Material search program, material search system, material search method, and material search device

Abstract

To reduce the time for searching for a material.SOLUTION: A material search program is to cause a plurality of calculators connected via a network to execute, in a decentralized manner, processing that includes: generating a first learning model that outputs a first physical property value through first machine learning using first training data; generating a second learning model that outputs a second physical property value through second machine learning using second training data; performing first transition learning, by using the second training data, for the first learning model so that it outputs the second physical property value; selecting a learning model that satisfies a standard from the second learning model and the first learning model for which the first transition learning is performed; determining a structure from which the evaluation value of the second physical property value is acquired on the basis of the selected learning model; acquiring the evaluation value in the determined structure; and adding the structure descriptor of the determined structure and the evaluation value to the second training data.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present invention relates to a material exploration program, a material exploration system, a material exploration method, and a material exploration apparatus. [Background technology]

[0002] The technology of searching for materials with desired physical properties is called material exploration. In material exploration, the structure of a material is changed in various ways to obtain the physical properties, and a structure with the desired physical properties is identified. When a human performs material exploration, the human synthesizes a material with a certain structure in an experiment and measures the physical properties of the material. Next, the human considers the measurement results and determines a structure that is expected to show even better physical properties. The human then synthesizes a material with the structure thus determined and measures its physical properties. By repeating this loop of synthesis, measurement, and consideration, the human gradually gains knowledge and can ultimately discover a material with suitable physical properties.

[0003] However, it is difficult for humans to search a vast search space quickly when searching for materials. Therefore, autonomous materials search, in which computers autonomously search for materials, has attracted attention.

[0004] Autonomous material exploration can be broadly divided into two types. The first type of autonomous material exploration is robotic exploration. In robotic exploration, an automatically moving robot synthesizes materials and measures their physical properties. Then, a computer uses machine learning to analyze the relationship between the structure of the synthesized material and its physical properties, and determines the structure of the next material to be synthesized based on the analysis results. For example, in Patent Document 1, a computer determines the structure of the next material to be synthesized and the synthesis process conditions by Bayesian optimization, which is one of the machine learning methods. Robotic exploration makes it possible for a robot to autonomously discover new materials with suitable physical properties without human intervention.

[0005] The second type of autonomous material search is simulation search, which uses a simulation instead of a robot. For example, in Non-Patent Document 1, a computer calculates the physical property values ​​of a material having a certain structure by a simulation called first-principles calculation. Then, the computer analyzes the calculated physical property values ​​and structure by Bayesian optimization to determine the structure of the material to be simulated next. As with robot search, simulation search also makes it possible for a computer to autonomously discover new materials with suitable physical property values ​​without human intervention.

[0006] However, there is room for further improvement in autonomous material exploration in terms of shortening the material exploration time. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Patent Publication No. 2021-43959 [Non-patent literature]

[0008] [Non-Patent Document 1] Iwasaki et al., “Machine learning autonomous identification of magnetic alloys beyond the Slater-Pauling limit”, Communications Materials, vol. 2, page 31, 2021 Summary of the Invention [Problem to be solved by the invention]

[0009] In one aspect, the present invention aims to reduce the search time for materials. [Means for solving the problem]

[0010] According to one aspect, the material exploration program includes: generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of the material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; The method distributes and executes a process including: performing a first transfer learning on the first learning model so as to output the second physical property value; selecting a learning model having characteristics that satisfy a criterion from a plurality of learning models including the first learning model on which the first transfer learning was performed and the second learning model; determining a structure from which an evaluation value of the second physical property value is to be obtained based on the selected learning model; obtaining the evaluation value for the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data, among a plurality of computers interconnected via a network.

[0011] In the above-mentioned material exploration program, the first machine learning may be machine learning using an ensemble neural network.

[0012] In the above-mentioned material search program, the evaluation value may be a calculated value of the second physical property value, and the process may include calculating the calculated value.

[0013] In the above-mentioned material exploration program, obtaining the evaluation value may include notifying one of the plurality of computers of the determined structure and an instruction to calculate the calculated value obtained from the structure, and obtaining the calculated value from the one of the computers.

[0014] In the above-mentioned material search program, the evaluation value may be an experimental value of the second physical property value, and the process may include an instruction to cause a robot to measure the experimental value.

[0015] In the above-mentioned material exploration program, the characteristic may be a generalization performance of each of the plurality of learning models.

[0016] In the above material exploration program, generating the first learning model may be executed by a first computer among the multiple computers, generating the second learning model may be executed by a second computer among the multiple computers, and the first computer may inquire of the second computer as to whether the second learning model has been generated, and when the second computer responds to the inquiry that the second learning model has been generated, selecting the learning model.

[0017] In the above material exploration program, generating the first learning model, generating the second learning model, performing the first transfer learning, selecting the learning model, determining the structure, obtaining the evaluation value, and adding the material descriptor and the evaluation value to the second training data may be repeated.

[0018] In the above material exploration program, different processes among generating the first learning model, generating the second learning model, performing the first transfer learning, selecting the learning model, determining the structure, obtaining the evaluation value, and adding the material descriptor and the evaluation value to the second training data may be executed by different computers among the multiple computers.

