Learning device, material search system, and learning method

Abstract

To provide a learning device capable of obtaining an appropriate learning model, a material search system including the learning device, and a learning method.SOLUTION: A learning device 2 comprises: a first feature amount calculation section 14; a second feature amount calculation section 18; and a similarity evaluation section 20. The first feature amount calculation section 14 obtains a first feature amount by using a first learning model for outputting the amount of feature of crystal structure of a predetermined material corresponding to graph data as a first feature amount with the graph data showing the crystal structure of the predetermined material as input. The second feature amount calculation section 18 obtains a second feature amount by using a second learning model for outputting crystal structure of a predetermined material corresponding to test data as a second feature amount with the test data showing the crystal structure of the predetermined material as the test data. The similarity evaluation section 20 updates the parameter of each learning model such that the similarity of a combination of respective feature amounts corresponding to the crystal structure of the same material becomes high and the similarity of a combination of respective feature amounts corresponding to the crystal structure of a different material becomes low.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a learning device, a material exploration system including the learning device, and a learning method. [Background technology]

[0002] BACKGROUND ART Conventionally, there is known an apparatus for analyzing the crystal structure of a material by evaluating the similarity between graph data that indicates the bonding relationships between atoms using nodes and edges (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-124964 Summary of the Invention [Problem to be solved by the invention]

[0004] The present inventors have been studying learning models for searching for the crystal structures of materials with desired physical properties. Their studies have revealed that learning models based on single-form information such as graph data have limited feature representation, making it difficult to obtain an appropriate learning model.

[0005] The present disclosure aims to provide a learning device capable of obtaining an appropriate learning model, a material search system including the learning device, and a learning method. [Means for solving the problem]

[0006] The invention described in claim 1 is A learning device, comprising: a first feature calculation unit (14) that receives graph data indicating a crystal structure of a predetermined material as input and calculates a first feature using a first learning model that outputs a feature of the crystal structure of the predetermined material corresponding to the graph data as a first feature; a second feature calculation unit (18) that receives text data indicating a crystal structure of a predetermined material as input and calculates the second feature using a second learning model that outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; The system is provided with a similarity evaluation unit (20) that updates both the parameters of the first learning model and the parameters of the second learning model so that the similarity between combinations of the first feature amount and the second feature amount corresponding to the crystal structure of the same material is high and the similarity between combinations of the first feature amount and the second feature amount corresponding to the crystal structure of different materials is low.

[0007] In this way, by learning not only the feature quantities of graph data showing crystal structures but also the feature quantities of graph data showing crystal structures and text data in combination, it becomes possible to learn material features that cannot be learned from a single information format. If feature quantities are learned from information based on multiple formats, as in the present disclosure, an appropriate learning model based on a variety of feature expressions can be obtained.

[0008] The invention described in claim 10 is A materials exploration system, comprising: A learning device (2) according to claim 1 or 2; a search device (4) that uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit of the learning device, as a learned learning model for material search, and searches for a crystal structure of a predetermined material having desired physical properties using the learning model; The system is equipped with a search device (4) that searches for a crystal structure of a predetermined material having desired physical properties using a learning model for material search that has been learned by the learning device.

[0009] In this way, by using an appropriate learning model based on various feature representations, it becomes possible to appropriately search for crystal structures with desired physical properties.

[0010] The invention described in claim 11 is A learning method using a computer (2), obtaining a first feature using a first learning model that receives graph data indicating a crystal structure of a predetermined material as input and outputs a feature of the crystal structure of the predetermined material corresponding to the graph data as a first feature; obtaining a second feature using a second learning model that receives text data indicating a crystal structure of a predetermined material as input and outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; The method includes updating both the parameters of the first learning model and the parameters of the second learning model so that the similarity between combinations of the first feature and the second feature corresponding to the crystal structure of the same material is high and the similarity between combinations of the first feature and the second feature corresponding to the crystal structure of different materials is low.

[0011] In this way, if features are learned using information based on multiple forms, an appropriate learning model based on diverse feature expressions can be obtained.

[0012] The reference symbols in parentheses attached to each component indicate an example of the correspondence between the component and the specific components described in the embodiments described below. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a schematic configuration diagram of a material exploration system according to a first embodiment. [Figure 2] 1 is a schematic configuration diagram of a learning device according to a first embodiment. [Figure 3] FIG. 10 is an explanatory diagram for explaining extraction of a first feature amount from graph data. [Figure 4] FIG. 10 is an explanatory diagram for explaining abstraction of specific written items included in text data. [Figure 5] FIG. 10 is an explanatory diagram for explaining extraction of a second feature amount from text data. [Figure 6] FIG. 10 is an explanatory diagram for explaining a loss function used in similarity evaluation of each feature amount. [Figure 7]FIG. 2 is an explanatory diagram for explaining the operation of the learning device according to the first embodiment. [Figure 8] FIG. 10 is an explanatory diagram illustrating a first modified example of a loss function. [Figure 9] FIG. 10 is an explanatory diagram illustrating a second modified example of the loss function. [Figure 10] FIG. 10 is a schematic configuration diagram of a learning device according to a second embodiment. [Figure 11] FIG. 10 is a schematic configuration diagram of a learning device according to a third embodiment. [Figure 12] FIG. 10 is a schematic configuration diagram of a learning device according to a fourth embodiment. [Figure 13] FIG. 10 is an explanatory diagram for explaining the operation of a learning device according to a fourth embodiment. [Figure 14] FIG. 10 is a schematic configuration diagram of a material exploration system according to a fifth embodiment. [Figure 15] FIG. 13 is an explanatory diagram for explaining the operation of a searcher according to the fifth embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the following embodiments, parts that are the same as or equivalent to those described in the preceding embodiments will be given the same reference numerals, and their description may be omitted. Furthermore, in the embodiments, when only some of the components are described, the components described in the preceding embodiments can be applied to the remaining components. The following embodiments can be partially combined with each other, even if not specifically stated, as long as there is no particular problem with the combination.

