Material performance prediction and standardization method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence material prediction, and discloses a material performance prediction and standardization method and system based on artificial intelligence, and the method comprises the steps: collecting artificial intelligence material experiment data, calculating material science simulation data, and carrying out the standardization processing of the data; based on the physical conservation law of artificial intelligence materials, constructing a physical model comprising partial differential equation constraints; a deep neural network is combined with a graph neural network to construct an artificial intelligence material performance prediction model. The system comprises a data acquisition module, a physical modeling module, an artificial intelligence training module, a model optimization module and a prediction and standardization module. By adopting the technical scheme of combining the deep neural network and the graph neural network, the technical effect of accurately predicting the material performance is achieved, and the complex nonlinear relation and the deep relation between the material crystal structure and the performance can be better captured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence material prediction technology, and specifically to an artificial intelligence-based material property prediction and standardization method and system. Background Art

[0002] Materials science is an important foundation for the development of modern industry and technology. Especially in the fields of aerospace, energy, electronics, manufacturing, etc., the prediction and optimization of material properties are crucial to promoting technological progress. With the continuous emergence of new materials and the increasing demand for complex performance, traditional experimental and theoretical calculation methods can no longer meet the demand for rapid and accurate prediction of material properties. In the past, the prediction of material properties mainly relied on empirical formulas, experimental data and physical modeling, but these methods have many limitations.

[0003] Existing material performance prediction methods mainly include empirical models based on experimental data and simulation methods based on computational materials science. In traditional experimental-based data-driven methods, researchers determine various performance indicators of materials through repeated experiments. However, the experimental data acquisition cycle is long, the cost is high, and it is difficult to cover the characteristics of all materials, which limits the wide applicability of the prediction results. Computational materials science predicts the performance of materials at the microscopic level through methods such as simulating molecular dynamics (MD) and density functional theory (DFT).

[0004] However, when existing material performance predictions rely on a single experimental data or physical model for prediction, they are often limited by insufficient data and low simulation accuracy. Traditional experimental methods require a lot of time and resources, and can only predict limited materials, making it difficult to meet the needs of new materials. At the same time, traditional physical models have limited predictive capabilities when facing complex material structures or multi-physical field coupling, and it is difficult to fully reflect the full picture of material performance. Most of the existing artificial intelligence models are purely data-driven, lack physical constraints, and the prior art lacks standardized data storage and sharing mechanisms. Although many material databases have accumulated a large amount of experimental and computational data, due to the inconsistent data storage format and poor cross-platform compatibility, data sharing between different research teams is difficult, limiting data reuse and resource sharing. Therefore, the present invention provides a material performance prediction and standardization method and system based on artificial intelligence to address the deficiencies in the prior art. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a material property prediction and standardization method and system based on artificial intelligence, which solves the problem that the existing material property prediction relies on a single experimental data or physical model for prediction, which is often limited by insufficient data and low simulation accuracy.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for predicting and standardizing material properties based on artificial intelligence, comprising the following steps:

[0007] Collect artificial intelligence material experimental data and computational materials science simulation data, and perform standardization processing on the data;

[0008] Based on the physical conservation laws of artificial intelligence materials, construct a physical model including partial differential equation constraints, and establish mathematical descriptions of the mechanical, thermal, and electromagnetic properties of materials;

[0009] Adopt a deep neural network combined with a graph neural network to construct a prediction model for the properties of artificial intelligence materials, and introduce a physical constraint loss function during the training process;

[0010] Optimize the prediction model for the properties of artificial intelligence materials based on the variational inference method, improve the generalization ability of the prediction in a low-data environment, and adopt a regularization strategy to reduce overfitting;

[0011] Based on the optimized prediction model, store the prediction results in a standardized data format, and construct a material database in combination with a knowledge graph for data sharing and reuse.

[0012] Preferably, the standardization processing includes the following steps:

[0013] Data unit conversion to ensure that the material experimental data and computational simulation data adopt a unified international system of units;

[0014] Format normalization, adopting a standardized data storage format;

[0015] Data cleaning, filling in missing data, removing abnormal data, and correcting measurement errors;

[0016] Feature normalization, adopting maximum-minimum normalization or standardization for numerical variables.

