Search method, information processing device, and program

Abstract

In this method for searching for a crystal structure of an inorganic material, a model (GM) is trained to include a first function for receiving, as an input, material information relating to the crystal structure of the inorganic material and outputting a latent variable in a latent space (LS) corresponding to the material information, and a second function for receiving, as an input, the latent variable in the latent space (LS) and outputting the sample information. The search method additionally includes identifying the crystal structure of an inorganic material having a score that is close to or greater than a prescribed target score using an optimization algorithm from the trained model.

Description

Search method, information processing device, and program CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application is based on Japanese Patent Application No. 2023-213958, filed on December 19, 2023, the contents of which are incorporated herein by reference.

[0002] The present disclosure relates to a method for searching for a crystal structure of an inorganic material having desired properties, an information processing device, and a program for use in the information processing device.

[0003] In the technical field where materials having desired properties are searched for based on information such as the structure of the material, it has been desirable to reduce the number of searches required to find a material having properties that meet the specified purpose.

[0004] In response to this, for example, Patent Document 1 discloses a technology for predicting a target property value of each material point in a search space based on a Gaussian process regression model algorithm that uses a target property value of a known material, whose target property value is known as a candidate material, and a feature vector. Note that the search space is a space in which points representing each candidate material are distributed in a space of dimensions of feature vectors corresponding to a plurality of candidate materials.

[0005] Japanese Patent Application Laid-Open No. 2023-69211

[0006] In recent years, as product performance requirements have increased, it has become essential to improve the performance of the materials themselves that make up the products. However, until now, inorganic materials with new properties have been discovered accidentally through experiments, etc., and the search for such materials has required a great deal of time. Therefore, the present inventors considered using the technology described in Patent Document 1 when searching for a crystal structure of an inorganic material with desired properties.

[0007] However, the technology described in Patent Document 1 uses target property values ​​and feature vectors of known materials whose target property values ​​are known as candidate materials, and the increased reliance on this known information makes it difficult to search for unknown structures that are not found in known material structures. Thus, there is still room for improvement in technology for searching for crystalline structures of inorganic materials having desired properties. The present disclosure aims to improve the search for crystalline structures of inorganic materials having desired properties.

[0008] According to one aspect of the present disclosure, a method for searching for a crystalline structure of an inorganic material includes: training a model including a first function that receives material information related to a crystalline structure of the inorganic material expressed in a predetermined notation as input and outputs a latent variable in a latent space corresponding to the material information; and a second function that receives the latent variable in the latent space as input and outputs the material information; and using an optimization algorithm from the trained model to identify a crystalline structure of the inorganic material having a score close to or exceeding a predetermined target score.

[0009] According to another aspect of the present disclosure, an information processing device for searching for a crystal structure of an inorganic material includes: a control unit; and a memory unit, wherein the control unit learns a model including a first function that receives material information regarding the crystal structure of the inorganic material expressed in a predetermined notation as input and outputs a latent variable in a latent space corresponding to the material information, and a second function that receives the latent variable in the latent space as input and outputs the material information, and identifies a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score from the learned model using an optimization algorithm.

[0010] According to yet another aspect of the present disclosure, a program used in an information processing device that searches for a crystal structure of an inorganic material causes the information processing device to function as: a model learning unit that learns a model including a first function that receives material information on the crystal structure of the inorganic material expressed in a predetermined notation as an input and outputs a latent variable in a latent space corresponding to the material information; and a second function that receives the latent variable in the latent space as an input and outputs the material information; and a structure identification unit that uses an optimization algorithm from the learned model to identify a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score.

[0011] In this way, if the crystal structure of an inorganic material is searched for using a model including the first function and the second function, it becomes easier to search for the crystal structure of an unknown inorganic material compared to searching from a data group consisting of known material information.

[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.