[0015] (First embodiment) This embodiment will be described with reference to Figures 1 to 7. In this embodiment, the crystal structure of an inorganic material is set as the learning target of a learning device 2, and an example will be described in which a material exploration system 1 including the learning device 2 is used to search for a crystal structure of an inorganic material having desired physical properties.

[0016] As shown in Figure 1, the materials exploration system 1 includes a learning device 2, a storage device 3 that stores a learning model for materials exploration that has been learned by the learning device 2, and an exploration device 4 that searches for a crystal structure with desired physical properties using the learning model for materials exploration.

[0017] The learning device 2 is a device for learning a first learning model for graph data including a graph encoder GEN and a second learning model for text data including a text encoder TEN, using learning data input from an external source.

[0018] The learning device 2 is configured by a computer including one or more processors and memory. The processor may be a general-purpose processor such as a CPU or GPU, or a dedicated processor specialized for learning. CPU is an abbreviation for Central Processing Unit. GPU is an abbreviation for Graphics Processing Unit. The memory is configured by a non-transitory tangible storage medium such as semiconductor memory, magnetic memory, or optical memory. The learning device 2 may be configured by at least one of a user terminal, a server, and a general-purpose or dedicated electronic device, or a combination of these.

[0019] As shown in FIG. 2, the learning device 2 is configured to include a graph data creation unit 12, a first feature calculation unit 14, a text data creation unit 16, a second feature calculation unit 18, and a similarity evaluation unit 20 as functional units for realizing various functions.

[0020] A common data format for crystallography called a CIF file is input to the learning device 2 as learning data. The CIF file contains information such as atomic coordinates, cell parameters, symmetry, and atomic species of the crystal structure. The CIF file is stored, for example, in an external database and provided to the learning device 2 from that database. CIF is an abbreviation for Crystallographic Information File.

[0021] The graph data creation unit 12 creates graph data showing the crystalline structures of inorganic materials. The graph data is data in which atoms constituting the inorganic material are represented by nodes and bonds between the atoms are represented by edges. The graph data creation unit 12 creates graph data showing the crystalline structures of a large number of inorganic materials based on information such as atomic coordinates, cell parameters, symmetry, and atomic species of the crystalline structures of the inorganic materials contained in the CIF.

[0022] The first feature calculation unit 14 receives the graph data created by the graph data creation unit 12 as input and calculates the first feature using a first learning model that outputs, as the first feature, a feature of the crystalline structure of the inorganic material corresponding to the graph data. For example, as shown in FIG. 3, the first feature calculation unit 14 inputs a batch of graph data to a graph encoder GEN, and embeds the first feature corresponding to each graph data from the graph encoder GEN into a fixed-length vector and outputs it. For example, when the batch size is B, the first feature calculation unit 14 outputs B feature vectors.

[0023] The first feature amount calculation unit 14 is configured by an embedded neural network that vectorizes graph data as a first feature amount. The first feature amount calculation unit 14 can be configured by, for example, CGCNN, ALIGNN, or the like. CGCNN is an abbreviation for Crystal Graph Convolutional Neural Network. Also, ALIGNN is an abbreviation for Atomistic Line Graph Neural Network. Note that the first feature amount calculation unit 14 may be configured by a Transformer.

[0024] The text data creation unit 16 creates text data indicating the crystalline structure of an inorganic material. The text data is data that expresses the crystalline structure of an inorganic material, etc., in natural language. The text data includes information related to the crystalline structure of the inorganic material, as well as information on the physical properties and characteristics of the inorganic material. The text data creation unit 16 of this embodiment includes a physical property calculation unit 161, a text conversion unit 162, and an abstraction processing unit 163.

[0025] The physical property calculation unit 161 calculates the physical properties of the inorganic material of interest to the user based on the CIF file and graph data. The physical property calculation unit 161 performs calculations to identify, for example, the space group, crystal structure, and arrangement structure of the inorganic material. Note that the physical property calculation unit 161 may be configured to determine various properties, such as the electrical properties and thermal properties, of the inorganic material through first-principles calculations, calculations equivalent to first-principles calculations, automated experiments, databases, etc.

[0026] The text conversion unit 162 converts information such as numerical values ​​and structures related to the characteristics obtained by the physical property calculation unit 161 into text written in a predetermined natural language and outputs the text. For example, the text conversion unit 162 converts the information obtained by the physical property calculation unit 161 into text written in English and outputs the text. It is desirable that the text conversion unit 162 is configured to include data describing physical properties related to a predetermined material that cannot be expressed in graph data in the text data. This is because it is expected that a wider variety of expressions can be learned by including physical properties related to a crystal structure that cannot be expressed in graph data showing the crystal structure in the text data. Note that the text conversion unit 162 may also be configured to output text written in a language other than English.

[0027] The physical property calculation unit 161 and the text conversion unit 162 can be realized by using, for example, a tool called Robocrystallographer, which automatically generates text that explains crystal structures.

[0028] Unlike organic materials, inorganic materials often represent substances made up of a combination of various elements, and there are many rules governing the atomic coordinates, cell parameters, occupancy rates, etc. For this reason, if specific descriptions in text data about inorganic materials are linked to graph data, there is a risk that information will leak, making it impossible to learn properly.