[0017] Preferably, the physical model with partial differential equation constraints includes:

[0018] Mechanical property model, using the stress-strain balance equation to describe the mechanical properties of artificial intelligence materials;

[0019] Thermal property model, using the heat conduction equation to describe the temperature distribution of artificial intelligence materials;

[0020] Electromagnetic property model, using Maxwell's equations to describe the electromagnetic properties of artificial intelligence materials.

[0021] Preferably, the physical constraint loss function includes:

[0022] Physical constraint loss based on partial differential equations, used to ensure that the performance prediction of artificial intelligence materials conforms to basic physical laws;

[0023] The physical constraint loss based on Noether's theorem is used to maintain the energy conservation and momentum conservation of the system;

[0024] The physical constraint loss based on gauge field theory is used for the generalization ability of the model under different coordinate transformations.

[0025] Preferably, the deep neural network adopts a multi-layer fully connected network structure to learn the non-linear performance relationship of materials, and the graph neural network is used to learn the crystal structure characteristics of materials and capture the mapping relationship between material structure and performance.

[0026] Preferably, the variational inference method adopts a variational Bayesian optimization strategy, introduces a divergence constraint to optimize the distribution of the prediction model parameters, and is used to keep the prediction model stable in a low-data environment.

[0027] Preferably, the standardized data format adopts a data storage method that conforms to international material standards, and constructs the association relationship of material attributes based on the knowledge graph to improve the reusability and interoperability of data.

[0028] Preferably, the prediction model of the artificial intelligence material performance is analyzed by using explainable artificial intelligence technology, including:

[0029] Using the SHAP method to analyze the key input features of model prediction;

[0030] Using the LIME method to provide local interpretability to enhance the transparency of prediction results.

[0031] Preferably, the material database combines blockchain technology for data storage to ensure the integrity, security and traceability of data and support dynamic data updates.

[0032] There is also provided an artificial intelligence-based material performance prediction and standardization system, including:

[0033] A data acquisition module for collecting experimental data of artificial intelligence materials and computational materials science simulation data and performing standardized processing on the data;

[0034] A physical modeling module for constructing a physical constraint model based on the physical conservation laws of materials, including mathematical descriptions of mechanical properties, thermal properties and electromagnetic properties;

[0035] An artificial intelligence training module for constructing a material performance prediction model by using a deep neural network combined with a graph neural network and introducing a physical constraint loss function during the training process;

[0036] A model optimization module for optimizing the material performance prediction model based on the variational inference method for the generalization ability and robustness of prediction;

[0037] A prediction and standardization module, which is used to store the optimized prediction results and construct a material database in combination with a knowledge graph for data sharing and reuse.

[0038] The present invention provides a method and system for predicting and standardizing material properties based on artificial intelligence. It has the following beneficial effects:

[0039] 1. The present invention adopts a technical solution that combines a deep neural network and a graph neural network, achieving the technical effect of accurately predicting material properties; compared with the single traditional machine learning method in the prior art, the deep neural network and the graph neural network can better capture complex non-linear relationships and the deep connections between the crystal structure and properties of materials, solving the inefficiency and limitations of traditional methods in processing material structure data.

[0040] 2. The present invention corrects the prediction results during the training process of the artificial intelligence model by introducing a physical constraint loss function, ensuring that the results conform to physical laws; compared with the pure data-driven method lacking physical constraints in the prior art, the present invention effectively avoids the problem of unreasonable predictions caused by the model ignoring physical laws, enhancing the physical consistency and scientificity of the prediction results.

[0041] 3. The present invention uses the variational inference method to optimize the model, solving the problem of poor generalization ability of the model when data is scarce; through this optimization scheme, the prediction accuracy of the model in a low-data environment is significantly improved. Compared with the overfitting problem that easily occurs in the prior art, the variational inference method effectively enhances the robustness and stability of the model, ensuring its reliability under different data conditions.