[0013] FIG. 1 is a schematic configuration diagram of an information processing device according to a first embodiment. FIG. 2 is an explanatory diagram for explaining the function of a model learning unit of the information processing device according to the first embodiment. FIG. 3 is an explanatory diagram for explaining the function of a structure identification unit of the information processing device according to the first embodiment. FIG. 4 is a flowchart showing the flow of a process for searching for a crystalline structure of an inorganic material executed by the information processing device according to the first embodiment. FIG. 5 is an explanatory diagram for explaining search results for a crystalline structure of an inorganic material. FIG. 6 is an explanatory diagram for explaining the relationship between the number of searches for a crystalline structure of an inorganic material and search results in the information processing device. FIG. 7 is an explanatory diagram for explaining the function of a model learning unit of an information processing device according to a second embodiment. FIG. 8 is an explanatory diagram for explaining the function of a structure identification unit of the information processing device according to the second embodiment.

[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 FIGS. 1 to 6. In this embodiment, an example will be described in which the crystalline structure of an inorganic material is the search target. In this embodiment, an information processing device 10 searches for a crystalline structure of an inorganic material having desired physical properties.

[0016] The information processing device 10 is configured by at least one of a user terminal, a server, and a general-purpose or dedicated electronic device, or a combination of these, etc. As shown in Fig. 1, the information processing device 10 is configured to include a control unit 20, a storage unit 30, an input unit 40, and an output unit 50.

[0017] The control unit 20 controls each unit of the information processing device 10 and executes processing related to the operation of the information processing device 10. The control unit 20 is configured by a control circuit including one or more processors. The processor is configured by a general-purpose processor such as a CPU or GPU, or a dedicated processor specialized for searching for material structures. Note that CPU is an abbreviation for Central Processing Unit. GPU is an abbreviation for Graphics Processing Unit. Note that the control unit 20 may be configured by circuits such as an FPGA or ASIC.

[0018] The storage unit 30 stores data used in the operation of the information processing device 10 and data obtained by the operation of the information processing device 10. The storage unit 30 is configured by a non-transitory physical storage medium such as a semiconductor memory, a magnetic memory, or an optical memory. Note that RAM is an abbreviation for Random Access Memory, and ROM is an abbreviation for Read Only Memory.

[0019] The storage unit 30 stores various types of information, such as experimental data ED used to search for the crystalline structure of inorganic materials, a database DB of crystalline structures, and the like, as material information relating to the crystalline structure of inorganic materials.

[0020] The material information includes the crystal structure itself, a powder diffraction pattern specific to the crystal structure, various physical property values, etc. Material information related to the crystal structure of an inorganic material is expressed in a predetermined notation method, such as the common data format for crystallography known as CIF and the notation method for crystal properties using Fourier transform known as FTCP.

[0021] CIF is a data format that stores crystal structures. It includes information such as atomic coordinates, cell parameters, symmetry, and atomic species of the crystal structure. CIF is an abbreviation for Crystallographic Information File.

[0022] FTCP is a notation that combines real and imaginary parts. The real part of FTCP consists of the type of atom, atomic occupancy, atomic coordinates, cell parameters, and atomic properties, while the imaginary part is the distance from the origin of each atom in reciprocal space and the discrete Fourier transform of the atomic properties along various spatial frequencies. FTCP is an abbreviation for Fourier-Transformed Crystal Properties.

[0023] The input unit 40 is an input interface such as a keyboard or a touch panel, and accepts input operations of data used in the operation of the information processing device 10. The input unit 40 is not essential to the information processing device 10, and may be connected to the information processing device 10 as an external input device, for example. A data file containing material information on the crystal structure of an inorganic material is input to the input unit 40, for example.

[0024] The output unit 50 is an interface for outputting data such as a display, and outputs data obtained by the operation of the information processing device 10. Note that the output unit 50 is not essential to the information processing device 10, and may be connected to the information processing device 10 as an external output device, for example.

[0025] The various functions of the information processing device 10 are realized by executing various programs stored in the storage unit 30 by the control unit 20 of the information processing device 10. For example, the function of searching for the crystalline structure of an inorganic material is realized by executing a program for searching for the crystalline structure of an inorganic material by the control unit 20 of the information processing device 10.