[0029] Taking this into consideration, the text data creation unit 16 of this embodiment includes an abstraction processing unit 163 that replaces specific descriptions included in the text data with abstract words based on predetermined rules. For example, the abstraction processing unit 163 replaces specific descriptions included in the text data with words that are higher concepts.

[0030] The abstraction processing unit 163 of this embodiment abstracts the genera and periods shown in the periodic table into a single expression, and also abstracts elements that exhibit specific physical properties into a single expression. For example, as shown in FIG. 4, the abstraction processing unit 163 abstracts the element names of inorganic materials included in text data into [METAL] and [NONMETAL]. Furthermore, for example, a composition formula such as FeZn(SiO3)2 is abstracted into "This material." This prevents information leaks between the crystal graph and the text.

[0031] Here, the abstraction processing unit 163 is not limited to abstracting element names, but may also abstract, for example, composition formulas, bond lengths / angles, etc. Furthermore, the abstraction processing unit 163 may also abstract the physical property values ​​of inorganic materials by, for example, defining a threshold value for electrical conductivity σ and labeling values ​​equal to or greater than the threshold as "high conductivity."

[0032] The second feature calculation unit 18 receives the text data created by the text data creation unit 16 as input and calculates the second feature using a second learning model that outputs, as the second feature, a feature of the crystalline structure of the inorganic material corresponding to the text data. For example, as shown in FIG. 5, the second feature calculation unit 18 inputs a batch of multiple text data to a text encoder TEN, and embeds the second feature corresponding to each piece of text data from the text encoder TEN into a fixed-length vector and outputs it. For example, when the batch size is B, the second feature calculation unit 18 outputs B feature vectors.

[0033] The second feature calculation unit 18 is configured by an embedded neural network that vectorizes text data as second features. The second feature calculation unit 18 is configured by, for example, BERT, MstSciBERT, or the like. BERT is a model that uses a Transformer encoder. BERT can also add tokens extracted from text data by named entity extraction. BERT is an abbreviation for Bidirectional Encoder Representations from Transformer.

[0034] The similarity evaluation unit 20 updates the parameters of both the first learning model and the second learning model by comparative learning between the output of the first learning model, which is the learning model of the first feature calculation unit 14, and the output of the second learning model, which is the learning model of the second feature calculation unit 18.

[0035] 6, the similarity evaluation unit 20 generates a B×B matrix by taking the inner product of a vector G corresponding to the first feature amount and a vector T corresponding to the second feature amount, and trains each learning model so that the diagonal components of the matrix have larger values ​​than the off-diagonal components. For example, the similarity evaluation unit 20 trains each learning model so that the similarity of the diagonal components in the matrix generated by taking the inner product of vector G and vector T approaches "1" and the similarity of the off-diagonal components approaches "0". As functional units for realizing such training, the similarity evaluation unit 20 has a similarity cost calculation unit 201 and a parameter update unit 202.

[0036] The similarity cost calculation unit 201 treats the first feature amount and the second feature amount as paired data and calculates the similarity cost L of the paired data using a predetermined loss function. The similarity cost calculation unit 201 of this embodiment calculates the similarity cost L using, as the loss function, a contrastive loss including a term combining the first feature amount and the second feature amount, as shown in Formula F1 in Fig. 6 .

[0037] The parameter update unit 202 updates the parameters of each learning model so that the similarity between combinations of the first feature amount and the second feature amount corresponding to the crystal structure of the same material is high and the similarity between combinations of the first feature amount and the second feature amount corresponding to the crystal structure of different materials is low. The parameter update unit 202 updates the parameters of each learning model by, for example, backpropagation.

[0038] The storage device 3 stores data used in the operation of the learning device 2 and the operation of the searching device 4, and data obtained by the operation of the learning device 2 and the operation of the searching device 4. For example, the storage device 3 stores a first learning model and a second learning model that have been learned by the learning device 2 as learning models for material searching. The storage device 3 is configured from a non-transitive tangible storage medium.

[0039] The searching device 4 searches for a crystal structure of an inorganic material having desired physical properties using the first learning model and the second learning model trained by the learning device 2 as learning models for material search. The searching device 4, for example, inputs text data describing the desired physical properties into the trained text encoder TEN to obtain second feature quantities corresponding to the desired physical properties. The searching device 4 then selects first feature quantities output from the trained graph encoder GEN that have a high similarity to the second feature quantities, and obtains a crystal structure of an inorganic material having desired physical properties based on the selected second feature quantities.

[0040] The operation of the learning device 2 configured in this manner and the material search system 1 including the learning device 2 will be described with reference to Figure 7. What is shown in Figure 7 is an example of the operation of the learning device 2 and the material search system 1, and the operation is not limited to this example.

[0041] As shown on the left side of Figure 7, the learning device 2 first uses externally input learning data to create graph data and text data showing the crystalline structure of an inorganic material. Then, through contrastive learning, the learning device 2 updates the parameters of the graph encoder GEN that constitutes the learning model for the graph data and the text encoder TEN that constitutes the learning model for the text data. This allows learning of a shared representation that associates the graph data with the text data.

[0042] Next, as shown on the right side of Figure 7, the materials exploration system 1 searches for a crystalline structure of an inorganic material having desired physical properties through a common feature space in different modalities. For example, the materials exploration system 1 can obtain a crystalline structure of an inorganic material corresponding to the desired physical properties using text data describing the desired physical properties. Furthermore, the materials exploration system 1 can obtain text explaining the physical properties of a crystalline structure from graph data of the crystalline structure predicted to have the desired physical properties.

[0043] Conventionally, methods such as physical property regression have been used to obtain feature quantities by aggregating adjacent information such as graph convolution between the nodes and edges of the crystal structure of a given material. Text data mining has also been used to visualize the relationships between knowledge. However, both of these methods rely on single-form information and its specific data structure, particularly training information organized in tables.