[0042] 4. The present invention stores data in a standardized format and combines knowledge graph technology, achieving the technical effect of improving data sharing and reuse; compared with the practice lacking a unified storage and sharing mechanism in the prior art, the technical solution of the present invention enables seamless docking of material data from different sources, improving the operability and cross-domain application value of material property prediction results globally. Description of the Drawings

[0043] Figure 1 It is a flowchart of the method steps of the present invention;

[0044] Figure 2 It is a system architecture diagram of the present invention. Detailed Embodiments

[0045] Next, in combination with the accompanying drawings of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] Please refer to the attached Figure 1 , the embodiments of the present invention provide an artificial intelligence-based material property prediction and standardization method, including the following steps:

[0047] S1. Collect artificial intelligence material experimental data and computational materials science simulation data, and perform standardization processing on the data;

[0048] S2. Based on the physical conservation laws of artificial intelligence materials, construct a physical model including partial differential equation constraints, and establish mathematical descriptions of the mechanical, thermal, and electromagnetic properties of materials;

[0049] S3. Use a deep neural network combined with a graph neural network to construct a prediction model for artificial intelligence material properties, and introduce a physical constraint loss function during the training process;

[0050] S4. Optimize the prediction model for artificial intelligence material properties based on the variational inference method, improve the generalization ability of the prediction in a low-data environment, and adopt a regularization strategy to reduce overfitting;

[0051] S5. Based on the optimized prediction model, store the prediction results in a standardized data format, and construct a material database in combination with a knowledge graph for data sharing and reuse.

[0052] For step S1, in this embodiment, the data sources relied on for material property prediction are extensive, including but not limited to experimental data and computational simulation data. Experimental data usually comes from various material experiments, and the variables included therein include mechanical properties, thermal properties, electromagnetic properties, etc. Computational simulation data is obtained through theory-based computational methods (such as density functional theory, molecular dynamics simulation, etc.), and usually includes computational results related to the internal structure of materials, electronic characteristics, and temperature, etc.

[0053] As an option, the experimental data and computational simulation data may be sourced from different databases, such as the MaterialsProject, AFLOW database, etc. During the data collection process, these data often have multiple formats, inconsistent units, missing values, and noisy data. In order to enable these data to be effectively used for machine learning modeling, it is necessary to perform standardization processing on them.

[0054] Specifically, the standardization processing in this embodiment includes the following aspects:

[0055] Data unit conversion: First, the units of all material experimental data and computational simulation data are unified into the International System of Units (SI units). The stress unit is converted from kPa to MPa, and the temperature is converted from Celsius to Kelvin (K). This step is the basis for ensuring the interoperability of data from different sources and providing consistent inputs for subsequent model training. Data unit conversion is crucial for ensuring the accuracy of calculation results, especially when there are differences in the unit systems of different experimental data.

[0056] Format normalization: All original data formats are converted through predefined standard formats. For example, the recording methods of experimental data may have different storage formats, such as CSV, Excel, or database tables, while computational simulation data may be stored in JSON or XML formats. For ease of processing, this data is converted into a unified standard format, such as the CIF format for database storage, in this step. This format unification not only facilitates subsequent processing but also ensures the reusability of the data.

[0057] Data cleaning: In some embodiments, the collected data may have missing values, outliers, or inaccurate data due to experimental errors. To ensure the quality and integrity of the data, the missing data is first filled. Common filling methods include filling based on the mean, median of the data, or through interpolation methods. For outliers, they are detected through statistical methods and removed. Incorrect data is corrected according to the actual situation to ensure that the data can most accurately reflect the true properties of the material.

[0058] Feature normalization: In some embodiments, to avoid instability in model training caused by differences in the dimensions of variables, all input data is subjected to feature normalization. Specifically, all numerical variables are transformed through min-max normalization or standardization methods. For example, for the stress values in mechanical property data, they are standardized so that their mean is 0 and the standard deviation is 1. This process effectively avoids variables with overly large numerical ranges having too much impact on model training and helps to accelerate the convergence of the model.

[0059] Through the above standardization process, all data is unified in terms of units, formats, and numerical ranges, enabling the artificial intelligence models adopted in subsequent steps to learn and predict material properties more efficiently and accurately.

[0060] In one possible implementation, the standardized data can also be stored in a cloud database in an automated manner for convenient access by different users and systems. For example, the standardized data can be uploaded to a service based on RESTful API for use by other developers or researchers, ensuring the sharing and availability of the data.

[0061] For step S2, in this embodiment, the properties of materials, such as mechanical properties, thermal properties, and electromagnetic properties, are constrained by a series of physical laws. By introducing these physical constraints, the prediction range of the model can be effectively limited, avoiding situations that do not conform to actual physical phenomena. To this end, the present invention designs a model based on physical constraints, and specifically uses partial differential equations (PDEs) to describe these properties.