[0026] Specifically, the control unit 20 includes functional units for realizing various functions of the information processing device 10. The control unit 20 of the present embodiment includes a model learning unit 21 and a structure identification unit 22 as functional units for realizing various functions of the information processing device 10.

[0027] As shown in FIG. 2, the model learning unit 21 receives as input material information relating to the crystalline structure of an inorganic material expressed in a predetermined notation, and learns a generative model GM including an encoder ENC and a decoder DEC.

[0028] The encoder ENC is a function that reduces the dimension of data and converts features in the data into latent variables. In this embodiment, the encoder ENC corresponds to a "first function" that receives material information related to the crystalline structure of an inorganic material as input and outputs latent variables in the latent space LS corresponding to the material information. The decoder DEC is a function that restores the original data from the features. The decoder DEC corresponds to a "second function" that receives latent variables in the latent space LS as input and outputs material information. The latent space LS is a space in which latent variables corresponding to features in the data are distributed. The latent space LS may be two-dimensional or multidimensional (more than two dimensions).

[0029] The generative model GM is a model including the above-mentioned "first function" and "second function." The generative model GM is constructed as an unsupervised machine learning model using a neural network. The generative model GM of this embodiment is configured as a learning model using a variational autoencoder called VAE. A variational autoencoder is an autoencoder AE having an encoder ENC and a decoder DEC, in which a probability distribution is introduced into the latent variables. In a variational autoencoder, features in the input data are used as latent variables in a probabilistic model. In a variational autoencoder, the encoder ENC calculates a mean vector and a variance vector, and samples latent variables from a normal distribution based on the calculated mean vector and variance vector. The variational autoencoder then decodes the sampled latent variables to output material information as data. Note that VAE is an abbreviation for Variational Auto Encoder.

[0030] The structure identification unit 22 uses an optimization algorithm from the generative model GM trained by the model learning unit 21 to identify a crystal structure having a score close to or exceeding a target score corresponding to the physical properties desired for the crystal structure of the inorganic material. In this embodiment, the structure identification unit 22 employs a Bayesian optimization algorithm as the optimization algorithm. The Bayesian optimization algorithm is a type of optimization algorithm that uses uncertainty to search for the next value to be searched for, and optimizes a black-box function using a Gaussian process.

[0031] 3, the structure identification unit 22 samples some of the latent variables in the latent space LS as candidate points and calculates the score of the crystal structure identified using the sampled candidate points. The structure identification unit 22 then repeats a process of comparing the score of the crystal structure identified based on the sampled candidate points with a target score, thereby identifying a desired crystal structure having a score close to or exceeding the target score.

[0032] Next, a series of processing steps for searching for the crystal structure of an inorganic material using the information processing device 10 will be described with reference to Fig. 4. The control routine shown in Fig. 4 is executed by the control unit 20 of the information processing device 10, for example, when a command signal instructing the information processing device 10 to start searching is input.

[0033] 4, in step S100, the information processing device 10 performs a process of digitizing material information related to the crystalline structure of inorganic materials stored in the experimental data ED and the database DB for use in training the generative model GM. In this process, the information related to the crystalline structure of inorganic materials is digitized using notation such as CIF or FTCP. Note that this process may be a process of digitizing material information related to the crystalline structure of inorganic materials in the information processing device 10, or a process of acquiring digitized material information from another device.

[0034] Next, in step S110, the information processing device 10 performs a learning process for the generative model GM. In this process, for example, the encoder ENC and decoder DEC of the generative model GM are trained so that the reconstruction error and the Karlback-Leiber divergence (so-called KLD) are used as error functions to minimize each error. Note that the reconstruction error is the difference between the data input to the generative model GM and the data reconstructed by the generative model GM. The Karlback-Leiber divergence indicates the distance between the distribution of the data input to the generative model GM and the distribution of the data output from the generative model GM.

