Method and application of material attribute property relationship prediction model based on mutual information

Through mutual information-based methods to learn the relationship between material attributes and properties from scientific literature, establish a neural network model, solve the problems of data acquisition and model generalization in new material discovery, and achieve fast and accurate material attribute prediction and screening. Shorten the cycle of discovery of new functional materials.

CN114464275BActive Publication Date: 2024-11-22HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210179184.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-25
Publication Date
2024-11-22
Estimated Expiration
2042-02-25

AI Technical Summary

Technical Problem

The existing machine learning methods have problems in the discovery of new materials, such as unreliable data acquisition and poor generalization capabilities, and it is difficult to efficiently learn complex material attribute properties relationships from massive text data, and traditional methods are only applicable to specific data sets. and materials.

Method used

Using mutual information-based methods, text data is collected and processed from scientific literature in the field of materials science, constructed corpus, learned the relationship between material properties and properties through neural network models, established prediction models, and used online weight updates to improve training efficiency , realize rapid prediction and filtering.

Benefits of technology

It realizes the rapid and accurate prediction of the relationship between material properties and properties when occupies less computing resources, shortens the cycle of discovery of new functional materials, and improves the universality and reliability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114464275B_ABST
    Figure CN114464275B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and application for establishing a material attribute property relationship prediction model based on mutual information, which belongs to the field of material big data, including: collecting scientific literature in the field of material science, and performing text processing to obtain a descriptor set #imgabs0# and converting each descriptor into a vector, obtaining a corpus #imgabs1# descriptor as a vocabulary describing material attributes or properties; calculating the mutual information between every two descriptor vectors in #imgabs2# as a material attribute property relationship; establishing a neural network model, using the vectorized descriptor as input, for predicting the mutual information between the input descriptor vector and each vector in #imgabs3# as a material attribute property relationship; using #imgabs4# and the calculated mutual information to train the model, and obtaining a prediction model for predicting material attribute property relationships. The present invention can quickly predict the quantitative relationship between material attributes and properties while occupying less computing resources, and further complete the screening and prediction of known materials and unknown materials.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of material big data, and more specifically, relates to a method for establishing a material attribute property relationship prediction model based on mutual information and its application. Background Art

[0002] Functional materials, that is, materials with certain specific properties, are the basis of industrial innovation. Accelerating the discovery of new functional materials has always been a research hotspot in materials science. The traditional trial-and-error method of discovering new materials mainly relies on scientists' expertise in physics and chemistry, and is easily limited by experimental resources. With the development of material genome technology, high-throughput calculations based on density functional theory (DFT) have accelerated the search process to a certain extent. However, with the increase in the scale of practical problems and the complexity of materials, high-throughput calculations require expensive computing resources and time.

[0003] The emergence of machine learning technology has brought an efficient solution to this dilemma. Compared with first-principles calculations, machine learning technology relies on learning the potential representation of existing data sets instead of solving complex quantum mechanical equations. More importantly, machine learning uses several orders of magnitude fewer computing resources than first-principles calculations. Due to these advantages, machine learning technology has successfully discovered many new functional materials. However, there are still two major problems in the application of machine learning technology to new material discovery: First, data acquisition is the source of machine learning. Since machine learning needs to learn models from data, the reliability of data acquisition directly affects the reliability of the final model performance. In existing machine learning methods, data often comes from some computational databases, such as Materials Project, Jarvis, etc., but there are many structurally unstable materials in these computational databases and the calculation errors of some material properties are relatively large, so the reliability cannot be effectively guaranteed; second, in existing machine learning methods, machine learning models are mostly obtained based on training for specific tasks and specific data sets, and these models often perform poorly on other tasks.

[0004] The massive text data published in academic journals is a reliable data set that can be mined in the field of materials science. These unstructured text data sets contain researchers' specific explanations of the complex relationship terms between the structure, physical and chemical properties of all research fields and all functional materials in the field of materials science. However, traditional machine learning and statistical methods cannot directly learn the explanations of these complex relationship terms, and the existing methods for material prediction or screening based on scientific literature are only applicable to materials that have appeared in the literature. Therefore, how to efficiently learn the explanations of complex relationship terms from these massive and reliable text data sets and develop a general material prediction and screening method is of great significance and application value. Summary of the invention

[0005] In response to the defects of the prior art and the need for improvement, the present invention provides a method for establishing a material attribute property relationship prediction model based on mutual information and its application, the purpose of which is to quickly predict functional materials with required properties while occupying fewer computing resources, and to achieve the prediction of functional materials with different properties, thereby shortening the cycle of discovering new functional materials.