[0062] As an option, the modeling of mechanical properties, thermal properties, and electromagnetic properties is respectively based on the stress-strain equilibrium equation, the heat conduction equation, and the Maxwell equation.

[0063] The modeling of mechanical properties is achieved through the stress-strain equilibrium equation. Specifically, there is a close relationship between the stress and deformation of materials under external forces. The stress-strain relationship can be described by the following equation:

[0064]

[0065] where σ represents the stress tensor, f is the external force density, is the divergence of the stress. This equation is used to describe the stress distribution of materials under external forces and its relationship with deformation.

[0066] Specifically, under external forces such as tension or compression, the degree of deformation of materials is measured by the strain ∈. By establishing the relationship between stress and strain, the mechanical behavior of materials under different external force conditions can be simulated.

[0067] The modeling of thermal properties is based on the heat conduction equation. In many engineering applications, the thermal response of materials is crucial for their properties, especially in heat exchange and energy transfer. The heat conduction equation is usually written as:

[0068]

[0069] where ρ is the density of the material, c p is the specific heat capacity, T is the temperature, k is the thermal conductivity, Q is the heat source term, is the heat conduction equation. This equation describes the heat transfer process of materials during the time change process and the temperature distribution under different heat source conditions. By solving this equation, the heat conduction performance of materials under different temperature gradients can be obtained, and then their thermal response characteristics can be predicted.

[0070] Electromagnetic properties are the characteristics exhibited by materials in an electromagnetic field, such as conductivity, permittivity, etc. The modeling of electromagnetic properties is mainly described by the Maxwell equation. The Maxwell equations include the relationship between the electric field and the magnetic field, and the general form is as follows:

[0071]

[0072] wherein, is the curl of the electric field E, E represents the electric field, and B represents the magnetic field. is the rate of change of the magnetic field B with respect to time. This equation is used to describe the relationship between the electric field and the magnetic field, and thus predict the behavior of materials in electromagnetic fields. For some specific materials, Maxwell's equations can predict their electromagnetic wave propagation, reflection, refraction and other characteristics.

[0073] The main purpose of introducing these physical constraints is to ensure that the predictions of the model conform to the basic laws of materials science and avoid predictions that do not conform to physical reality. For example, when simulating mechanical properties, if the stress-strain relationship is not considered, it may lead to unreasonable material deformation predicted by the model. The thermal and electromagnetic property models avoid incorrect predictions of the model by constraining the heat conduction and electromagnetic response of materials.

[0074] In one possible implementation, physical modeling can also perform microscopic modeling in combination with the macroscopic structure of the material. For example, when simulating complex materials, it may be necessary to consider multi-level and multi-scale modeling methods, and further improve the prediction accuracy through accurate modeling of the microscopic structure.

[0075] To effectively incorporate these physical constraints into the training of the artificial intelligence model, the present invention designs a physical constraint loss function. This loss function includes physical constraints in multiple aspects, including constraints of partial differential equations, physical laws such as energy conservation and momentum conservation. During the training process, the loss function will correct the prediction results of the model to ensure that they meet the requirements of the physical model.

[0076] Specifically, the partial differential equation constraint can ensure that the model output is consistent with the actual physical laws, and energy conservation and momentum conservation ensure the conservation of the system during the prediction process. For example, for the mechanical property model, the stress prediction result of the model will be corrected through the loss function during the training process to make it conform to the stress-strain equilibrium equation.

[0077] By introducing these physical constraints, the material property prediction method of the present invention can avoid the biases that may occur in purely data-driven models, making the final prediction results more reliable.

[0078] For step S3, in this embodiment, a deep neural network (DNN) combined with a graph neural network (GNN) is used to construct an artificial intelligence-based material property prediction model, and a physical constraint loss function is introduced during the training process. An accurate and physically consistent material property prediction model is established, and the data-driven method is combined with the principles of physics to optimize the accuracy and reliability of the prediction.

[0079] Generally, deep neural networks and graph neural networks play important roles in material property prediction. Deep neural networks can handle complex non-linear relationships, while graph neural networks can effectively learn the crystal structure of materials and the mapping relationship between the structure and properties. Combining the advantages of both, the present invention can learn complex laws from a large amount of experimental and computational simulation data.

