the knowledge graph, calculating the similarity between it and known materials; updating the adjacency matrix of the knowledge graph using the similarity information; and updating the adjacency matrix in the graph convolution operation using the relational data of the new material and its similar materials to ensure that the material similarity is reflected in the learning of the graph neural network.

[0126] In this step, the system adds the physical property data of the new material to the knowledge graph and calculates the similarity between the new material and existing materials to enhance the graph structure of the graph neural network. The formula is as follows:

[0127] ;

[0128] in: For material similarity based on knowledge graphs; This represents the physical property vector of the new material. This is a vector of the physical properties of existing materials.

[0129] The adjacency matrix of the knowledge graph is updated based on the similarity between new and existing materials. The updated adjacency matrix reflects the relationship between new and existing materials, and information is propagated through graph convolution operations. The formula is as follows:

[0130] ;

[0131] in, This is the updated adjacency matrix; Given the existing adjacency matrix; This is newly added similarity information, representing the relationship between the new material and existing materials.

[0132] Analysis of experimental results:

[0133] When conducting experiments using publicly available material property datasets, all experimental data are analyzed using pre-defined formulations and material property parameters, such as... Figure 3 As shown in the figure, the upper part represents the change in training algebras and training / testing errors, while the lower part represents the change in model gradients. The obtained prediction results cover multiple performance indicators, such as gloss, hardness, and density. The error between the predicted and actual values ​​remains within a low range, between 0.05 and 0.15 (after 30 training and testing generations). This indicates that the method of this invention can accurately reflect the actual material properties.

[0134] This invention, based on incremental learning and knowledge distillation techniques, exhibits superior prediction accuracy. Especially with specific formulations and the addition of new materials, the model can quickly adapt and generate relatively accurate predictions. The predicted values ​​in the figure generally match the trends of the actual values, particularly in some stable regions, where the model effectively captures performance change trends, demonstrating strong fitting ability and adaptability.

[0135] Compared to traditional material prediction methods, this invention consumes fewer computational resources. Through an incremental learning mechanism, the introduction of new materials does not require retraining the entire model from scratch, reducing computational costs and improving training efficiency. This method also offers high computational flexibility, enabling rapid processing of new data and continuous improvement in prediction accuracy.

[0136] Secondly, embodiments of the present invention also provide a dynamic material performance prediction system based on incremental learning and knowledge distillation, including a data acquisition module, a dynamic material encoder module, a similarity calculation module, a physical property vector calculation module, a structural feature vector generation module, a performance prediction module, a dynamic fusion module, and an incremental update module.

[0137] The data acquisition module is used to acquire input data and normalize it; an incremental learning framework is used to combine the characteristics of the new material with existing training data for training.

[0138] The primary purpose of this module is to support incremental learning, which involves adding new material data to the model incrementally without retraining it from scratch. Using this incremental learning strategy, the model is only partially updated each time a new material is added, avoiding the consumption of significant computational resources. The new material's knowledge is updated by merging with the embedding vectors of existing materials, thereby enhancing the model's adaptability to new data. During incremental training, the model continuously learns from new formulation and material performance data, constantly improving its predictive capabilities.

[0139] The dynamic material encoder module is used to acquire the property data of new materials, generate embedding vectors for new materials through the dynamic material encoder, merge the embedding vectors with the embedding vectors of existing materials, and update the embedding table in the dynamic material encoder.

[0140] The core task of this module is to address the cold-start problem of new materials and generate embedding vectors for subsequent prediction tasks. A neural network structure is used to encode the physical properties of new materials, generating their embedding vectors. Independent embedding representations are created for each new material and updated using data from existing materials. The dynamic material encoder continuously updates the embedding table during training, ensuring that embeddings of new materials are promptly incorporated into the training model. This approach effectively enables rapid integration of new materials and allows the model to adapt to new inputs.

[0141] The similarity calculation module calculates a similarity metric based on the property data of the new material, and searches for known materials that are similar to the new material.