[0019] In the above-mentioned material exploration program, the processing may further include: generating a third learning model that outputs the third physical property value corresponding to the structure by a third machine learning using a plurality of third material descriptors corresponding to each of the plurality of structures and a third physical property value of the material for each of the third material descriptors as third training data; generating a fourth learning model that outputs the first physical property value and the third physical property value corresponding to the structure by a fourth machine learning using the first training data and the third training data; performing a second transfer learning on the third learning model using the second training data to output the second physical property value corresponding to the structure; and performing a third transfer learning on the fourth learning model using the second training data to output the second physical property value corresponding to the structure, and the plurality of learning models may include the third learning model on which the second transfer learning has been performed and the fourth learning model on which the third transfer learning has been performed.

[0020] In the above-mentioned material exploration program, determining the target structure may include calculating a predicted value and variance of the second physical property value using the selected learning model, generating an acquisition function based on the predicted value and the variance, and determining as the target structure a structure for which the acquisition function is maximum or minimum.

[0021] According to one aspect, a material exploration method includes: generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; A process including: performing a first transfer learning on the first learning model so as to output the second physical property value; selecting a learning model having characteristics that satisfy a criterion from a plurality of learning models including the first learning model on which the first transfer learning was performed and the second learning model; determining a structure from which an evaluation value of the second physical property value is to be obtained based on the selected learning model; obtaining the evaluation value for the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data.

[0022] According to one aspect, the material exploration system generates a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of the material and a first physical property value of the material for each of the first material descriptors as first training data, generates a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data, and generates a second learning model that outputs the second physical property values ​​corresponding to the structures by using the second training data. A process including: performing a first transfer learning on the first learning model so as to output the second physical property value; selecting a learning model having characteristics that satisfy a criterion from a plurality of learning models including the first learning model on which the first transfer learning was performed and the second learning model; determining a structure from which an evaluation value of the second physical property value is to be obtained based on the selected learning model; obtaining the evaluation value for the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data.

[0023] According to one aspect, the material exploration apparatus generates a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data, generates a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data, and generates a second learning model that outputs the second physical property values ​​corresponding to the structures by using the second training data. The method distributes and executes a process including: performing a first transfer learning on the first learning model so as to output a second physical property value; selecting a learning model having characteristics that satisfy a criterion from a plurality of learning models including the first learning model on which the first transfer learning was performed and the second learning model; determining a structure from which an evaluation value of the second physical property value is to be obtained based on the selected learning model; obtaining the evaluation value for the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data, among a plurality of computers interconnected via a network. Effect of the Invention

[0024] According to the present invention, the time required for searching for materials can be shortened. [Brief description of the drawings]

[0025] [Figure 1] FIG. 1 is a system configuration diagram of a material exploration system according to the first embodiment. [Diagram 2] FIG. 2 is a functional configuration diagram of each computer according to the first embodiment. [Diagram 3] FIG. 3 is a schematic diagram of training data according to the first embodiment. [Figure 4] FIG. 4 is a schematic diagram of a learning model consisting of an ensemble neural network according to the first embodiment. [Diagram 5]FIG. 5 is a schematic diagram of a neural network according to the first embodiment. [Figure 6] FIG. 6 is a schematic diagram of a materials exploration system realized by two computers according to the first embodiment. [Figure 7] FIG. 7 is a flowchart showing the processing performed by each computer in the first embodiment. [Figure 8] FIG. 8 is a schematic diagram of a material exploration system according to the second embodiment. [Figure 9] FIG. 9 is a schematic diagram of a neural network constituting an ensemble neural network of the fourth learning model according to the second embodiment. [Figure 10] FIG. 10 is a flowchart (part 1) showing the processing performed by each computer in the second embodiment. [Figure 11] FIG. 11 is a flowchart (part 2) showing the processing performed by each computer in the second embodiment. [Figure 12] FIG. 12 is a diagram obtained by investigating how much the search time is reduced in the second embodiment. [Figure 13] FIG. 13 is a system configuration diagram of a material exploration system according to the third embodiment. [Figure 14] FIG. 14 is a flowchart showing the processing performed by each computer in the third embodiment. [Figure 15] FIG. 15 is a system configuration diagram of a material exploration system according to the fourth embodiment. [Figure 16] FIG. 16 is a functional configuration diagram of a material exploration apparatus according to the fourth embodiment. [Figure 17] FIG. 2 is a diagram showing the hardware configuration of each of the computers according to the first to third embodiments and the material exploration apparatus according to the fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Note that like elements are given like reference numerals and their description will be omitted.

[0027] (First embodiment) In this embodiment, a material having desired physical property values ​​is searched for while utilizing the computational resources scattered at each base as follows.

[0028] Fig. 1 is a system configuration diagram of a material exploration system according to this embodiment. As shown in Fig. 1, a material exploration system 10 according to this embodiment includes first to n-th computers 11a to 11n connected to each other via a network 12. Each of the computers 11a to 11n is a computer such as a PC (Personal Computer), a server, or a parallel computer for performing first-principles calculations or machine learning.

[0029] In this example, it is assumed that the target physical property values ​​are different for each of the computers 11a to 11n. The target physical property values ​​are the physical property values ​​to be predicted by machine learning and the physical property values ​​to be calculated by first-principles calculation. For example, when the target physical property value of the computer 11a is the spin polarizability, the target physical property value of the computer 11b is the Curie temperature.