[0035] Here, among inorganic materials, those used for a specific purpose are referred to as target materials, and those other than the target materials are referred to as non-target materials. Target materials include, for example, materials for oxidation-reduction purposes, materials for hydrogen storage purposes, magnetic materials, materials for weight reduction purposes, electret materials, CO ２ Examples of such materials include reduction materials, conductive materials, materials for heat management, biosensing materials, and surface treatment materials.

[0036] In the learning process of the generative model GM, the encoder ENC and decoder DEC are trained using target data related to the target material and non-target data related to the non-target material. A generative model GM trained based on the target material and non-target materials may be less susceptible to accuracy degradation due to extrapolation than a generative model GM trained based only on the target material used in inorganic material applications. For this reason, when searching for the crystal structure of an unknown inorganic material, it is preferable to train the generative model GM based on the target material and non-target materials.

[0037] Next, in step S120, the information processing device 10 sets a target score corresponding to the physical property information desired for the crystalline structure of the inorganic material and a set number of times that serves as an upper limit for the number of times that the crystalline structure is searched. The target score and the number of times that the crystalline structure is searched may be set to values ​​that are set in advance in the storage unit 30, or may be set to values ​​that are input via the input unit 40. The target score is set as a physical property that aims to achieve stability of the crystalline structure of the inorganic material, such as formation energy. The target score may be set as a score that is desired to converge to, or may be set as a score that is desired to be exceeded.

[0038] Next, in step S130, the information processing device 10 extracts candidate points of latent variables to be sampled from the latent space LS using a Bayesian optimization algorithm from the trained generative model GM. Then, in step S140, the information processing device 10 samples candidate points of latent variables from the latent space LS.

[0039] Next, in step S150, the information processing device 10 outputs material information of the crystalline structure of the inorganic material corresponding to the latent variables sampled from the latent space LS from the decoder DEC, and calculates a score for the output material information. The score for the material information is calculated, for example, by calculation processing or numerical simulation for quantifying various properties of the material information. Note that the score for the material information may be calculated using a machine learning model such as a surrogate model, or may be obtained externally as a quantified score calculated by a personal task.

[0040] Next, in step S160, the information processing device 10 compares the score of the crystal structure sampled from the latent space LS with the target score to determine whether the score of the sampled crystal structure is appropriate. Specifically, the information processing device 10 determines whether the difference between the score of the crystal structure sampled from the latent space LS and the target score is within a preset range or exceeds the target score, or whether the number of searches is a preset number.

[0041] If the difference between the score of the crystal structure sampled from the latent space LS and the target score is outside a preset range or the score is equal to or less than the target score, and the number of searches is less than the preset number, the information processing device 10 returns to step S130. That is, the information processing device 10 extracts candidate points of the latent variables to be sampled next from the latent space LS using a Bayesian optimization algorithm based on the generative model GM.

[0042] On the other hand, if the difference between the score of the sampled crystal structure and the target score is within a preset range or exceeds the target score, or if the number of searches has reached a preset number, the information processing device 10 proceeds to step S170. In step S170, the information processing device 10 outputs the search results, indicating that the search for the crystal structure has been completed.

[0043] The search results are output via the output unit 50 connected to the control unit 20. For example, as shown in FIG. 5 , the search results are displayed on a display constituting the output unit 50 as a two-dimensional image in which latent variables sampled from the latent space LS and scores corresponding to the latent variables are superimposed on the latent space LS. In this way, it is desirable that the search results be provided in a visually comprehensible manner. Note that the output of the search results is not limited to two-dimensional images, and may be multidimensional images, etc.

[0044] The search technology for the crystalline structure of an inorganic material according to the present embodiment described above includes a method for searching for the crystalline structure of an inorganic material, an information processing device 10 for searching for the crystalline structure of an inorganic material, and a program for searching for the crystalline structure of an inorganic material. The search technology of the present invention has the following features. Specifically, the search technology of the present invention trains a generative model GM including an encoder ENC that receives material information about the crystalline structure of the inorganic material as input and outputs latent variables in a latent space LS, and a decoder DEC that receives the latent variables in the latent space LS as input and outputs material information. The search technology of the present invention then uses an optimization algorithm from the trained generative model GM to identify a crystalline structure of the inorganic material that has a score close to or exceeding a predetermined target score.