[0142] This module utilizes a knowledge graph, enabling the system to manage and leverage the similarity relationships between materials. This provides auxiliary information for predicting new materials, enhancing the model's predictive capabilities. The knowledge graph constructs a similarity relationship diagram based on the physical property similarities between materials. Employing a graph neural network approach, it combines graph structure information with material physical properties and formulation information to improve prediction accuracy. Whenever a new material is added, the graph updates the corresponding nodes and edges, reflecting changes in material similarity in real time, thereby improving the model's accurate prediction of new materials.

[0143] The physical property vector calculation module is used to calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; based on the Mahalanobis distance, the known materials and the new material are sorted, and the known material most similar to the new material is selected as the candidate.

[0144] This module is responsible for calculating the similarity between materials and providing auxiliary signals for predicting the performance of new materials based on this similarity. Multiple similarity calculation methods are employed, such as Mahalanobis distance and cosine similarity, to evaluate the similarity between materials from two dimensions: attribute space and structural features, respectively. By calculating similarity weights, the model is guided on how to perform weighted fusion of prediction results from similar materials when predicting new materials. The similarity calculation is not only based on the physical properties of the materials but also incorporates structural information embedded in the graph neural network, thus comprehensively considering multiple aspects of the material's characteristics.

[0145] The structural feature vector generation module uses a graph neural network to embed the structural information of the material to generate a structural feature vector; it calculates the structural similarity between the new material and known materials; based on the calculated structural similarity, it selects known materials with high similarity as teacher signals to assist in the performance prediction of the new material.

[0146] This module improves the predictive performance of the student model on new material data by passing the prediction information from the "teacher model" to the "student model," especially when sufficient new material data is lacking. The student model learns by using the output of the teacher model's predictions of similar materials (i.e., the teacher signal). The similarity between the new material and similar materials is calculated, and the prediction weights of the student model for the new material are adjusted accordingly. The prediction performance is improved by weighted fusion of the outputs of the student and teacher models. During the weight fusion process, a temperature control mechanism is used to dynamically adjust the prediction weights of the student and teacher models based on the similarity confidence level, thereby optimizing the prediction results.

[0147] The performance prediction module is used to obtain the basic performance prediction results of the new material and use them as the prediction of the student model. The basic performance prediction results of the new material and the weighted average of the teacher signal are combined and dynamically fused based on similarity confidence to obtain the enhanced performance prediction results.

[0148] This module is responsible for predicting multiple performance targets, including not only material properties (such as mechanical and thermal properties) but also costs. A multi-objective prediction head was designed using a multi-objective learning method, comprising performance prediction and cost prediction. Multiple output tasks are trained simultaneously within the same model, improving training efficiency and model generalization ability by sharing network parts. The output layer adapts to the needs of different targets, generating prediction results relevant to each task.

[0149] The dynamic fusion module is used to dynamically adjust the weighting ratio of teacher signals and student models, and fuse them with the confidence score of the maximum similarity to obtain the final performance prediction result.

[0150] This module achieves more accurate predictions by dynamically fusing the predictions from student and teacher models, especially when dealing with new materials. It utilizes a similarity-based confidence mapping to weightedly fuse the predictions from the teacher and student models. Through dynamic control of similarity, the influence of the teacher model and the predictive ability of the student model are balanced in the prediction of new materials, thereby optimizing the overall prediction results. A temperature parameter is used to control the mapping from similarity to the fusion coefficient, enabling fine-tuning of the prediction results.

[0151] The incremental update module is used to add the attribute data of new materials to the knowledge graph, calculate the similarity between them and known materials, update the adjacency matrix of the knowledge graph using the similarity information, and update the adjacency matrix in the graph convolution operation using the relationship data of the new materials and their similar materials, so as to ensure that the material similarity is reflected in the learning of the graph neural network.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic material property prediction method based on incremental learning and knowledge distillation, characterized in that, Includes the following steps: S1: Acquire input data and normalize it; use an incremental learning framework to combine the characteristics of the new material with existing training data for training; S2: Obtain the attribute data of the new material, use the dynamic material encoder to generate an embedding vector for the new material, merge the embedding vector generated for the new material with the embedding vector of the existing material, and update the embedding table in the dynamic material encoder. S3: Calculate similarity metrics based on the property data of the new material to find known materials that are similar to the new material; S4: Calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; rank the known materials and the new material based on the Mahalanobis distance, and select the known material most similar to the new material as the candidate. S5: Use graph neural networks to embed the structural information of materials to generate structural feature vectors; calculate the structural similarity between the new material and known materials; based on the calculated structural similarity, select known materials with high similarity as teacher signals to assist in the performance prediction of the new material; S6: Obtain the basic performance prediction results of the new material and use them as the prediction of the student model; combine the basic performance prediction results of the new material and the weighted average of the teacher signal, and perform dynamic fusion based on similarity confidence to obtain the enhanced performance prediction results; S7: By dynamically adjusting the weighting ratio of the teacher signal and the student model, and combining it with the confidence of the maximum similarity, the final performance prediction result is obtained.

2. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, In S1, the input data is normalized to obtain a mean of 0 and a standard deviation of 1.

3. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, The input data obtained in S1 is used to map the time series data into a representation in a high-dimensional feature space through the embedding layer.

4. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, In S2, a dynamic material encoder is used to generate an embedding vector for the new material, including: using the dynamic material encoder to map the physical property data of the new material into a low-dimensional embedding space to generate an embedding vector for the new material.

5. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 4, characterized in that, The generation of the embedding vector is expressed by the following formula: ； The newly generated embedding vectors are merged with the embedding vectors of the known materials to obtain the updated embedding table: ； In the above formula, Represents the embedding vector of the new material. These are the physical property data of the new material. It is the mapping function of the dynamic material encoder; This is the updated material embedding table. It is the embedding vector of the known material.

6. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, In S2, the property data of the new material is obtained, including its chemical composition and physical properties.

7. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, In S5, based on the structural feature vectors generated by graph neural network embedding, cosine similarity is used to measure the structural similarity between new materials and known materials.

8. The dynamic material property prediction method based on incremental learning and knowledge distillation according to claim 1, characterized in that, The method also includes incremental updates of the knowledge graph, specifically: adding attribute data of the new material to the knowledge graph, calculating the similarity between the new material and known materials; updating the adjacency matrix of the knowledge graph using the similarity information; and updating the adjacency matrix in the graph convolution operation using the relationship data of the new material and its similar materials, ensuring that the material similarity is reflected in the learning of the graph neural network.

9. A dynamic material property prediction system based on incremental learning and knowledge distillation, characterized in that, include: The data acquisition module is used to acquire input data and perform normalization processing on the data; An incremental learning framework is used to combine the characteristics of the new material with existing training data for training. The dynamic material encoder module is used to acquire the property data of new materials, generate embedding vectors for new materials through the dynamic material encoder, merge the embedding vectors generated for new materials with the embedding vectors of existing materials, and update the embedding table in the dynamic material encoder. The similarity calculation module calculates a similarity metric based on the property data of the new material and searches for known materials that are similar to the new material. The physical property vector calculation module is used to calculate the physical property vector of the material and calculate the Mahalanobis distance between the new material and the known materials; based on the Mahalanobis distance, the known materials and the new material are sorted, and the known material most similar to the new material is selected as the candidate. The structural feature vector generation module uses a graph neural network to embed the structural information of the material to generate a structural feature vector; it calculates the structural similarity between the new material and known materials; based on the calculated structural similarity, it selects known materials with high similarity as teacher signals to assist in the performance prediction of the new material. The performance prediction module is used to obtain the basic performance prediction results of new materials and use them as predictions for the student model. By combining the basic performance prediction results of the new material with the weighted average of the teacher's signal, dynamic fusion is performed based on similarity confidence to obtain the enhanced performance prediction results; The dynamic fusion module is used to dynamically adjust the weighting ratio of teacher signals and student models, and fuse them with the confidence score of the maximum similarity to obtain the final performance prediction result.

10. The dynamic material property prediction system based on incremental learning and knowledge distillation according to claim 9, characterized in that, It also includes an incremental update module, which is used to add the attribute data of new materials to the knowledge graph and calculate the similarity between them and known materials; And the adjacency matrix of the knowledge graph is updated using similarity information; By using relational data of new materials and their similar materials, the adjacency matrix in the graph convolution operation is updated to ensure that material similarity is reflected in the learning of the graph neural network.

Images