[0044] In contrast, the learning device 2 of this embodiment performs contrastive learning between graph information and text information for the crystal structure of a given material, learning a feature space common to both as a shared concept. Specifically, the learning device 2 acquires a single feature space through contrastive learning between feature vectors obtained by aggregating the nodes and edges of the graphed crystal structure and feature vectors in which text describing facts associated with the crystal structure is embedded without input shaping such as tabulation. This allows for a common representation between the chemical structure and the text as a feature of the crystal structure that cannot be learned from single-form information alone. As a result, an appropriate learning model based on diverse feature representations can be obtained.

[0045] Furthermore, the learning device 2 and material search system 1 of this embodiment have the following features.

[0046] (1) Inorganic materials are often expressed as a combination of various elements, and if specific descriptions in text data are linked to graph data, there is a risk that information will leak and appropriate learning will not be possible.

[0047] In contrast, the learning device 2 of this embodiment includes an abstraction processing unit 163 that replaces specific descriptions contained in text data with abstract words based on predetermined rules. Replacing specific descriptions contained in text data with abstract words in this way reduces information leakage and enables appropriate learning. Reducing information leakage also serves to abstract crystals, each of which is a unique substance, at a granularity specified by the designer, providing a framework for providing design guidelines for new functional materials as a group of candidates. Note that the purpose of text abstraction here is not to avoid, for example, linking unique names with physical properties, but rather to prevent unique text expressions from corresponding to graph information.

[0048] Here, the learning device 2 of this embodiment is preferably configured to automatically generate text based on information about the physical properties of interest obtained from first-principles calculations or databases. For example, the learning device 2 is preferably configured to automatically generate text describing physical property information, such as geometric aspects of crystals, from local bonds to coordination structures and global symmetries such as space groups, from calculations and literature, in addition to text information described in publicly available literature. This is because, compared to simply collecting text from the web, it is expected that appropriate learning will be performed by using text data that reflects the physical properties of interest and eliminates unnecessary information. In addition to text that contains a large amount of information other than the physical properties of interest, it is also possible to add teacher labels that assume regression to the natural language input framework.

[0049] (2) The learning device 2 of this embodiment learns the crystalline structure of an inorganic material. The graph data is data that shows the crystalline structure of the inorganic material. The text data is data that includes information related to the crystalline structure of the inorganic material. This makes it possible to obtain an appropriate learning model based on a variety of feature representations related to the crystalline structure of the inorganic material. Learning models based on a variety of feature representations are suitable for the crystalline structure of inorganic materials, which often represent substances using a combination of various elements and have many rules governing the atomic coordinates, cell parameters, occupancies, and other aspects of the crystalline structure.

[0050] (3) The similarity evaluation unit 20 of the learning device 2 updates both the parameters of the first learning model and the parameters of the second learning model using a contrastive loss including a term combining the first feature amount and the second feature amount as a loss function. This allows the parameters of each learning model to be updated so that the similarity between graph data and text data showing the same crystal structure increases, and the similarity between graph data and text data showing different crystal structures decreases.

[0051] (4) The materials exploration system 1 includes a learning device 2 and an exploration device 4 that searches for a crystal structure of an inorganic material with desired physical properties using a learning model for materials exploration that has been trained by the learning device 2. In this way, by using an appropriate learning model based on various feature representations, it becomes possible to appropriately search for a crystal structure with desired physical properties.

[0052] (Modification of the first embodiment) The similarity evaluation unit 20 of the learning device 2 uses contrastive loss as a loss function to update both the parameters of the first learning model and the parameters of the second learning model, but is not limited to this.

[0053] (First Modification) The learning device 2 may be configured to update the parameters of each learning model using, as a loss function, a cross-contrastive loss including not only a term combining the first feature and the second feature as shown in FIG. 8 but also a term combining different second feature quantities.

[0054] Graph data has fewer ambiguous elements than text data, making it easier to clearly distinguish between them. Therefore, the accuracy of determining similarity can be improved by using a loss function that includes a term that combines first features corresponding to graph data. This contributes to improving the classification accuracy of features for graph data and text data.

[0055] (Second Modification) Furthermore, the learning device 2 may be configured to update the parameters of each learning model using, as a loss function, a dual contrastive loss including a term combining different first features and a term combining different second features, as shown in Fig. 9. In this way, the use of a loss function including a term combining second features corresponding to text data can also be expected to improve the accuracy of similarity determination.

[0056] (Second embodiment) Next, a second embodiment will be described with reference to Fig. 10. In this embodiment, differences from the first embodiment will be mainly described.

[0057] 10, the composition formula of the inorganic material is input as learning data to the learning device 2 of this embodiment instead of the CIF. The composition formula of the inorganic material is stored, for example, in an external database and provided to the learning device 2 from the database.

[0058] The learning device 2 includes a structure search unit 21 that searches for the crystal structure of an inorganic material from the composition formula of the inorganic material. The structure search unit 21 searches for the crystal structure of the inorganic material corresponding to the composition formula of the inorganic material through an external database.

[0059] The graph data creating unit 12 creates graph data showing the crystal structures of a large number of inorganic materials based on the information on the crystal structures of the inorganic materials corresponding to the composition formulas of the inorganic materials obtained by the structure searching unit 21 .

[0060] On the other hand, the text data creation unit 16 includes a text conversion unit 162 and an abstraction processing unit 163. The text conversion unit 162 of this embodiment converts information on the crystalline structure of the inorganic material corresponding to the composition formula of the inorganic material into text written in a predetermined natural language and outputs the text.