[0030] In the following, it is assumed that at least two of the computers 11a to 11n are managed by different administrators. The locations of the computers 11a to 11n are not particularly limited. All of the computers 11a to 11n may be installed in Japan or outside Japan. Some of the computers 11a to 11n may be installed in Japan, and the remaining computers may be installed outside Japan.

[0031] The network 12 is, for example, the Internet, a LAN (Local Area Network), or a public telephone network.

[0032] 2 is a functional block diagram of each of the computers 11a to 11n according to this embodiment. As shown in FIG. 2, each of the computers 11a to 11n includes a communication unit 14, a control unit 15, and a storage unit 16.

[0033] The communication unit 14 is an interface for connecting each of the computers 11a to 11n to the network 12.

[0034] The memory unit 16 is a processing unit that stores training data 27, a learning model 28, physical property values ​​29, and structure data 30.

[0035] The training data 27 is data for learning the relationship between the structure and physical properties of a certain material by machine learning.

[0036] FIG. 3 is a schematic diagram of the training data 27. As shown in FIG. 3, the training data 27 includes a plurality of pairs 33 of material descriptors 32 and physical property values ​​29. The material descriptors 32 are explanatory variables input to a machine learning learning model, and are data correlated with the structure of a certain material. For example, the average atomic weight of a material in a certain structure may be adopted as the material descriptor 32. The material descriptor 32 may also be expressed in the form of a character string or a graph. As an example, the machine learning unit 22 generates the material descriptor 32 from the structure data 30.

[0037] In the example of Fig. 3, it is assumed that r material descriptors 32 are generated from r pieces of structure data 30. The dimensions of each material descriptor 32 are not particularly limited, but here, n parameters x1, x2, ...x n Assume an n-dimensional material descriptor 32 with elements x1, x2, ... x n The superscripted symbols (1), (2), ... (r) attached to each parameter in the above are symbols corresponding to the r pieces of original structure data 30. In the following, these superscripted symbols will be omitted for the sake of simplicity.

[0038] The structure data 30 is data that indicates the structure of a certain material. When the material is a periodic system, the structure data 30 is data that associates the atomic species contained in the simulation cell with their positions, as shown in the figure. According to this format, the number of atomic species contained in the simulation cell can also be specified, so the composition of the material can also be specified from the structure data 30.

[0039] The format of the structure data 30 is not limited to this, and any format that is compatible with the calculation code of the first-principles calculation may be adopted. For example, when the material is a molecule, a Z matrix may be adopted as the structure data 30.

[0040] The physical property values ​​29 are calculated from the structural data 30. Examples of such physical property values ​​include spin polarizability, Curie temperature, magnetic moment, band gap, cohesive energy, Gibbs free energy, Helmholtz free energy, and enthalpy.

[0041] In each set 33 of the training data 27, a physical property value 29 calculated from certain structure data 30 and a material descriptor 32 generated from that structure data 30 are associated with each other.

[0042] As described above, since the target physical property value differs for each of the computers 11a to 11n, the physical property value 29 in one computer differs from the physical property value 29 in another computer. As a result, the computers 11a to 11n learn different types of physical property values ​​29.

[0043] Referring again to Fig. 2, the learning model 28 is a model that, when a material descriptor 32 is input, outputs a physical property value 29 corresponding to the structure indicated by the structural data 30 from which the material descriptor 32 is generated. In this embodiment, an ensemble neural network is adopted as the learning model 28.

[0044] 4 is a schematic diagram of a learning model 28 consisting of an ensemble neural network. As shown in FIG 4, the learning model 28 includes a plurality of neural networks 35, and outputs the average value of the outputs of the neural networks 35.

[0045] Fig. 5 is a schematic diagram of each neural network 35. As shown in Fig. 5, the neural network 35 calculates the explanatory variables (material descriptors) x1, x2, ... x nThe model includes an input layer 36, a hidden layer 37, and an output layer 38. As described above, the model outputs a response variable y corresponding to n explanatory variables (material descriptors) x1, x2, ...x n is data generated from one piece of structure data 30 (see FIG. 3). The objective variable y is a physical property value 29. As described above, the computers 11a to 11n learn different types of physical property values ​​29, and therefore the type of the objective variable y differs for each of the computers 11a to 11n.

[0046] The input layer 36 is the explanatory variables x1, x2, ... x n is input to the hidden layer 37. The hidden layer 37 is a layer including weight coefficients adjusted by learning, and outputs values ​​according to the weight coefficients. In this example, the hidden layer 37 has three layers, but the number of layers of the hidden layer 37 is not particularly limited. However, in this embodiment, transfer learning is performed to update the weight coefficients of the hidden layer 37 as described below, so it is preferable that at least one hidden layer 37 exists so that transfer learning is possible.

[0047] The output layer 38 is a layer that outputs the objective variable y based on the output of the last hidden layer 37 .

[0048] Referring again to Fig. 2, the physical property values ​​29 are the same data as those included in the training data 27 (see Fig. 3). Since the physical property values ​​29 are also included in the training data 27, they do not need to be stored in the storage unit 16. Furthermore, the structure data 30 is data used to generate the material descriptor 32.

[0049] The control unit 15 is a processing unit that controls each of the computers 11a to 11n. As an example, the control unit 15 includes a calculation unit 21, a machine learning unit 22, a transfer learning unit 23, a model selection unit 24, a structure determination unit 25, and an evaluation value acquisition unit 26.