[0006] To achieve the above object, according to one aspect of the present invention, a method for establishing a material attribute property relationship prediction model based on mutual information is provided, comprising:

[0007] Collect scientific literature in the field of materials science and perform text processing to obtain a set of descriptors And the descriptor set Each descriptor in is converted into a vector to obtain the corpus Descriptors are words used to describe the properties or characteristics of a material;

[0008] Calculate the corpus separately The mutual information between every two descriptor vectors in is used as the corresponding material property relationship;

[0009] Build a neural network model; the neural network model takes the vectorized descriptor as input and is used to predict the input descriptor vector and the corpus The mutual information between the vectors in is used as the corresponding material property relationship;

[0010] Using Corpus The neural network model is trained with the calculated mutual information to obtain a prediction model for predicting the relationship between material properties.

[0011] The present invention collects scientific literature in the field of materials science to construct a data set, thereby ensuring the reliability of the data; mutual information is used to define the intrinsic relationship between material attributes and properties distributed in unstructured text, which can effectively measure the relationship between material attributes and properties. However, the corpus constructed based on scientific literature is too large, and in practical applications, it is impossible to calculate the relationship between material attributes and various properties by traversing. Based on this, the present invention further uses the established data set and the calculated mutual information to train a neural network model, and uses the neural network model to learn the complex relationships in massive text data sets. After the training is completed, the material attributes are input into the obtained calculation model, and the relationship between the material attributes and various properties can be extracted from the model output to complete the subsequent material prediction and material screening. This process occupies few resources, has high computational efficiency, does not rely on the prior knowledge of experts, and has high versatility.

[0012] Furthermore, text processing includes:

[0013] Sentence segmentation; word segmentation; remove punctuation marks; remove stop words that are not related to material science knowledge.

[0014] Furthermore, text processing also includes:

[0015] Standardize the chemical formula of materials; recognize phrases;

[0016] and, the vocabulary used to describe the material attributes or properties, including words, and identified phrases.

[0017] When the present invention performs text processing on scientific literature, in addition to identifying words that describe material attributes or properties, it will further identify phrases used to describe material attributes or properties for building a corpus, which can more deeply mine the complex relationships in massive text data and help improve the model's prediction accuracy for material attribute property relationships.

[0018] Furthermore, scientific literature in the field of materials science includes scientific literature related to the field of materials science published in academic journals.

[0019] Scientific literature related to the field of materials science published in academic journals has extremely high reliability. The scientific literature in the field of materials collected by the present invention includes these documents, which can further ensure the reliability of the model training data.

[0020] Furthermore, when training the neural network model, an online weight updating method is adopted.

[0021] Since the corpus constructed from massive text data is very large, the present invention adopts an online weight update method to train the neural network model, which can effectively improve the training efficiency of the model while ensuring the model training effect.

[0022] Furthermore, each word is converted into a vector by one-hot encoding.

[0023] According to another aspect of the present invention, there is provided a method for predicting a material attribute property relationship, comprising:

[0024] For a target material, obtain a vocabulary v for describing its material properties;

[0025] like The vocabulary v is converted into a corresponding vocabulary vector, and the vocabulary vector is input into the prediction model established by the material attribute property relationship prediction model establishment method based on mutual information provided by the present invention, and the material attribute property relationship between the material attribute of the target material and each property is extracted from the model output;

[0026] like Then split the vocabulary v into vocabulary set S, For each word in the vocabulary set S, it is converted into a corresponding vocabulary vector, and the vocabulary vector is input into the prediction model established by the material attribute property relationship prediction model establishment method based on mutual information provided by the present invention, and the material attribute property relationship corresponding to each word in the vocabulary set S is superimposed, that is, the mutual information is summed, as the material attribute property relationship between the material attribute of the target material and each property.