[0061] The rest of the configuration is the same as in the first embodiment. The learning device 2 of this embodiment can obtain the same effects as in the first embodiment that are achieved by a configuration that is common to or equivalent to the first embodiment.

[0062] (Third embodiment) Next, a third embodiment will be described with reference to Fig. 11. In this embodiment, differences from the first embodiment will be mainly described.

[0063] 11, the learning device 2 has a fine-tuning unit 22. The fine-tuning unit 22 uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit 20, as pre-training models, and further updates the parameters of the pre-training models. The fine-tuning unit 22 of this embodiment includes a screening unit 24, a data generation unit 26, a pre-training unit 28, and an error calculation unit 30 as functional units for realizing various functions.

[0064] The screening unit 24 selects physical property values ​​that satisfy predetermined conditions from information such as numerical values ​​and structures related to the characteristics calculated by the physical property calculation unit 161. The screening unit 24 sets a threshold value for the target physical property value, and abstracts and selects the property value as "high" or the like if the property value is equal to or greater than the threshold value. For example, if the electrical conductivity σ is equal to or greater than the threshold value, the screening unit 24 abstracts the property value as "high conductivity" or the like.

[0065] The data generating unit 26 stores data with conditions attached to some of the physical property values ​​related to the inorganic material in a database such as the storage device 3. The data generating unit 26 stores the data related to the physical property values ​​selected by the screening unit 24 in the database as the above-mentioned conditional data.

[0066] The pre-learning unit 28 uses the first learning model and the second learning model whose parameters have been updated by the similarity evaluation unit 20 as pre-learning models, inputs the data generated by the data generation unit 26 into the pre-learning models, and outputs features corresponding to the data.

[0067] The error calculation unit 30 calculates the error between the input to the pre-trained model and the output to the pre-trained model as a regression error, and updates the parameters of the pre-trained model so that the error converges. The error calculation unit 30 updates the parameters of the pre-trained model using, for example, an error backpropagation method. The error calculation unit 30 can be configured using a Transformer and a GCN. GCN is an abbreviation for Graph Convolutional Network.

[0068] The rest of the configuration is the same as in the first embodiment. The learning device 2 of this embodiment can obtain the same effects as in the first embodiment that are achieved by a configuration that is common to or equivalent to the first embodiment.

[0069] Furthermore, the learning device 2 of this embodiment has the following features.

[0070] (1) The learning device 2 includes a fine-tuning unit 22 that uses data with conditions attached to some of the physical property values ​​related to inorganic materials to further update the parameters of the first learning model and the second learning model whose parameters have been updated by the similarity evaluation unit 20. In this way, by updating the parameters of the pre-learning model using information with conditions attached to some of the physical property values ​​related to the crystal structure, it is possible to construct a learning model specialized for specific physical property values.

[0071] Here, the learning device 2 may be configured to calculate the dimensionless figure of merit ZT for, for example, a thermoelectric material, create a database of only those materials that are equal to or greater than a target value, and perform fine tuning using the database. Also, the learning device 2 may be configured to fine-tune, for example, physical properties that are highly similar in a pre-learned model.

[0072] (Fourth embodiment) Next, a fourth embodiment will be described with reference to Figures 12 and 13. In this embodiment, differences from the first embodiment will be mainly described.

[0073] 12, the learning device 2 has a fine-tuning unit 22A for further updating the parameters of the pre-training model. The fine-tuning unit 22A of this embodiment includes, as functional units for realizing various functions, a first feature output unit 32, a text output unit 34, a second feature output unit 36, a candidate selection unit 38, a pre-training unit 40, a physical property output unit 42, and an error calculation unit 44.

[0074] The first feature output unit 32 inputs the CIF file to the first learning model whose parameters have been updated by the similarity evaluation unit 20, and outputs a predetermined number of first features. Note that the first feature output unit 32 may be configured to output first features using graph data obtained from information other than the CIF file as input.

[0075] The text output unit 34 automatically outputs text data describing the physical properties desired by the user. For example, the text output unit 34 outputs text data describing multiple physical properties, such as "This substance has a high electrical conductivity σ and a low thermal conductivity."

[0076] The second feature output unit 36 ​​inputs text data describing desired physical properties to the second learning model whose parameters have been updated by the similarity evaluation unit 20, and outputs second features corresponding to the text data.

[0077] The candidate selection unit 38 selects, from among the first features output from the first feature output unit 32, those that are similar to the second features output from the second feature output unit 36 ​​as candidate substances having desired physical properties.

[0078] The pre-learning unit 40 uses the first learning model and the second learning model whose parameters have been updated by the similarity evaluation unit 20 as pre-learning models, inputs data regarding the candidate substances selected by the candidate selection unit 38 into the pre-learning models, and outputs features corresponding to the candidate substances.

[0079] The physical property output unit 42 outputs the physical properties, etc. of the candidate substance selected by the candidate selection unit 38. The physical property output unit 42 outputs the physical properties, etc. of the candidate substance through, for example, first-principles calculation, calculation equivalent to first-principles calculation, automatic experiment, database, etc.

[0080] The error calculation unit 44 calculates the error between the output to the pre-trained model and the output of the physical property output unit 42 as a regression error, and updates the parameters of the pre-trained model so that the error converges. The error calculation unit 44 updates the parameters of the pre-trained model by, for example, the error backpropagation method.

[0081] The learning device 2 configured in this way includes the fine tuning unit 22, and thereby further updates the parameters of the pre-learning model, for example, as shown in FIG.