[0050] The calculation unit 21 is a processing unit that calculates the physical property values ​​of the material of the structure indicated by the structure data 30 by first-principles calculation. The first-principles calculation may be an electronic state calculation or a phonon calculation. Examples of the electronic state calculation method include the density functional method, the GW method, the Hartree-Fock method, the MP (Meller-Plesset) method, the CI (Configuration Interaction) method, the MCSCF (Multi-Configurational Self-Consistent Field) method, the CC (Coupled Cluster) method, and the quantum Monte Carlo method. Furthermore, a molecular dynamics method may be adopted instead of the first-principles calculation.

[0051] As described above, the target physical property value of the first-principles calculation is different for each of the computers 11a to 11n. The administrator of each of the computers 11a to 11n describes the calculation conditions in an input file (not shown) such that the target physical property value is calculated with a predetermined calculation accuracy, and stores the input file in the storage unit 16 of each of the computers 11a to 11n. Then, the calculation unit 21 of each of the computers 11a to 11n performs calculation according to the calculation conditions in the input file stored in the computer. When the target material is a periodic system, such calculation conditions include, for example, the roughness of the k-mesh, the cutoff energy of the plane wave basis, and the type of pseudopotential. When the target material is a molecule, the size and type of the basis set are also examples of the calculation conditions.

[0052] The machine learning unit 22 is a processing unit that generates a learning model 28 by machine learning using training data 27 of its own device. As an example, the machine learning unit 22 generates material descriptors 32 from structure data 30 stored in the storage unit 16. Then, the machine learning unit 22 generates a plurality of sets 33 by associating the generated material descriptors 32 with corresponding physical property values ​​29, thereby generating training data 27 including the plurality of sets 33.

[0053] Furthermore, the machine learning unit 22 generates a learning model 28 that outputs a physical property value 29 corresponding to the structure indicated by the structural data 30 from which the material descriptor 32 was generated.

[0054] The transfer learning unit 23 is a processing unit that performs transfer learning on the learning model 28 generated by the machine learning unit 22, using training data 27 stored in a computer different from the own device. For example, consider a case where the objective variable of the learning model 28 of the computer 11a is the spin polarizability, and the physical property value 29 of the learning model 28 of the computer 11b is the Curie temperature. In this case, the transfer learning unit 23 of the computer 11a performs transfer learning on the learning model 28 of the own device, by using the training data 27 of the computer 11b, so as to output the Curie temperature as the objective variable.

[0055] In transfer learning, transfer learning unit 23 fixes the weight coefficients of two layers on the input layer 36 side among multiple hidden layers 37 of neural network 35 (see FIG. 5), and updates the weight coefficient of one layer on the output layer 38 side. As a result, learning model 28 becomes a model that outputs physical property values ​​29 of training data 27 of a computer different from its own device while reflecting the trend of physical property values ​​before transfer learning.

[0056] The model selection unit 24 is a processing unit that selects a learning model having characteristics that satisfy a criterion from among a plurality of learning models including the learning model 28 that has been subjected to transfer learning by the transfer learning unit 23. The learning models 28 that the model selection unit 24 of a certain computer selects as candidates for selection are the learning model 28 of its own device and the learning models 28 of other computers.

[0057] Moreover, the characteristics that are the criterion for selection include, for example, the generalization performance of the learning model 28. As an example, the model selection unit 24 inputs a material descriptor 32 that does not exist in the training data 27 used for learning to the learning model 28 as test data, and thereby obtains the physical property value output by the learning model 28. The test data is, for example, stored in advance by the user in the storage unit 16. Then, the model selection unit 24 calculates the generalization error, which is the difference between the physical property value and the correct answer value, as the generalization performance for each of the multiple structure data. Note that the correct answer data is, for example, an experimental value or a calculated value by first-principles calculation. And, as the criterion for selection, for example, the criterion that the generalization performance is the highest among the multiple learning models 28.

[0058] The structure determination unit 25 is a processing unit that determines a structure from which an evaluation value of a physical property value is to be obtained. The physical property value is a physical property value learned by the learning model 28 through transfer learning. For example, when transfer learning is performed on the learning model 28 so as to output the Curie temperature as described above, the Curie temperature becomes the physical property value. The evaluation value of the physical property value is an experimental value of the physical property value or a calculated value of the physical property value by first-principles calculation. As an example, the structure determination unit 25 determines the structure by Bayesian optimization as follows.

[0059] First, the structure determination unit 25 acquires the predicted value and variance of the physical property value 29 by inputting the material descriptor 32 generated from the multiple structure data 30 stored in the storage unit 16 to the learning model 28 selected by the model selection unit 24. Then, the structure determination unit 25 generates a UCB (Upper Confidence Bond) function as an acquisition function based on the acquired predicted value and variance of the physical property value 29. Note that the acquisition function is not limited to UCB, and EI (Expected Improvement) or PI (Probability Improvement) may be adopted as the acquisition function.

[0060] The structure determination unit 25 identifies the structure data 30 for which the UCB function is maximum or minimum, and determines the structure indicated by the structure data 30 as the structure to be input next to the calculation unit 21. The value of the UCB function that is the criterion for determining the structure depends on the type of physical property value. For example, if one wishes to find a material with as high a physical property value as possible, one identifies the structure data 30 for which the UCB function is maximum. On the other hand, if one wishes to find a material with as low a physical property value as possible, one identifies the structure data 30 for which the UCB function is minimum.