[0027] The prediction model established by the method for establishing a material attribute property relationship prediction model based on mutual information provided by the present invention can quickly and accurately predict the relationship between the input material attribute and various properties. The material attribute property relationship prediction method provided by the present invention is based on the model. For known materials, that is, materials whose material attributes have appeared in scientific literature, the vocabulary describing the material attribute is vectorized and then input into the model, so that the relationship between the material attribute and various properties can be quickly obtained; for unknown materials, that is, materials whose material attributes have not appeared in scientific literature, based on the calculation rules of mutual information, through vocabulary decomposition, the existing vocabulary in the corpus is used to calculate the relationship corresponding to each vocabulary obtained by decomposition, and then the relationship between the material attribute of the material and the target property is obtained by simple superposition. In general, the material attribute property prediction method provided by the present invention, whether it is a known material or an unknown material, can quickly predict the relationship between the material attribute and various properties while occupying fewer computing resources.

[0028] According to another aspect of the present invention, there is provided a material screening method comprising:

[0029] For multiple materials to be screened, the material attribute property relationship prediction method provided by the present invention is used to obtain the material attributes of each material and the material attribute property relationship between each property, and the material attribute property relationship between each material and the target property is extracted therefrom, and some materials with the strongest material attribute property relationship are screened out as materials with the target property.

[0030] The material screening method provided by the present invention can quickly calculate the relationship between material properties and target properties for both known materials and unknown materials while occupying fewer computing resources, and then quickly complete the material screening by simple sorting. Therefore, the material screening method provided by the present invention can quickly screen out functional materials with required properties while occupying fewer computing resources.

[0031] According to another aspect of the present invention, there is provided a material prediction method, comprising:

[0032] For the material to be predicted, the material attribute property relationship prediction method provided by the present invention is used to obtain the material attribute property relationship between its material attributes and various properties, and the properties corresponding to the material attribute property relationship that is higher than a preset second threshold are predicted as the properties that the material to be predicted may have.

[0033] The material prediction method provided by the present invention can quickly calculate the relationship between material properties and target properties for both known and unknown materials while occupying less computing resources, and then combine threshold judgment to predict the possible properties of the material. Therefore, the material prediction method provided by the present invention can predict the possible properties of unknown materials on the one hand, and predict the undiscovered properties of known materials on the other hand.

[0034] According to another aspect of the present invention, a computer-readable storage medium is provided, comprising a stored computer program, which, when executed by a processor, controls the device where the computer-readable storage medium is located to execute the material attribute property relationship prediction model establishment method based on mutual information provided by the present invention, and / or the material attribute property relationship prediction method provided by the present invention, and / or the material screening method provided by the present invention, and / or the material prediction method provided by the present invention.

[0035] In general, the above technical scheme conceived by the present invention collects a large amount of scientific literature related to the field of materials science that has been published in academic journals as a data set, uses mutual information to define the intrinsic relationship between the properties of materials distributed in unstructured text, and establishes a calculation model based on mutual information. The calculation model is implemented using a neural network method, and finally the prediction and screening of functional materials with different properties can be performed through the trained calculation model. This model can quickly predict functional materials with required properties with less computing resources, and can achieve the prediction of functional materials with different properties based on the same one, shortening the cycle of new functional material discovery, and providing guidance for scientific researchers in material selection during experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A method for establishing a material attribute property relationship prediction model based on mutual information provided by the present invention;

[0037] Figure 2 A schematic diagram of the structure of a neural network model for calculating mutual information provided by the present invention;

[0038] Figure 3A distribution diagram of mutual information scores between two-dimensional materials and two-dimensional photodetector properties in the Computational 2DMaterials Database (C2DB) database calculated using the material attribute property relationship prediction method provided by the present invention;

[0039] Figure 4 A distribution diagram of the mutual information scores of two-dimensional materials and two-dimensional magnetic properties in the Computational 2DMaterials Database (C2DB) database calculated by the material attribute property relationship prediction method provided by the present invention;

[0040] Figure 5 A trend chart showing the verification rates of the top 10, top 20, and top 30 materials screened out using the material screening method provided by the present invention and reported as two-dimensional photodetectors in papers from 2010 to 2020;

[0041] Figure 6 This is a material labeling diagram of the top 30 materials predicted by the material prediction method provided by the present invention and reported as two-dimensional magnetic materials in papers after 2017, wherein the materials marked with material chemical formulas represent two-dimensional materials experimentally reported as having two-dimensional magnetism after 2017. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0043] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects but not necessarily to describe a specific order or sequence.