[0082] Specifically, the learning device 2 inputs text data describing desired physical properties to a second learning model whose parameters have been updated by the similarity evaluation unit 20, and embeds the first feature amount output from the text encoder TEN into a fixed-length vector and outputs the same. The learning device 2 also inputs graph data to a first learning model whose parameters have been updated by the similarity evaluation unit 20. The learning device 2 then embeds the first feature amount output from the graph encoder GEN into a fixed-length vector and outputs the same.

[0083] Next, the learning device 2 calculates the similarity between the first feature and the second feature, and selects, from the first feature, those that are similar to the second feature as candidate substances.The learning device 2 then calculates the error between the output to the pre-trained model and the output of the physical property output unit 42 as a regression error, and updates the parameters of the pre-trained model so that the error converges.

[0084] The rest of the configuration is the same as in the first embodiment. The learning device 2 of this embodiment can obtain the same effects as in the first embodiment that are achieved by a configuration that is common to or equivalent to the first embodiment.

[0085] Furthermore, the learning device 2 of this embodiment has the following features.

[0086] (1) The fine-tuning unit 22A inputs text describing desired physical properties into the second learning model, whose parameters have been updated by the similarity evaluation unit 20, to obtain second features. Furthermore, the fine-tuning unit 22A selects, as candidate substances, those first features similar to the second features corresponding to the desired physical properties from among the first features output from the first learning model, whose parameters have been updated by the similarity evaluation unit 20. The fine-tuning unit 22A then further updates the parameters of the pre-training model using data related to the candidate substances. In this way, by updating the parameters of the pre-training model using information on candidate substances predicted to have the desired physical properties, it is possible to construct a learning model specialized for the desired physical properties.

[0087] In particular, the learning device 2 of this embodiment can search for desired physical properties using text data that ambiguously describes multiple physical properties, rather than directly searching for them in a database, etc. Furthermore, using text data that ambiguously describes multiple physical properties is expected to enable efficient searches.

[0088] (Fifth embodiment) Next, a fifth embodiment will be described with reference to Figures 14 and 15. In this embodiment, differences from the first embodiment will be mainly described.

[0089] The search device 4 of this embodiment is configured to input text data describing desired physical properties and search for a crystal structure of an inorganic material corresponding to the desired physical properties using its own pre-trained model. As shown in Fig. 14, the search device 4 includes a property input unit 46, a structure selection unit 48, an evaluation unit 50, and an identification unit 52 as functional units for realizing the search function for inorganic materials.

[0090] The characteristic input unit 46 is a functional unit for accepting input of characteristics desired by the user. The characteristic input unit 46 includes a text input unit 461 and a second information output unit 462. The text input unit 461 accepts text data describing desired physical properties. The second information output unit 462 outputs second feature values ​​corresponding to the text data describing the desired physical properties using a second learning model whose parameters have been updated by the learning device 2. Specifically, the second information output unit 462 inputs the text data describing the desired physical properties to a text encoder TEN, and embeds second feature values ​​corresponding to each piece of text data from the text encoder TEN into a fixed-length vector and outputs the vector.

[0091] The structure selection unit 48 is a functional unit for selecting candidates for a first feature to be compared with the output information from the characteristic input unit 46. The structure selection unit 48 includes a sampling unit 481, a graph generation unit 482, and a first information output unit 483.

[0092] The sampling unit 481 defines a fixed-length vector in which a first feature corresponding to the graph data is embedded as a latent space z, and samples some of the latent variables in the latent space z as candidate points. The latent space z is a space in which latent variables corresponding to the feature in the data are distributed. To improve search efficiency, it is desirable to use a latent space z in which a first feature obtained by inputting candidate substances predicted to have desired physical properties into the graph encoder GEN is embedded in a fixed-length vector. Note that the latent space z may be two-dimensional or multidimensional (more than two dimensions).

[0093] The graph generation unit 482 receives the latent variables in the latent space z as input and generates graph data indicating the crystalline structure of the inorganic material. Specifically, the graph generation unit 482 generates the graph data based on the candidate points sampled by the sampling unit 481.

[0094] The first information output unit 483 outputs the second feature amount corresponding to the graph data generated by the graph generation unit 482, using the first learning model whose parameters have been updated by the learning device 2. Specifically, the first information output unit 483 inputs the graph data to the graph encoder GEN, and embeds the first feature amount corresponding to each piece of graph data from the graph encoder GEN into a fixed-length vector and outputs it.

[0095] The evaluation unit 50 evaluates whether the second feature amount output from the characteristic input unit 46 and the first feature amount output from the structure selection unit 48 are similar. The evaluation unit 50 calculates a distance or divergence D indicating the similarity between the feature amounts, and determines whether the feature amounts are similar based on the magnitude of the distance or divergence D. If the feature amounts are not similar, the evaluation unit 50 determines whether the first feature amount and the second feature amount corresponding to the graph data of a newly sampled candidate point are similar.

[0096] The identification unit 52 identifies a second feature similar to the first feature corresponding to the text data describing the desired characteristics by the evaluation unit 50, and outputs the crystal structure of the inorganic material corresponding to the identified second feature as an inorganic material having the desired physical properties.

[0097] The searcher device 4 configured in this manner uses a second learning model whose parameters have been updated by the learning device 2 to find second feature amounts corresponding to text data describing desired physical properties, for example, as shown in FIG.

[0098] Furthermore, the search device 4 generates a latent space z by adding perturbations or the like to the first feature values ​​obtained by inputting candidate substances predicted to have desired properties into the graph encoder GEN and embedding the result in a fixed-length vector. The search device 4 samples some of the latent variables in the latent space z as candidate points and generates graph data based on the sampled candidate points. The search device 4 then inputs the graph data corresponding to the candidate points into the graph encoder GEN and obtains the first feature values ​​corresponding to each piece of graph data from the graph encoder GEN.