[0061] In this embodiment, an ensemble neural network (see FIG. 4) is adopted as the learning model 28. The ensemble neural network has the advantage that both Bayesian optimization by the structure determination unit 25 and transfer learning by the transfer learning unit 23 are possible.

[0062] The evaluation value acquisition unit 26 is a processing unit that acquires an evaluation value of a physical property value corresponding to the structure determined by the structure determination unit 25. For example, the evaluation value acquisition unit 26 acquires, as an evaluation value, a calculation value calculated by the calculation unit 21 of a computer different from its own device through first-principles calculation or an experimental value obtained through an experiment by an experimental robot (not shown) connected to the network 12.

[0063] The evaluation value acquisition unit 26 also determines whether to end the material search. For example, the evaluation value acquisition unit 26 determines to end the material search when the number of times the evaluation value has been acquired exceeds a preset threshold, and determines not to end the material search when the number of times the evaluation value has been acquired does not exceed a preset threshold. Alternatively, the evaluation value acquisition unit 26 may determine to end the material search when the evaluation value is within an acceptable range preset by the user, and may determine not to end the material search when the evaluation value is outside the acceptable range. For example, when a material with a higher Curie temperature is to be searched for, the acceptable range is a range equal to or greater than a threshold preset by the user for the Curie temperature.

[0064] Furthermore, when it is determined not to end the material search, the evaluation value acquisition unit 26 adds the material descriptor 32 generated from the structure data 30 of the structure determined by the structure determination unit 25 and the evaluation value obtained from the structure to the training data 27. The training data 27 to which the material descriptor 32 is added may be training data 27 in the same computer as the evaluation value acquisition unit 26, or may be training data 27 used by the transfer learning unit 23 for transfer learning.

[0065] The material exploration system 10 according to this embodiment utilizes the resources of each of the computers 11a to 11n by distributing the processing of each of the steps of the material exploration program among the computers 11a to 11n (see FIG. 1). The number of the computers 11a to 11n is not limited, but the following describes a case where the material exploration method is performed by two computers, the first computer 11a and the second computer 11b, among the multiple computers 11a to 11n.

[0066] Fig. 6 is a schematic diagram of a materials exploration system 10 realized by two computers 11a and 11b. In the following, the training data and the learning model of the first computer 11a are respectively referred to as first training data 27 and first learning model 28. Moreover, the training data and the learning model of the second computer 11b are respectively referred to as second training data 27 and second learning model 28.

[0067] In this example, it is assumed that the target physical property values ​​of the first calculator 11a and the second calculator 11b are different, and the target physical property value of the first calculator 11a is the spin polarizability, and the target physical property value of the second calculator 11b is the Curie temperature.

[0068] In this case, the objective variable of the first learning model 28 is the spin polarizability, and the physical property value 29 (see FIG. 3) of the first training data 27 is also the spin polarizability. On the other hand, the objective variable of the second learning model 28 is the Curie temperature, and the physical property value 29 (see FIG. 3) of the second training data 28 is also the Curie temperature.

[0069] In this embodiment, a material search is performed as follows by causing the first learning model 28, which has learned the spin polarizability, to learn the Curie temperature by transfer learning using the second training data 28.

[0070] FIG. 7 is a flow chart showing the processing performed by each of the computers 11a and 11b in FIG. 6 in this embodiment.

[0071] This process is executed, for example, when a user issues a start instruction to a material search program running in each of the computers 11a and 11b.

[0072] First, the machine learning unit 22 of the first computer 11a generates a first learning model 28 with the spin polarizability as the objective variable (step S1). For example, the machine learning unit 22 of the first computer 11a generates the first learning model 28 by machine learning using the first training data 27 in which the physical property value 29 (see FIG. 3) is the spin polarizability. The material descriptor 32 (see FIG. 3) included in the first training data 27 is generated in advance by the machine learning unit 22 of the first computer 11a. The material descriptor 32 and the physical property value 29 included in the first training data 27 are examples of the first material descriptor and the first physical property value, respectively. The machine learning for generating the first learning model 28 is an example of the first machine learning.

[0073] Next, the machine learning unit 22 of the second computer 11b generates a second learning model 28 with the Curie temperature as the objective variable (step S3). For example, the machine learning unit 22 of the second computer 11b generates the second learning model 28 by machine learning using the second training data 27 in which the physical property value 29 (see FIG. 3) is the Curie temperature. The material descriptor 32 (see FIG. 3) included in the second training data 27 is generated in advance by the machine learning unit 22 of the second computer 11b. The material descriptor 32 and the physical property value 29 included in the second training data 27 are examples of the second material descriptor and the second physical property value, respectively. The machine learning for generating the second learning model 28 is an example of the second machine learning.

[0074] Furthermore, the timing for executing step S3 is not limited to after execution of step S1, as long as step S3 is completed by the time step S7 described below is executed.