[0080] As an option, graph neural networks are particularly suitable for dealing with the crystal structure characteristics of materials. The crystal structure of materials determines the basis of their mechanical, thermal, electromagnetic and other properties. Therefore, in graph neural networks, each atom or molecule is regarded as a node in the graph, and the interaction between atoms is regarded as an edge. Such a structure can effectively capture the complex relationship between the structure and properties.

[0081] Specifically, a deep neural network (DNN) captures the complex characteristics of material properties through multi-layer non-linear transformations, especially excelling in large-scale data processing. A graph neural network (GNN) can perform convolutional operations on the graph structure based on the structural data of materials, updating the information of each node through the relationship between adjacent nodes, thereby learning deeper structural characteristics.

[0082] During the training process, in order to ensure the physical consistency of the model, this embodiment introduces a physical constraint loss function. Specifically, the loss function includes the following important physical constraint terms:

[0083] Physical constraint loss based on partial differential equations: This term is used to ensure that the prediction results of material properties conform to known physical laws. In the prediction of mechanical properties, by introducing the stress-strain balance equation constraint, the output results of the model should be consistent with the physical laws. Such physical constraints make the model follow basic physical laws during prediction by adding additional loss terms. This loss term is usually expressed as:

[0084]

[0085] where σ represents the stress tensor, f is the external force density, represents the divergence of stress, and the loss function corrects the prediction of the model by minimizing this term.

[0086] Physical constraint loss based on Noether's theorem: Noether's theorem emphasizes the conservation of physical systems. In the prediction of material properties, energy conservation and momentum conservation are very important constraints. By incorporating these conservation conditions into the loss function, it can be ensured that the prediction results of the model conform to the basic conservation laws of physical systems. The physical constraint loss of Noether's theorem usually includes restrictions on the changes in the energy and momentum of the system:

[0087]

[0088] Among them, \(L\) Noether represents the conserved quantity obtained from Noether's theorem, where \(\mathcal{L}\) represents the Lagrangian, \(q\) is the generalized coordinate, \(\dot{q}\) is the generalized velocity, \(\frac{\partial\mathcal{L}}{\partial\dot{q}}\) is the partial derivative of the Lagrangian with respect to the generalized velocity, and \(\frac{\partial\mathcal{L}}{\partial q}\) is the partial derivative of the Lagrangian with respect to the generalized coordinate. This loss function is used to ensure that the prediction results follow the conservation laws.

[0089] Physical constraint loss based on gauge field theory: This constraint is used to enhance the generalization ability of the model under different coordinate transformations, ensuring that the model remains stable under various physical transformations. By introducing terms related to gauge field theory into the loss function, the adaptability and robustness of the model can be effectively improved. This loss term ensures that the output of the model is not affected by coordinate transformations by introducing constraint terms.

[0090] In a possible implementation, the optimization of the loss function adopts a weighted strategy, where different physical constraints are assigned different weights according to their importance in a specific application to ensure a balanced impact of each constraint term on the final optimization result.

[0091] During the actual training process, the model will learn from the standardized material data, and the input features include the chemical composition, crystal structure of the material, and other factors related to physical properties. By continuously optimizing the loss function, the model gradually learns how to accurately predict the mechanical, thermal, and electromagnetic properties of the material.

[0092] The loss function in this embodiment not only enables the model to learn complex non - linear relationships but also ensures the reliability and physical consistency of the prediction results through the introduction of physical constraints. During the training process, the introduction of the loss function makes the model not only rely on the statistical characteristics of the data but also follow the basic physical laws such as mechanics and thermotics, thus enhancing the credibility of the model in real - world applications.

[0093] For step S4, in this embodiment, by introducing a variational inference optimization strategy, the generalization ability of the model in a low - data environment can be improved, and a regularization strategy is adopted to reduce overfitting. The main objective of this step is to improve the stability and robustness of the model to ensure that it can make accurate predictions when facing new data.

[0094] Generally, artificial intelligence models, especially deep - learning models, usually require a large amount of data for effective training. For the material property prediction task, experimental data may be limited, which requires the use of some advanced optimization techniques to enhance the generalization ability of the model. The variational inference method is one of the very effective strategies.

[0095] As an alternative, Variational Inference (VI) is a technique for optimizing the parameters of complex models by introducing prior knowledge and using approximate inference methods. In this embodiment, variational inference is applied to the parameter optimization of artificial intelligence models to enhance their predictive ability by inferring the latent parameter distribution of the models in the case of scarce data.