[0045] In this way, if the crystalline structure of an inorganic material is searched for using a generative model GM including an encoder ENC and a decoder DEC, it becomes easier to search for the crystalline structure of an unknown inorganic material compared to searching from a data group composed of known material information.

[0046] Here, in the real space, when considering the ternary compound Ax, By, and Cz, 10 ６ It is necessary to optimize a space with more than 100 dimensions. Such optimization is extremely difficult and unrealistic.

[0047] In contrast, in the present invention, materials are represented in a latent space LS with a reduced number of dimensions, and multiple parameters can be optimized with a reduced number of dimensions, which has the advantage of making it easier to explore the crystal structure of unknown inorganic materials.

[0048] Furthermore, according to this embodiment, the following effects can be obtained.

[0049] (1) When training the generative model GM, the encoder ENC and decoder DEC are trained using target data related to the target material and non-target data related to the non-target material. A generative model GM trained based on the target material and non-target materials may be less susceptible to accuracy degradation due to extrapolation than a generative model GM trained based only on the target material used in inorganic material applications. For this reason, when searching for the crystal structure of an unknown inorganic material, it is preferable to train the generative model GM based on the target material and non-target materials.

[0050] (2) In this embodiment, a variational autoencoder is used as the generative model GM. In this way, if the generative model GM is configured using a variational autoencoder, latent variables sampled from a normal distribution are decoded, which has the advantage of making it easier to identify effective crystal structures from the latent space LS.

[0051] (3) The optimization algorithm used to search for crystalline materials of inorganic materials is a Bayesian optimization algorithm. When identifying the crystalline structure of an inorganic material, some of the latent variables in the latent space LS are sampled as candidate points. Then, the score of the crystalline structure of the inorganic material identified using the sampled candidate points is calculated, and the calculated score is compared with a target score. This process is repeated to identify a crystalline structure of the inorganic material having a score close to or exceeding the target score.

[0052] In this way, by sampling latent variables according to a Bayesian optimization algorithm, the number of trials required to identify a crystal structure having a score close to or exceeding the target score can be reduced compared to when latent variables are sampled randomly.

[0053] 6 shows the relationship between the number of searches for the crystalline structure of an inorganic material and the search results of the information processing device 10. Specifically, FIG. 6 shows the search results when searching for a crystalline structure of an inorganic material with a low formation energy. As shown in FIG. 6, the information processing device 10 of the present invention was able to identify a crystalline structure of an inorganic material predicted to have a low formation energy with fewer searches than when randomly sampling latent variables.

[0054] Second Embodiment Next, a second embodiment will be described with reference to Figures 7 and 8. In this embodiment, differences from the first embodiment will be mainly described.

[0055] As shown in Fig. 7, the generative model GM of this embodiment is configured as a learning model that uses a conditional variational autoencoder called CVAE. The conditional variational autoencoder is a variational autoencoder that is trained by adding a condition vector as a condition variable to the variational autoencoder. In the conditional variational autoencoder of this embodiment, in addition to material information related to the crystalline structure of the inorganic material, some of the characteristics of the crystalline structure of the inorganic material are assigned as condition vectors to the variational autoencoder for training.

[0056] Here, in three-dimensional lattice structures, crystal systems are classified into seven types, namely, rhombohedral, cubic, hexagonal, monoclinic, orthorhombic, tetragonal, and triclinic, based on cell parameters that represent the size of the unit cell. In this embodiment, the cell parameters are used as condition vectors among the information contained in the CIF. The size of the unit cell is determined by the lengths of the lattice sides (e.g., x, y, z) and the angles between the lattice sides (e.g., α, β, γ).

[0057] The control unit 20 of this embodiment receives as input material information and a condition vector relating to the crystalline structure of an inorganic material expressed in a predetermined notation, and trains a generative model GM including an encoder ENC and a decoder DEC.