[0075] Next, the transfer learning unit 23 of the first computer 11a performs the first transfer learning so as to output the Curie temperature as a response variable for the first learning model 28 of the own device (step S5). For example, the transfer learning unit 23 of the first computer 11a requests the second training data 27 of the Curie temperature from the control unit 15 of the second computer 11b, and acquires the second training data 27 in response to the request. Then, the transfer learning unit 23 of the first computer 11a performs the transfer learning using the acquired second training data 27 so as to output the Curie temperature as a response variable for the first learning model 28 of the own device. As described above, in the transfer learning, among the multiple hidden layers 37 of the neural network 35 (see FIG. 5), the weight coefficients of two layers on the input layer 36 side are fixed, and the weight coefficient of one layer on the output layer 38 side is updated.

[0076] Next, the model selection unit 24 of the first computer 11a selects a learning model whose characteristics satisfy the criteria from among the first learning model 28 and the second learning model 28 that have undergone transfer learning (step S7). For example, the model selection unit 24 of the first computer 11a acquires the second learning model 28 from the second computer 11b via the network 12. Then, the model selection unit 24 of the first computer 11a calculates the generalization performance for each Curie temperature of the second learning model 28 and the first learning model 28 that has undergone transfer learning, using test data for the Curie temperature stored by the user in the storage unit 16 of the first computer 11a. After that, the model selection unit 24 of the first computer 11a selects the model with the highest generalization performance from among the second learning model 28 and the first learning model 28 that has undergone transfer learning.

[0077] In addition, the model selection unit 24 of the second computer 11b may obtain test data from the first computer 11a, and use the test data to calculate the generalization performance of the second learning model 28 for the Curie temperature. In this case, the model selection unit 24 of the first computer 11a may obtain the generalization performance from the second computer 11b. This can save the time required for the model selection unit 24 of the first computer 11a to obtain the second learning model 28. Furthermore, since there is no need to transmit the second learning model 28 to the network 12, it is also possible to suppress pressure on the bandwidth of the network 12.

[0078] Also, if the second learning model 28 has not yet been generated at the end of step S5, the model selection unit 24 of the first computer 11a may wait without executing step S7. For example, the model selection unit 24 of the first computer 11a periodically notifies the second computer 11b of an inquiry as to whether the second learning model 28 has been generated. Then, when the second computer 11b responds to the inquiry that the second learning model 28 has not been generated, the model selection unit 24 of the first computer 11a waits without executing step S7. On the other hand, when the second computer 11b responds that the second learning model 28 has been generated, the model selection unit 24 of the first computer 11a executes step S7. This allows the processes of the first computer 11a and the second computer 11b to be synchronized.

[0079] Next, the structure determination unit 25 of the first computer 11a determines a target structure for obtaining an evaluation value of the Curie temperature (step S9). For example, the structure determination unit 25 generates a plurality of material descriptors 32 from a plurality of structure data 30 in the storage unit 16 of the device itself. Then, the structure determination unit 25 inputs the generated plurality of material descriptors 32 to the learning model 28 selected in step S7, and obtains the predicted value and variance of the physical property value 29 output from the learning model 28 for the input. Then, the structure determination unit 25 generates a UCB function based on the predicted value and variance of the physical property value 29, and identifies the structure data 30 for which the UCB function is maximum. The structure indicated by the structure data 30 for which the UCB function is maximum in this way is a structure dissimilar to the structure present in the training data 27, and is a structure expected to have a Curie temperature as high as possible, and is determined as a target for obtaining an evaluation value. Note that, as described above, EI or PI may be adopted as the acquisition function instead of the UCB function.

[0080] In addition, since the acquisition function is generated based on the learning model 28 selected in step S7 as the model with the highest generalization performance and prediction accuracy, it is possible to increase the possibility of determining a structure among multiple structures that has the highest possible Curie temperature.

[0081] Next, the calculation unit 21 of the second computer 11b performs a first-principles calculation based on the structure determined by the structure determination unit 25, and calculates the Curie temperature corresponding to the structure as an evaluation value (step S11). For example, the structure determination unit 25 of the first computer 11a notifies the calculation unit 21 of the second computer 11b of the structure determined in step S9 by including the structure in an instruction to calculate the Curie temperature obtained from the structure. Then, the calculation unit 21 of the second computer 11b, which has received the notification, calculates the calculated value of the Curie temperature obtained from the notified structure by first-principles calculation. At this time, the calculation unit 21 performs the calculation according to calculation conditions previously set by the administrator of the second computer 11b so that the Curie temperature is calculated with a predetermined calculation accuracy.

[0082] First-principles calculations can obtain evaluation values ​​more quickly than experiments, thereby shortening the time required for material search.

[0083] Moreover, since an input file that defines the calculation conditions for calculating the Curie temperature with a predetermined calculation accuracy is stored in advance in the storage unit 16 of the second calculator 11b, the administrator does not need to reset the calculation conditions. If there is no problem in resetting the calculation conditions, the calculation unit 21 of the first calculator 11a may calculate the evaluation value instead of the second calculator 11b.

[0084] The structure used in the first-principles calculation is the structure indicated by the structural data 30 for which the acquired function such as the UCB function is maximized in step S7. Therefore, the evaluation value obtained by the first-principles calculation in step S11 is expected to be a Curie temperature as large as possible for a structure dissimilar to the structure included in the training data 27.

[0085] Next, the evaluation value acquisition unit 26 of the first calculator 11a acquires the evaluation value calculated by the calculation unit 21 of the second calculator 11b (step S13).