[0096] Specifically, the core idea of variational inference is to approximate the posterior distribution p(θ|x) of the model by introducing a variational distribution q(θ), thereby avoiding directly calculating the high-dimensional and complex posterior distribution. The optimization process can be achieved by minimizing the KL divergence of variational inference, and the definition of KL divergence is as follows:

[0097]

[0098] where D KL represents the KL divergence, q(θ) is the variational distribution, p(θ|x) is the posterior distribution, θ is the model parameter, x is the observed data, is the ratio of logarithms. By minimizing the KL divergence, variational inference minimizes the difference between the approximate distribution q(θ) and the true posterior distribution p(θ|x), thereby optimizing the model parameters.

[0099] In some embodiments, variational inference further enhances the stability of the model through Bayesian optimization methods. Bayesian optimization uses a probabilistic model to describe the uncertainty of the prediction task and effectively guides the model learning by selecting the input with the maximum amount of information. Bayesian optimization not only improves the prediction accuracy of the model but also provides reasonable inferences when the data is insufficient. Specifically, Bayesian optimization can be performed through the following steps:

[0100] Select an appropriate prior distribution: Select a suitable prior distribution according to the characteristics of the material properties. Usually, the Gaussian process or Markov chain Monte Carlo method is selected as the prior.

[0101] Design an optimization objective function: The optimization objective function is usually designed as a combination of the loss function of the model and the regularization term.

[0102] Iterative optimization: By continuously optimizing the objective function, gradually converge to the optimal parameters.

[0103] In a possible implementation, this step also uses regularization techniques to further reduce overfitting. Overfitting usually occurs when the amount of data is small. The model performs well on the training data but has poor generalization ability on new data. To solve this problem, this embodiment controls the complexity of the model by introducing L2 regularization (also known as Ridge regularization), thereby avoiding overfitting. The L2 regularization loss function is usually written as:

[0104]

[0105] where λ is the regularization parameter, and θ i is the i-th parameter of the model. By minimizing this regularized loss, the model parameters can be prevented from being too large, thus making the model more generalizable.

[0106] In addition, in some embodiments, the Dropout method may also be used to further reduce overfitting. Dropout is a commonly used regularization technique in deep learning. Its basic idea is to randomly discard a part of the neurons in the neural network to reduce the over-dependence on certain specific neurons. In this way, the training process of the model becomes more robust and can better adapt to unseen data.

[0107] By combining variational inference, Bayesian optimization, and regularization techniques, the material property prediction model of this embodiment can maintain high prediction accuracy and robustness in the case of scarce data. Variational inference optimizes the model parameters, enabling effective inference in a low-data environment, while the regularization method ensures the generalization ability of the model and avoids overfitting.

[0108] For step S5, in this embodiment, the prediction results are stored in a standardized data format, and a material database is constructed in combination with a knowledge graph for data sharing and reuse. The goal of this step is to convert the prediction results into sustainable data and ensure the compatibility and operability of different material data globally, thereby promoting the wide application of material property prediction technology in scientific research and industrial applications.

[0109] Generally, the prediction results generated after the optimization of the material property prediction model are highly data-dependent, and these results must be stored in a standardized format for subsequent query, analysis, and application. Standardized storage can not only improve the interoperability of data but also support multi-party data sharing across different fields. By adopting a standardized data storage format, the present invention ensures the long-term availability of the prediction results and provides valuable data resources for future material research.

[0110] As an option, the standardized data format used in this embodiment conforms to international material standards (such as CIF, XML, or JSON formats). These formats have been widely used in the field of materials science and can support data exchange and analysis between different software systems. For example, the CIF format is usually used to represent the crystal structure information of materials, while the JSON and XML formats are suitable for storing multi-dimensional data sets, such as the performance parameters of materials, experimental conditions, and related calculation data. These standardized formats help to avoid compatibility issues between different data sources, enabling researchers and engineers worldwide to easily share and utilize data.

[0111] Specifically, the stored prediction results include various performance parameters of materials, such as mechanical, thermal, and electromagnetic properties, etc. Taking mechanical properties as an example, the standardized storage format will include performance indicators such as the elastic modulus, yield strength, and hardness of the material, as well as data such as the calculated stress-strain curve. These data will be recorded in a unified format and can be accessed and used through corresponding query interfaces (such as APIs).