[0099] Next, the search device 4 determines whether the first feature corresponding to the text data describing the desired characteristic is similar to the second feature corresponding to the graph data sampled from the latent space z. By repeating this determination, the search device 4 identifies a second feature similar to the first feature corresponding to the text data describing the desired characteristic. For example, as shown in formula F4 in FIG. 15 , the search device 4 searches for latent variables with small divergence D, which indicates the similarity between each feature, and identifies the latent variable with the smallest divergence D as the second feature having the desired characteristic. Then, the search device 4 outputs the crystal structure of the inorganic material corresponding to the identified second feature as the crystal structure of the inorganic material having the desired physical properties.

[0100] The rest of the configuration is the same as in the first embodiment. The learning device 2 and material exploration system 1 of this embodiment can obtain the same effects as in the first embodiment that are achieved by a configuration that is common to or equivalent to the first embodiment.

[0101] (Modification of the fifth embodiment) As in the fifth embodiment, in order to improve search efficiency, it is desirable that the search device 4 input a candidate substance predicted to have desired properties into the graph encoder GEN, and embed the first feature amount obtained by this into a fixed-length vector, and use this as the latent space z, but this is not limited to this. For example, the search device 4 may input graph data in an external database into the graph encoder GEN, and embed the first feature amount obtained by this into a fixed-length vector, and use this as the latent space z.

[0102] (Other embodiments) Although typical embodiments of the present disclosure have been described above, the present disclosure is not limited to the above-described embodiments. The present disclosure can be modified in various ways, for example, as follows.

[0103] As in the above-described embodiment, the learning device 2 preferably includes an abstraction processing unit 163 that replaces specific descriptions contained in text data with abstract words based on predetermined rules, but is not limited to this. The learning device 2 may not include the abstraction processing unit 163 and may output second features corresponding to the text data without abstracting specific descriptions contained in the text data.

[0104] Furthermore, it is desirable for learning device 2 to automatically generate text based on information about the physical properties of interest obtained from first-principles calculations or a database, but this is not a limitation. Learning device 2 may also be configured to automatically collect text data from the web, for example.

[0105] As in the above-described embodiment, the learning device 2 of the present disclosure is suitable for, but not limited to, the crystalline structure of inorganic materials, which often represent substances made up of a combination of various elements. The learning device 2 may also be used to learn, for example, the crystalline structure of organic materials.

[0106] Although the learning device 2 in the above-described embodiment is configured to update the parameters of each learning model using a contrastive loss including a term combining the first feature amount and the second feature amount, the learning device 2 is not limited to this. The learning device 2 may also be configured to update the parameters of each learning model using a loss function other than the contrastive loss.

[0107] The description of the learning device 2 in the above embodiment also includes a description of a learning method using a computer, in addition to the learning device 2. According to the above learning method, an appropriate learning model based on a variety of feature representations can be obtained.

[0108] In the above-described embodiments, it goes without saying that the elements constituting the embodiments are not necessarily essential unless they are specifically stated as essential or are clearly considered essential in principle.

[0109] In the above-described embodiments, when numerical values ​​such as the number, values, amounts, ranges, etc. of components of the embodiments are mentioned, they are not limited to the specific numbers unless they are specifically stated as essential or are clearly limited to a specific number in principle.

[0110] In the above-described embodiments, when referring to the shapes, positional relationships, etc. of components, etc., the shapes, positional relationships, etc. are not limited to those unless otherwise specified or when they are fundamentally limited to specific shapes, positional relationships, etc.

[0111] The learning device 2 and learning method of the present disclosure may be implemented on a special-purpose computer by configuring a processor and memory programmed to execute one or more functions embodied in a computer program. The learning device 2 and learning method of the present disclosure may be implemented on a special-purpose computer by configuring a processor with one or more dedicated hardware logic circuits. The learning device 2 and learning method of the present disclosure may be implemented on one or more special-purpose computers configured with a combination of a processor and memory programmed to execute one or more functions and a processor configured with one or more hardware logic circuits. Furthermore, the computer program may be stored as instructions executed by a computer on a computer-readable non-transitory tangible recording medium.

[0112] [Aspects of the present disclosure] [First viewpoint] A learning device, comprising: a first feature calculation unit (14) that receives graph data indicating a crystal structure of a predetermined material as an input and calculates the first feature using a first learning model that outputs a feature of the crystal structure of the predetermined material corresponding to the graph data as a first feature; a second feature calculation unit (18) that receives text data indicating a crystal structure of the predetermined material as input and calculates the second feature using a second learning model that outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; a similarity evaluation unit (20) that updates both the parameters of the first learning model and the parameters of the second learning model so that a similarity between a combination of the first feature amount and the second feature amount corresponding to the crystal structure of the same material is high and a similarity between a combination of the first feature amount and the second feature amount corresponding to the crystal structure of a different material is low; A learning device comprising:

[0113] [Second perspective] an abstraction processing unit (163) that replaces specific descriptions included in the text data with abstract words based on predetermined rules; The learning device according to a first aspect, wherein the second feature calculation unit calculates the second feature using as input the text data abstracted by the abstraction processing unit.

[0114] [Third Perspective] the predetermined material is an inorganic material, the graph data is data representing a crystal structure of the inorganic material, The learning device according to the first or second aspect, wherein the text data is data including information related to the crystal structure of the inorganic material.

[0115] [Fourth viewpoint] The learning device according to any one of the first to third aspects, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using a contrastive loss including a term combining the first feature and the second feature as a loss function.

[0116] [Fifth viewpoint] The learning device according to a fourth aspect, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using the loss function including a term that combines different first features.