[0058] In addition, the control unit 20 uses an optimization algorithm from the generative model GM that has been trained by the model training unit 21 to identify a crystal structure that has a score that is close to or exceeds a target score that corresponds to the physical properties desired for the crystal structure of the inorganic material.

[0059] 8, the control unit 20 samples some of the latent variables in the latent space LS to which the condition vector has been added as candidate points, and calculates the score of the crystal structure identified using the sampled candidate points. The structure identification unit 22 then repeats a process of comparing the score of the identified crystal structure based on the sampled candidate points with a target score, thereby identifying a desired crystal structure having a score close to or exceeding the target score.

[0060] The other points are the same as those in the first embodiment. The technology for searching for the crystalline structure of an inorganic material according to the present embodiment can obtain the same effects as those in the first embodiment, which are achieved by a configuration common to or equivalent to the first embodiment.

[0061] Furthermore, the technology for searching for the crystalline structure of an inorganic material according to this embodiment includes the following features.

[0062] (1) In this embodiment, a conditional variational autoencoder is used as the generative model GM. In this way, configuring the generative model GM with a conditional variational autoencoder allows constraints to be imposed on the output from the decoder DEC, which has the advantage of making it easier to identify effective crystal structures even from a sparse latent space LS.

[0063] (Modification of Second Embodiment) In the second embodiment, cell parameters are used as an example of the condition vector, but the condition vector is not limited to this. The condition vector may be other information such as atomic coordinates, symmetry, and atomic species of the crystal structure, instead of the cell parameters.

[0064] Other Embodiments Although typical embodiments of the present disclosure have been described above, the present disclosure is not limited to the above-described embodiments and can be modified in various ways, for example, as follows.

[0065] As in the above embodiment, when learning the generation model GM, it is desirable to use target data related to the target material and non-target data related to the non-target material, but this is not limited to this. The learning process of the generation model GM may be configured to use only target data related to the target material, for example.

[0066] As in the above-described embodiment, the generative model GM is preferably configured as a variational autoencoder or a conditional variational autoencoder, but is not limited to this. The generative model GM may also be configured as other models, such as an autoencoder AE or a transformer. Furthermore, in this embodiment, an example of searching for the crystalline structure of an inorganic material using a generative model GM including an encoder ENC and a decoder DEC has been described, but the technology for searching for the crystalline structure of an inorganic material is not limited to this. The crystalline structure of an inorganic material may also be searched for using a "model" including a "first function" and a "second function" that is different from the above-described generative model.

[0067] In the above embodiment, the Bayesian optimization algorithm is used as an example of the optimization algorithm used to search for crystalline materials of inorganic materials, but the optimization algorithm is not limited to this. For example, the optimization algorithm may be another algorithm such as gradient descent.

[0068] 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.

[0069] 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.

[0070] 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.

[0071] The controller and method of the present disclosure may be implemented on a special-purpose computer by configuring a processor and memory programmed to perform one or more functions embodied in a computer program. The controller and 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 controller and method of the present disclosure may be implemented on one or more special-purpose computers configured with a processor and memory programmed to perform one or more functions in combination with a processor configured with one or more hardware logic circuits. The computer program may also be stored on a computer-readable non-transitory tangible storage medium as instructions executed by a computer.

[0072] [Aspects of the present disclosure]

[0073] [First Aspect] A method for searching for a crystalline structure of an inorganic material, the method comprising: training a model (GM) including a first function that receives material information on a crystalline structure of the inorganic material expressed in a predetermined notation as input and outputs latent variables on a latent space (LS) corresponding to the material information; and a second function that receives the latent variables on the latent space as input and outputs the material information; and using an optimization algorithm from the trained model to identify a crystalline structure of the inorganic material that has a score close to or exceeding a predetermined target score.

[0074] [Second Aspect] When, among the inorganic materials, those used for a predetermined purpose are defined as target materials and those other than the target materials are defined as non-target materials, the searching method according to the first aspect, wherein by learning the model, the first function and the second function are learned using target data relating to the target material and non-target data relating to the non-target material.