[0112] In some embodiments, in order to further improve the reusability and scalability of data, this embodiment also adopts knowledge graph technology to construct a materials database. A knowledge graph is a graph structure constructed by nodes (entities) and edges (relationships) and is used to express the complex relationships between material properties. By storing information such as the chemical composition, crystal structure, and physical properties of materials in a graphical manner, the knowledge graph can reveal the internal connections between materials and provide more reasoning and analysis functions.

[0113] Specifically, the nodes in the knowledge graph represent different materials or material characteristics, while the edges represent the relationships between these material characteristics. For example, one node may represent the chemical composition of a material, another node represents the mechanical properties of the material, and the edge between them can represent the association between the chemical composition and the mechanical properties. By constructing such a knowledge graph, users can not only easily search for and obtain the performance data of different materials but also discover the potential laws between different material characteristics.

[0114] In a possible implementation, the materials database combines blockchain technology for data storage. This method can ensure the security, integrity, and traceability of data. Through decentralization and encryption technologies, blockchain technology can effectively prevent data tampering and loss, ensuring the true reliability of data. In addition, blockchain can also achieve data transparency, enabling researchers worldwide to share and verify prediction results, and at the same time, it can also automate data access and use through smart contracts.

[0115] By storing the material property prediction results in a standardized format and integrating knowledge graph and blockchain technologies, this embodiment realizes the efficient management and sharing of material data. This not only improves the accessibility of data but also enhances the collaboration efficiency among different research institutions and the industrial sector during the material R & D process. With the continuous discovery of new materials and the accumulation of performance data, the material database will be continuously optimized and updated, providing solid data support for global material science and industrial applications.

[0116] In addition, the database of this embodiment can also be customized and extended according to different requirements. For example, in some applications, it may be necessary to deeply mine the properties of specific materials, while in other applications, it may be necessary to predict and optimize the properties of materials from a global perspective. The flexibility of the knowledge graph and blockchain makes it possible to meet these requirements, thus enhancing the flexibility and adaptability of the material data management system.

[0117] The artificial intelligence-based material property prediction and standardization system described below can be correspondingly referred to the artificial intelligence-based material property prediction and standardization method described above.

[0118] Please refer to the attached Figure 2 , the present invention also provides an artificial intelligence-based material property prediction and standardization system, including:

[0119] A data acquisition module for collecting data from laboratory experiments and computational materials science simulations. Its task is to ensure the comprehensiveness and accuracy of the data, covering multi-dimensional property data of different materials. After data acquisition, the data will be standardized, including noise removal, unit unification, and format conversion, to ensure that the data can be seamlessly input into subsequent modules for analysis and modeling. This module also supports the automatic update and maintenance of data to ensure that the material property database always remains up-to-date.

[0120] A physical modeling module constructs mathematical descriptions of mechanical, thermal, and electromagnetic properties based on the physical conservation laws of materials. By establishing accurate physical models, it ensures that the prediction of material properties can follow known physical laws, such as elastic modulus, thermal conductivity, magnetic permeability, etc. The physical modeling module also improves the accuracy of the model through a combination of symbolic operations and numerical solutions, thus providing a reliable physical basis for subsequent artificial intelligence model training. In addition, the module can handle complex multi-physical field coupling problems and simulate the performance of materials under different environments.

[0121] The artificial intelligence training module constructs a material property prediction model by combining deep neural networks and graph neural networks. Through the powerful expressive ability of neural networks, the model can identify complex non-linear relationships between material properties and perform efficient learning and prediction. During the training process, a physical constraint loss function is combined to ensure that the model always follows physical laws during learning. This constraint not only improves the scientific nature of the model but also avoids unreasonable predictions that may be caused by data-driven methods. Through this method, the artificial intelligence model can maintain physical consistency while ensuring the accuracy of data-driven methods.

[0122] The model optimization module optimizes the material property prediction model through variational inference methods, thereby improving the generalization ability and robustness of the model. Variational inference can efficiently handle uncertainty and avoid overfitting problems that may occur in traditional methods. In addition, this module automatically adjusts the model parameters, enabling the model to maintain a high prediction accuracy with limited sample data and adapt to different types of materials and performance prediction tasks. In this way, the optimization module ensures the adaptability and reliability of the prediction model for new material data.