[0117] [Sixth viewpoint] The learning device according to a fourth aspect, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using the loss function including a term combining different first features and a term combining different second features.

[0118] [Seventh viewpoint] The learning device according to any one of the first to sixth aspects, wherein the text data includes data describing physical properties of the specified material that cannot be represented by the graph data.

[0119] [Eighth viewpoint] The learning device according to any one of the first to seventh aspects, further comprising a fine tuning unit (22) that uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit, as pre-learning models, and further updates the parameters of the pre-learning models using data that has conditions attached to some of the physical property values ​​related to the specified material.

[0120] [Ninth viewpoint] a fine tuning unit (22A) that sets the first learning model and the second learning model, the parameters of which have been updated by the similarity evaluation unit, as pre-trained models, and further updates the parameters of the pre-trained models; The fine tuning unit A learning device according to any one of the first to seventh aspects, wherein the second feature is obtained by inputting text describing desired physical properties into the second learning model whose parameters have been updated by the similarity evaluation unit, and selecting as candidate substances from the first feature values ​​output from the first learning model whose parameters have been updated by the similarity evaluation unit those that are similar to the second feature value corresponding to the desired physical properties, and further updating the parameters of the pre-learning model using data regarding the candidate substances.

[0121] [10th viewpoint] A materials exploration system, comprising: A learning device (2) according to claim 1 or 2; A material exploration system comprising: a search device (4) that uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit of the learning device, as learned learning models for material exploration, and searches for a crystal structure of the specified material having desired physical properties using the learning models.

[0122] [11th viewpoint] A learning method using a computer (2), obtaining the first feature amount using a first learning model that receives graph data indicating a crystal structure of a predetermined material as input and outputs a feature amount of the crystal structure of the predetermined material corresponding to the graph data as a first feature amount; obtaining the second feature using a second learning model that receives text data indicating a crystal structure of the predetermined material as input and outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; updating both the parameters of the first learning model and the parameters of the second learning model so that the similarity between combinations of the first feature and the second feature corresponding to the crystal structures of the same material is high and the similarity between combinations of the first feature and the second feature corresponding to the crystal structures of different materials is low. [Explanation of symbols]

[0123] 2 Learning device 14 First feature calculation unit 18 Second feature calculation unit 20 Similarity Evaluation Section

Claims

1. A learning device, comprising: a first feature calculation unit (14) that receives graph data indicating a crystal structure of a predetermined material as an input and calculates the first feature using a first learning model that outputs a feature of the crystal structure of the predetermined material corresponding to the graph data as a first feature; a second feature calculation unit (18) that receives text data indicating a crystal structure of the predetermined material as input and calculates the second feature using a second learning model that outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; a similarity evaluation unit (20) that updates both the parameters of the first learning model and the parameters of the second learning model so that a similarity between a combination of the first feature amount and the second feature amount corresponding to the crystal structure of the same material is high and a similarity between a combination of the first feature amount and the second feature amount corresponding to the crystal structure of a different material is low; A learning device comprising:

2. an abstraction processing unit (163) that replaces specific descriptions included in the text data with abstract words based on predetermined rules; The learning device according to claim 1 , wherein the second feature amount calculation unit calculates the second feature amount using the text data abstracted by the abstraction processing unit as an input.

3. the predetermined material is an inorganic material, the graph data is data representing a crystal structure of the inorganic material, The learning device according to claim 1 , wherein the text data includes information related to the crystal structure of the inorganic material.

4. 3. The learning device according to claim 1, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using a contrastive loss including a term combining the first feature and the second feature as a loss function.

5. 5. The learning device according to claim 4, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using the loss function including a term that combines different first features.

6. 5. The learning device according to claim 4, wherein the similarity evaluation unit updates both the parameters of the first learning model and the parameters of the second learning model using the loss function including a term combining different first features and a term combining different second features.

7. 3. The learning device according to claim 1, wherein the text data includes data describing physical properties of the predetermined material that cannot be expressed by the graph data.

8. The learning device described in claim 1 or 2, further comprising a fine-tuning unit (22) that uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit, as pre-learning models, and further updates the parameters of the pre-learning model using data that conditions some of the physical property values ​​related to the specified material.

9. a fine tuning unit (22A) that sets the first learning model and the second learning model, the parameters of which have been updated by the similarity evaluation unit, as pre-learned models and further updates the parameters of the pre-learned models; The fine tuning unit A learning device as described in claim 1 or 2, wherein text describing desired physical properties is input into the second learning model whose parameters have been updated by the similarity evaluation unit to obtain the second feature, and from the first feature output from the first learning model whose parameters have been updated by the similarity evaluation unit, one similar to the second feature corresponding to the desired physical property is selected as a candidate substance, and further the parameters of the pre-learning model are updated using data regarding the candidate substance.

10. A materials exploration system, comprising: A learning device (2) according to claim 1 or 2; a search device (4) that uses the first learning model and the second learning model, whose parameters have been updated by the similarity evaluation unit of the learning device, as learned learning models for material search, and searches for a crystal structure of the specified material having desired physical properties using the learning models.

11. A learning method using a computer (2), obtaining the first feature amount using a first learning model that receives graph data indicating a crystal structure of a predetermined material as input and outputs a feature amount of the crystal structure of the predetermined material corresponding to the graph data as a first feature amount; obtaining the second feature using a second learning model that receives text data indicating a crystal structure of the predetermined material as input and outputs the crystal structure of the predetermined material corresponding to the text data as a second feature; updating both the parameters of the first learning model and the parameters of the second learning model so that the similarity between combinations of the first feature and the second feature corresponding to the crystal structure of the same material is high, and the similarity between combinations of the first feature and the second feature corresponding to the crystal structure of different materials is low.

Images