[0044] In order to solve the technical problem that the prior art is unable to learn the interpretation of complex relationship terms from massive reliable text data sets, thereby completing material prediction and screening, the present invention provides a method for establishing a material attribute property relationship prediction model based on mutual information and its application, the overall idea is: using scientific literature in the field of materials science to construct training data for machine learning models, and using mutual information to define the intrinsic relationship of material attribute properties distributed in unstructured texts, establishing and training a neural network model, so that the model can learn the complex relationships in massive text data sets during the training process, and finally obtain a computational model that can quickly predict the relationship between the attributes and properties of materials, based on which material screening and material prediction can be quickly completed while occupying fewer computing resources.

[0045] Before explaining the technical solution of the present invention in detail, the relevant technical terms are briefly introduced as follows:

[0046] Material properties: description of the material structure;

[0047] Material properties: description of the material's functionality;

[0048] Descriptor: A word used to describe a material attribute or material property;

[0049] Functional materials: materials with certain specific properties;

[0050] Material screening: For known materials, screen out materials with specific properties;

[0051] Material prediction: predicting the possible properties of a material, including two aspects: first, for unknown materials, predicting the possible properties of the material; second, for known materials, predicting the possible properties of the material that have not yet been discovered;

[0052] Mutual Information: Mutual Information is a useful information measure in information theory. It can be regarded as the amount of information contained in a random variable about another random variable, or the uncertainty of a random variable reduced by knowing another random variable. In the present invention, mutual information is used to define the intrinsic relationship between material properties distributed in unstructured text. The calculation formula of mutual information is as follows:

[0053]

[0054] In the formula, and Represents two sets, Representing a collection and collection The mutual information between and Respectively represent sets and collection Elements in and Represented and The joint probability between The product of their respective marginal probabilities.

[0055] The following are examples.

[0056] Embodiment 1:

[0057] A method for establishing a material attribute property relationship prediction model based on mutual information, such as Figure 1 As shown, including:

[0058] Collect scientific literature in the field of materials science and perform text processing to obtain a set of descriptors And the descriptor set Each descriptor in is converted into a vector to obtain the corpus Descriptors are words used to describe the properties or characteristics of a material;

[0059] Calculate the corpus separately The mutual information between every two descriptor vectors in is used as the corresponding material property relationship;

[0060] Build a neural network model; the neural network model takes the vectorized descriptor as input and is used to predict the input descriptor vector and the corpus The mutual information between the vectors in is used as the corresponding material property relationship;

[0061] Using Corpus The neural network model is trained with the calculated mutual information to obtain a prediction model for predicting the relationship between material properties.

[0062] This embodiment constructs a data set through scientific literature in the field of materials science, which ensures the reliability of the data and avoids dependence on expert prior knowledge; as an optional implementation method, the scientific literature collected in this embodiment is specifically approximately 4.62 million English scientific literature related to the field of materials science between 1920 and 2020, and the relevant literature is collected through Elsevier's Scopus standard application program interface; it should be noted that scientific literature published in other languages ​​can also be applied to the present invention.

[0063] After collecting relevant scientific literature, the specific method of text preprocessing in this embodiment is as follows: first, the nltk software package is used to sentence and word segment the collected text data, and some punctuation marks and stop words with high frequency that are irrelevant to the expression of material science knowledge are removed. Finally, the massive text data will be processed into a series of words and descriptors used to describe the attributes or properties of the material; considering that the chemical formulas of the same material may be inconsistent in different documents, this embodiment further uses the pymatgen software package to standardize the material chemical formulas in the descriptors; in addition, considering that words may be combined into phrases to describe the attributes or properties of the material, this embodiment further uses the gensim software package to identify the phrases in the descriptors and connect them through special symbols for subsequent use; optionally, this embodiment uses the special symbol "-" to connect these words with potential phrase structures representing the attributes of professional knowledge in the material field, such as "two-dimensional-materials" and "solar-cells".