[0123] The prediction and standardization module is responsible for storing the optimized prediction results and archiving them in a standardized data format (such as CIF, XML, JSON, etc.) for future query, analysis, and sharing. The material database constructed through knowledge graph technology can not only store and organize multi-dimensional material property data but also reveal the internal relationships between different material characteristics. This database provides a channel for data sharing and reuse among different research institutions, reducing duplicate labor and improving research efficiency. In addition, the module also uses blockchain technology to ensure the security, integrity, and traceability of data, ensuring the sharing and verification of prediction results globally.

[0124] The system of this embodiment can be used to execute the above method embodiment, and its principle and technical effects are similar, which will not be elaborated here.

[0125] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A material property prediction and standardization method based on artificial intelligence, characterized in that: The following steps are involved: Collect artificial intelligence material experimental data and computational materials simulation data, and standardize the data; Based on the physical conservation laws of artificial intelligence materials, a physical model including partial differential equation constraints is constructed to establish a mathematical description of the mechanical, thermal, and electromagnetic properties of materials; A deep neural network combined with a graph neural network is used to build an artificial intelligence material performance prediction model, and a physical constraint loss function is introduced during the training process; Optimize the prediction model of artificial intelligence material properties based on variational inference methods, improve the generalization ability of prediction in low-data environments, and use regularization strategies to reduce overfitting; Based on the optimized prediction model, the prediction results are stored in a standardized data format, and a material database is built in combination with the knowledge graph for data sharing and reuse.

2. The artificial intelligence-based material property prediction and standardization method according to claim 1, characterized in that: The standardization process comprises the following steps: Data unit conversion to ensure that material experimental data and computational simulation data use a unified international system of units; Format normalization, using standardized data storage format; Data cleaning, filling in missing data, removing abnormal data, and correcting measurement errors; Feature normalization, using maximum and minimum normalization or standardization for numerical variables.

3. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The physical model of the partial differential equation constraint includes: Mechanical properties model, which uses stress-strain equilibrium equation to describe the mechanical properties of artificial intelligence materials; Thermal performance model, which uses heat conduction equation to describe the temperature distribution of AI materials; The electromagnetic performance model uses Maxwell's equations to describe the electromagnetic properties of artificial intelligence materials.

4. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The physical constraint loss function includes: Physical constraint losses based on partial differential equations are used to ensure that the performance predictions of AI materials are consistent with the basic laws of physics; Physical constraint loss based on Noether's theorem to maintain energy conservation and momentum conservation of the system; Physical constraint loss based on gauge field theory is used to generalize the model under different coordinate transformations.

5. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The deep neural network adopts a multi-layer fully connected network structure to learn the nonlinear performance relationship of the material, and the graph neural network is used to learn the crystal structure characteristics of the material and capture the mapping relationship between the material structure and performance.

6. The artificial intelligence-based material property prediction and standardization method according to claim 1, characterized in that: The variational inference method adopts a variational Bayesian optimization strategy and introduces a divergence constraint to optimize the distribution of prediction model parameters so as to keep the prediction model stable in a low-data environment.

7. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The standardized data format adopts a data storage method that complies with international material standards and constructs associations between material properties based on a knowledge graph to improve the reusability and interoperability of data.

8. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The prediction model of artificial intelligence material performance is analyzed using explainable artificial intelligence technology, which includes: The SHAP method was used to analyze the key input features of the model prediction; The LIME method is used to provide local interpretability to enhance the transparency of prediction results.

9. The material property prediction and standardization method based on artificial intelligence according to claim 1, characterized in that: The material database combines blockchain technology for data storage to ensure data integrity, security and traceability, and supports dynamic data updates.

10. A material property prediction and standardization system based on artificial intelligence, applied to the material property prediction and standardization method based on artificial intelligence according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, used to collect experimental data of artificial intelligence materials and computational materials simulation data, and to standardize the data; A physical modeling module, used to build a physical constraint model based on the physical conservation laws of materials, wherein the physical conservation laws include mathematical descriptions of mechanical properties, thermal properties, and electromagnetic properties; An artificial intelligence training module, which uses a deep neural network combined with a graph neural network to build a material performance prediction model and introduces a physical constraint loss function during the training process; Model optimization module, which is used to optimize the material property prediction model based on variational inference method to improve the generalization ability and robustness of prediction; The prediction and standardization module is used to store the optimized prediction results and build a material database in combination with the knowledge graph for data sharing and reuse.

Images