[0086] Next, the evaluation value acquisition unit 26 of the first computer 11a determines whether to end the material search (step S15). For example, the evaluation value acquisition unit 26 determines to end the material search (YES) when the number of times step S13 for acquiring the evaluation value is executed exceeds a preset threshold, and determines not to end the material search (NO) when the number of times is not exceeded. Alternatively, the evaluation value acquisition unit 26 may determine to end the material search (YES) when the evaluation value is in a range equal to or greater than a threshold value preset by the user for the Curie temperature, and may determine not to end the material search (NO) when the evaluation value is in a range less than the threshold.

[0087] Here, if it is determined that the material search is not to be ended (NO), the process proceeds to step S17, where the evaluation value acquisition unit 26 of the first computer 11a updates the second training data 27 of the second computer 11b. For example, the evaluation value acquisition unit 26 of the first computer 11a generates a material descriptor 32 corresponding to the structure determined by the structure determination unit 25. Then, the evaluation value acquisition unit 26 adds the generated material descriptor 32 and the evaluation value acquired in step S13 to the second training data 27 of the second computer 11b. After that, the process returns to step S1 again.

[0088] On the other hand, if it is determined that the material search is to be ended (YES), the process ends.

[0089] According to the present embodiment described above, the first transfer learning (step S5) and the updating of the training data (step S17) are repeatedly executed until it is determined in step S15 that the material search is to be ended (YES).

[0090] Among these, in the first transfer learning (step S5), the first learning model 28 that has already learned the spin polarizability is made to learn the Curie temperature. The spin polarizability and the Curie temperature are not independent physical property values, but have some correlation invisible to humans. Therefore, if the first learning model 28 that has learned the trend of the spin polarizability is made to learn the Curie temperature by the transfer learning, the prediction accuracy of the first learning model 28 when predicting the Curie temperature is likely to be improved due to the trend of the Curie temperature inherent in the trend of the spin polarizability. As a result, since the possibility of determining a structure that exhibits a high Curie temperature in step S9 increases, even if the number of times steps S1 to S17 are repeatedly executed is reduced, a material with a high Curie temperature exceeding the standard set by the user can be found, and the search time required for material search can be shortened. This provides the technical effect that a new material having excellent physical properties such as a high Curie temperature can be found in a short time, and a new device equipped with the new material can be developed in a short time.

[0091] Furthermore, in step S17, material descriptors 32 corresponding to structures expected to have high Curie temperatures and evaluation values ​​of the Curie temperatures obtained from the structures are added to the second training data 27. As a result, material descriptors 32 that provide high Curie temperatures are accumulated in the second training data 27, and it becomes possible to discover structures that exhibit high Curie temperatures by using the second learning model 28 generated from the second training data 27.

[0092] Moreover, the first transfer learning (step S5) for the learning model 28 is performed by the transfer learning unit 23 of the first computer 11a using the second training data 27 of the second computer 11b. This makes it possible to utilize the knowledge held by each of the computers 11a and 11b, such as the learning model 28 and the training data 27, and the computing capabilities of each of the computers 11a and 11b, in the autonomous material search.

[0093] The execution entity of each of the above steps S1 to S17 is merely an example, and each step may be executed by each of the computers 11a and 11b in a distributed manner. For example, different processes among the plurality of steps S1 to S17 may be executed by the first computer 11a and the second computer 11b, respectively. This allows a material search to be performed while effectively utilizing the computational resources of each of the computers 11a and 11b in the same manner as described above. The execution entity of each of the steps S1 to S17 may be set in advance by the user in the material search program, for example.

[0094] Second embodiment In the first embodiment, an example in which a material search is performed using two computers out of the multiple computers 11a to 11n has been described, whereas in the present embodiment, an example in which a material search is performed using three computers will be described.

[0095] Fig. 8 is a schematic diagram of a material exploration system 10 according to this embodiment. In Fig. 8, the same elements as those described in the first embodiment are denoted by the same reference numerals as those in the first embodiment, and their description will be omitted below.

[0096] As shown in FIG. 8, a material exploration system 10 according to this embodiment includes first to third computers 11a to 11c.

[0097] The storage unit 16 of the first computer 11a includes, in addition to the first learning model 28 described in the first embodiment, a fourth learning model 28 in which the spin polarizability and the magnetic moment are the objective variables. The fourth learning model 28 is an ensemble neural network composed of a plurality of neural networks 28 as shown in FIG.

[0098] FIG. 9 is a schematic diagram of a neural network 35 that constitutes the ensemble neural network of the fourth learning model 28.

[0099] As shown in FIG. 9, this neural network 35 has explanatory variables x1, x2, ... x n In addition to the input layer 36 and hidden layer 37 to which are input, there are two output layers 38. One of the output layers 38 outputs the spin polarization as the objective variable y1, and the other outputs the magnetic moment as the objective variable y2.

[0100] Referring again to Fig. 8, the target physical property value of the second calculator 11b is the Curie temperature, similarly to the first embodiment.

[0101] On the other hand, the target physical property value of the third computer 11c is the magnetic moment. Hereinafter, the training data and the learning model of the third computer 11c are respectively referred to as the third training data 27 and the third learning model 28. In this case, the objective variable of the third learning model 28 is the magnetic moment, and the evaluation value 29 (see FIG. 3) of the third training data 27 is the magnetic moment.