[0064] After the above text processing, the massive unstructured text data collected in this embodiment is processed into words used to describe the properties or characteristics of the material. These words are descriptors, including words and phrases. These descriptors constitute a descriptor set. To simplify the processing, as an optional implementation, in this embodiment, the descriptor set When converting each descriptor in to a vector, one-hot encoding is used. Specifically, for each descriptor, a position is assigned to it, and a vector of length N (N is the total number of descriptors) is used to represent a single descriptor. In this vector, the element at the position corresponding to the descriptor is 1, and the other elements are 0; for example, the vectors obtained by converting the three descriptors "2D", "MoS2" and "photodetectors" are as follows Figure 2 As shown; after vectorization, these sparse vectors form a matrix And the corpus

[0065] Based on the definition of the relationship between the attributes and properties of materials in the present invention, after obtaining the corpus Afterwards, this embodiment will calculate the mutual information based on the above formula for the corpus The mutual information is calculated for each two vectors in as the corresponding material attribute property relationship; it is easy to understand that the larger the mutual information value, the stronger the relationship between the material attributes and properties.

[0066] In order to improve the training efficiency as much as possible while ensuring the model calculation effect, as an optional implementation method, Figure 2As shown, the neural network model established in this embodiment is a three-layer structure, including an input layer, a hidden layer and an output layer; the number of nodes in the input layer is N, which is used to input a descriptor vector of length N; the number of nodes in the output layer is N, and each node corresponds to N descriptors one by one, and the output information is the input descriptor vector and the corpus The mutual information between each vector in is used as the corresponding material property relationship; Represents the established neural network model. During training, each descriptor vector is input into the neural network model. In the last layer, the softmax activation function is used to output the relative probability values ​​of the material and target attributes. The relevant calculation can be expressed as:

[0067]

[0068] In the formula, m represents a material, c represents a target property, represents the property set, MI(m,c) represents the mutual information value of the material and target properties, Represents the inner product value of the input layer parameters and output layer parameters of the neural network; it should be noted that, Figure 2 The neural network model structure shown is only an optional implementation and should not be understood as the only limitation of the present invention. Other neural network models that can accurately complete mutual information calculation after training can also be used in the present invention.

[0069] Because the corpus In this embodiment, the neural network is trained by online weight updating. After the training, a prediction model for predicting the relationship between material properties can be obtained. The trained neural network model and the input layer parameters and output layer parameters of the neural network are saved. In subsequent applications, the relevant parameters are loaded and the prediction model can be used to calculate the mutual information.

[0070] It is easy to understand that as scientific literature in the field of materials science continues to be published, the model can also be continuously updated and trained. The specific training methods are similar and will not be repeated here.

[0071] Embodiment 2:

[0072] A material property relationship prediction method, comprising:

[0073] For a target material, obtain a vocabulary v for describing its material properties;

[0074] like That is, if the target material is a known material, the vocabulary v is converted into a corresponding vocabulary vector, and the vocabulary vector is input into the prediction model established by the material attribute property relationship prediction model establishment method based on mutual information provided in the above embodiment 1, and the material attribute property relationship between the material attribute of the target material and each property is extracted from the model output;

[0075] like That is, if the target material is an unknown material, the vocabulary v is split into a vocabulary set S. For each word in the vocabulary set S, convert it into a corresponding vocabulary vector, and input the vocabulary vector into the prediction model established by the material attribute property relationship prediction model establishment method based on mutual information provided in the above-mentioned embodiment 1, and superimpose the material attribute property relationship corresponding to each word in the vocabulary set S as the material attribute property relationship between the material attribute of the target material and each property.