[0075] [Third Aspect] The search method according to the first or second aspect, wherein the model uses a variational autoencoder.

[0076] [Fourth Aspect] The search method according to the first or second aspect, wherein the model uses a conditional variational autoencoder.

[0077] [Fifth Aspect] The optimization algorithm is a Bayesian optimization algorithm, and the identifying the crystalline structure of the inorganic material includes repeating a process of sampling some of the latent variables in the latent space as candidate points, determining the score of the crystalline structure of the inorganic material identified using the sampled candidate points, and comparing the determined score with the target score, thereby identifying the crystalline structure of the inorganic material having the score close to or exceeding the target score.

[0078] [Sixth Aspect] An information processing device for searching for a crystal structure of an inorganic material, comprising: a control unit (20); and a storage unit (30), wherein the control unit learns a model (GM) including a first function that receives material information on a crystal structure of the inorganic material expressed in a predetermined notation as an input and outputs a latent variable on a latent space (LS) corresponding to the material information, and a second function that receives the latent variable on the latent space as an input and outputs the material information, and identifies a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score from the learned model using an optimization algorithm.

[0079] [Seventh Aspect] A program used in an information processing device (10) that searches for a crystal structure of an inorganic material, causing the information processing device to function as: a model learning unit (21) that learns a model (GM) including a first function that receives material information on a crystal structure of the inorganic material expressed in a predetermined notation as an input and outputs latent variables on a latent space (LS) that correspond to the material information; and a second function that receives the latent variables on the latent space as an input and outputs the material information; and a structure identification unit (22) that uses an optimization algorithm from the learned model to identify a crystal structure of the inorganic material that has a score close to or exceeding a predetermined target score.

Claims

1. A method for searching for a crystal structure of an inorganic material, comprising: training a model (GM) including a first function that receives material information on a crystal structure of the inorganic material expressed in a predetermined notation and outputs a latent variable in a latent space (LS) corresponding to the material information, and a second function that receives the latent variable in the latent space as input and outputs the material information; and using an optimization algorithm from the trained model to identify a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score.

2. The exploration method of claim 1, wherein, among the inorganic materials, those used for a specific purpose are defined as target materials, and those other than the target materials are defined as non-target materials, and in learning the model, the first function and the second function are learned using target data related to the target material and non-target data related to the non-target material.

3. The search method according to claim 1, wherein the model is a variational autoencoder.

4. The search method according to claim 1, wherein the model is a conditional variational autoencoder.

5. The search method according to claim 1, wherein the optimization algorithm is a Bayesian optimization algorithm, and the identifying of the crystalline structure of the inorganic material includes repeating a process of sampling a portion of the latent variables in the latent space as candidate points, determining the score of the crystalline structure of the inorganic material identified using the sampled candidate points, and comparing the determined score with the target score, thereby identifying the crystalline structure of the inorganic material having the score close to or exceeding the target score.

6. An information processing device for searching for a crystal structure of an inorganic material, comprising: a control unit (20); and a memory unit (30), wherein the control unit learns a model (GM) including a first function that receives material information on the crystal structure of the inorganic material expressed in a predetermined notation and outputs a latent variable on a latent space (LS) corresponding to the material information, and a second function that receives the latent variable on the latent space as input and outputs the material information, and identifies a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score from the learned model using an optimization algorithm.

7. A program for use in an information processing device (10) that searches for a crystal structure of an inorganic material, the program causing the information processing device to function as a model learning unit (21) that learns a model (GM) including a first function that receives material information on the crystal structure of the inorganic material expressed in a predetermined notation and outputs a latent variable in a latent space (LS) corresponding to the material information, and a second function that receives the latent variable in the latent space as input and outputs the material information, and a structure identification unit (22) that uses an optimization algorithm from the learned model to identify a crystal structure of the inorganic material having a score close to or exceeding a predetermined target score.

Images