[0076] The prediction model established in the above-mentioned embodiment 1 can quickly and accurately predict the relationship between the input material attributes and various properties. The material attribute property relationship prediction method provided in this embodiment is based on this model. For known materials, that is, materials whose material attributes have appeared in scientific literature, the vocabulary describing the material attributes is vectorized and input into the model, and the relationship between the material attributes and various properties can be quickly obtained; for unknown materials, that is, materials whose material attributes have not appeared in scientific literature, based on the calculation rules of mutual information, through vocabulary decomposition, the existing vocabulary in the corpus is used to calculate the relationship between the decomposed vocabulary, and then the relationship between the material attributes of the material and the target properties is obtained by simple superposition. In general, the material attribute property prediction method provided in this embodiment, whether it is a known material or an unknown material, can quickly predict the relationship between the material attributes and various properties while occupying less computing resources.

[0077] Embodiment 3:

[0078] A material screening method, comprising:

[0079] For multiple materials to be screened, the material attribute property relationship prediction method provided in this embodiment is used to obtain the material attributes of each material and the material attribute property relationship between each property, and the material attribute property relationship between each material and the target property is extracted therefrom, and some materials with the strongest material attribute property relationship are screened out as materials with the target property.

[0080] The material screening method provided in this embodiment can quickly calculate the relationship between material properties and target properties for both known and unknown materials while occupying fewer computing resources, and then quickly complete material screening through simple sorting. Therefore, the material screening method provided in this embodiment can quickly screen out functional materials with required properties while occupying fewer computing resources.

[0081] Embodiment 4:

[0082] A material prediction method, comprising:

[0083] For the material to be predicted, the material attribute property relationship prediction method provided in this embodiment is used to obtain the material attribute property relationship between its material attributes and various properties, and the properties corresponding to the material attribute property relationship that is higher than a preset second threshold are predicted as the properties that the material to be predicted may have.

[0084] The material prediction method provided in this embodiment can quickly calculate the relationship between material properties and target properties for both known and unknown materials while occupying less computing resources, and then combine the threshold judgment to predict the possible properties of the material. Therefore, the material prediction method provided in this embodiment can predict the possible properties of unknown materials on the one hand, and predict the undiscovered properties of known materials on the other hand.

[0085] Embodiment 5:

[0086] A computer-readable storage medium includes a stored computer program. When the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the material attribute property relationship prediction model establishment method based on mutual information provided in the above-mentioned embodiment 1, and / or the material attribute property relationship prediction method provided in the above-mentioned embodiment 2, and / or the material screening method provided in the above-mentioned embodiment 3, and / or the material prediction method provided in the above-mentioned embodiment 4.

[0087] The effects that can be achieved by the present invention are further described below in conjunction with some specific application examples.

[0088] Figure 3 The mutual information score distribution diagram of the two-dimensional materials in the Computational 2D Materials Database (C2DB) database and the two-dimensional photoelectric properties is calculated using the material screening method provided in the above embodiment 3, and the five materials with the highest mutual information scores are screened out. These five materials are distributed in Figure 3MoS2, MoSe2, MoTe2, WSe2, and WS2 in the upper right corner. These five materials are typical two-dimensional photodetector materials that have been experimentally reported.

[0089] Figure 4 The two-dimensional magnetic property is taken as the target property, and the mutual information score distribution diagram of the two-dimensional materials in the Computational 2D Materials Database (C2DB) database and the two-dimensional magnetic property is calculated by the material screening method provided in the above embodiment 3, and the five materials with the highest mutual information scores are screened out. These five materials are distributed in Figure 4 CrI3, Fe3GeTe2, MnBi2Te4, CrSiTe3, Cr2Ge2Te6 in the upper right corner are distributed Figure 4 The upper right corner of the figure shows the five materials that are typical two-dimensional magnetic materials that have been experimentally confirmed.

[0090] according to Figure 3 and Figure 4 It can be seen from the results shown that the material screening method provided in the above embodiment can accurately screen out materials with target properties.

[0091] Figure 5 This is a trend chart of the verification rates of the top 10, 20, and 30 materials with the mutual information scores obtained by the material prediction method provided in Example 4 and reported as two-dimensional photodetectors in papers from 2010 to 2020. It can be seen that the accuracy rate of the predicted materials being reported as two-dimensional photodetector materials in the past 10 years exceeds 60%.

[0092] Figure 6 This is a material marker diagram of 30 materials predicted by the material prediction method provided in Example 4 and reported as two-dimensional magnetic materials in papers after 2017; since it was not until 2017 that the experimental demonstration of magnetic order in Cr2Ge2Te6 and CrI3 solved the long-standing question about whether there is magnetic order in two-dimensional materials for the first time, the literature before 2016 was selected for training prediction during verification. It can be seen from the figure that the accuracy rate of the predicted materials reported to have two-dimensional magnetism in experiments in the past three years exceeds 50%.

[0093] The above results show that the present invention can accurately screen out functional materials with target properties and can also accurately predict unknown properties of materials.

[0094] It will be easily understood by those skilled in the art that 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 and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for establishing a material attribute property relationship prediction model based on mutual information, characterized in that: include: Collect scientific literature in the field of materials science and perform text processing to obtain a set of descriptors And the descriptor set Each descriptor in is converted into a vector to obtain the corpus The descriptor set Descriptors in are words used to describe material attributes or properties; Calculate the corpus separately The mutual information between every two descriptor vectors in is used as the corresponding material property relationship; Establish a neural network model; the neural network model takes the vectorized descriptor as input and is used to predict the input descriptor vector and the corpus The mutual information between the vectors in is used as the corresponding material property relationship; Using the corpus The neural network model is trained based on the calculated mutual information to obtain a prediction model for predicting the relationship between material attributes.

2. The method for establishing a material attribute property relationship prediction model based on mutual information according to claim 1, characterized in that: The text processing includes: Sentence segmentation; word segmentation; remove punctuation marks; remove stop words that are not related to material science knowledge.

3. The method for establishing a material attribute property relationship prediction model based on mutual information according to claim 2, characterized in that: The text processing also includes: Standardize the chemical formula of materials; recognize phrases; Also, the vocabulary used to describe material attributes or properties includes words and recognized phrases.

4. The method for establishing a material attribute property relationship prediction model based on mutual information according to any one of claims 1 to 3, characterized in that: The scientific literature in the field of materials science includes scientific literature related to the field of materials science published in academic journals.

5. The method for establishing a material attribute property relationship prediction model based on mutual information according to any one of claims 1 to 3, characterized in that: When training the neural network model, an online weight updating method is adopted.

6. The method for establishing a material attribute property relationship prediction model based on mutual information according to any one of claims 1 to 3, characterized in that: The way to convert each descriptor into a vector is one-hot encoding.

7. A method for predicting the relationship between material attributes, characterized in that: include: For a target material, obtain a vocabulary v for describing its material properties; like The vocabulary v is converted into a corresponding vocabulary vector, and the vocabulary vector is input into the prediction model established by the method for establishing a material attribute property relationship prediction model based on mutual information according to any one of claims 1 to 6, and the material attribute property relationship between the material attribute of the target material and each property is extracted from the model output; like Then the vocabulary v is split into a vocabulary set S, For each word in the vocabulary set S, convert it into a corresponding vocabulary vector, and input the vocabulary vector into the prediction model established by the method for establishing a material attribute property relationship prediction model based on mutual information as described in any one of claims 1 to 6, and superimpose the material attribute property relationships corresponding to each word in the vocabulary set S as the material attribute property relationships between the material attributes of the target material and each property.

8. A material screening method, characterized in that: include: For multiple materials to be screened, the material property property relationship prediction method described in claim 7 is used to obtain the material properties of each material and the material property property relationship between each property, and the material property property relationship between each material and the target property is extracted therefrom, and some materials with the strongest material property property relationship are screened out as materials with the target property.

9. A material prediction method, characterized in that: include: For the material to be predicted, the material attribute property relationship prediction method described in claim 7 is used to obtain the material attribute property relationship between its material attributes and various properties, and the properties corresponding to the material attribute property relationship that is higher than a preset second threshold are predicted as the properties that the material to be predicted may have.

10. A computer-readable storage medium, characterized in that: It includes a stored computer program, which, when executed by a processor, controls the device where the computer-readable storage medium is located to execute the material attribute property relationship prediction model establishment method based on mutual information as described in any one of claims 1 to 6, and / or the material attribute property relationship prediction method as described in claim 7, and / or the material screening method as described in claim 8, and / or the material prediction method as described in claim 9.