[0102] 10 and 11 are flow charts showing the processing performed by each of the computers 11a to 11c in this embodiment.

[0103] This process is executed, for example, when a user issues a start instruction to a material search program running in each of the computers 11a to 11c.

[0104] First, similarly to the first embodiment, the machine learning unit 22 of the first computer 11a generates a first learning model 28 in which the spin polarizability is an objective variable (step S1).

[0105] Next, similarly to the first embodiment, the machine learning unit 22 of the second computer 11b generates a second learning model 28 with the Curie temperature as the objective variable (step S3).

[0106] Next, the machine learning unit 22 of the third computer 11c generates a third learning model 28 with the magnetic moment as the objective variable (step S23). For example, the machine learning unit 22 of the third computer 11c generates the third learning model 28 by machine learning using the third training data 27 in which the physical property value 29 (see FIG. 3) is the magnetic moment. The material descriptor 32 (see FIG. 3) included in the third training data 27 is generated in advance by the machine learning unit 22 of the third computer 11c. The material descriptor 32 and the physical property value 29 included in the third training data 27 are examples of the third material descriptor and the third physical property value, respectively. The machine learning for generating the third learning model 28 is an example of the third machine learning.

[0107] The order of execution of steps S1, S3, and S23 is arbitrary and is not limited to the above.

[0108] Next, the machine learning unit 22 of the first computer 11a generates a fourth learning model 28 with the spin polarizability and the magnetic moment as objective variables (step S25). For example, the machine learning unit 22 of the first computer 11a acquires the third training data 27 of the magnetic moment from the third computer 11c. Then, the machine learning unit 22 of the first computer 11a generates the fourth learning model 28 by machine learning using the acquired third training data 27 and the first training data 27 of the spin polarizability stored in the memory unit 16 of the own device, and stores it in the memory unit 16 of the own device. Note that the machine learning for generating the fourth learning model 28 is an example of the fourth machine learning.

[0109] Next, in the same manner as in the first embodiment, the transfer learning unit 23 of the first computer 11a performs transfer learning on the first learning model 28 so as to output the Curie temperature as the objective variable (step S5).

[0110] Next, the transfer learning unit 23 of the third computer 11c performs second transfer learning so as to output the Curie temperature as a response variable for the third learning model 28 of the own device (step S27). For example, the transfer learning unit 23 of the third computer 11c requests the second training data 27 of the Curie temperature from the control unit 15 of the second computer 11b, and acquires the second training data 27 in response to the request. Then, the transfer learning unit 23 of the third computer 11c performs transfer learning using the acquired second training data 27 so as to output the Curie temperature as a response variable for the third learning model 28 of the own device.

[0111] Next, the transfer learning unit 23 of the first computer 11a performs third transfer learning so as to output the Curie temperature as a response variable for the fourth learning model 28 of the own device (step S29). For example, the transfer learning unit 23 of the first computer 11a requests the second training data 27 of the Curie temperature from the control unit 15 of the second computer 11b, and acquires the second training data 27 in response to the request. Then, the transfer learning unit 23 of the first computer 11a performs transfer learning using the acquired second training data 27 so as to output the Curie temperature as a response variable for the fourth learning model 28 of the own device.

[0112] The order of execution of steps S5, S25, S27, and S29 is arbitrary and is not limited to the above.

[0113] Next, the model selection unit 24 of the first computer 11a selects a learning model whose characteristics satisfy the criteria from among the multiple learning models 28 (step S7). The learning models that are candidates for selection are the second learning model 28, and the first learning model 28, the third learning model 28, and the fourth learning model 28 on which the first to third transfer learnings have been performed.

[0114] For example, the model selection unit 24 of the first computer 11a acquires the second learning model 28 from the second computer 11b, and acquires the third learning model 28 that has undergone transfer learning from the third computer 11c. Then, the model selection unit 24 of the first computer 11a calculates the generalization performance for each Curie temperature of the first to fourth learning models 28 using test data for the Curie temperature stored in the first computer 11a, and selects the model with the highest generalization performance.

[0115] In addition, the model selection unit 24 of the second computer 11b may calculate the generalization performance for the Curie temperature of the second learning model 28, and the model selection unit 24 of the third computer 11c may calculate the generalization performance for the Curie temperature of the third learning model 28. In this case, the test data of the Curie temperature required for calculating the generalization performance may be acquired by each of the computers 11b and 11c from the first computer 11a. Then, the model selection unit 24 of the first computer 11a may acquire the calculated generalization performance from each of the computers 11b and 11c. This can save the time required for the model selection unit 24 of the first computer 11a to acquire the second learning model 28 and the third learning model 28. Furthermore, it is also possible to suppress the bandwidth of the network 12 from being compressed by the second learning model 28 and the third learning model 28.

[0116] Next, the structure determination unit 25 of the first computer 11a determines a structure from which an evaluation value of the Curie temperature is to be obtained by using the learning model 28 selected in step S7 in the same manner as in the first embodiment (step S9).

[0117] Next, the calculation unit 21 of the second computer 11b performs a first-principles calculation based on the structure determined by the structure determination unit 25, and calculates the Curie temperature corresponding to the structure as an evaluation value (step S11). Here, similar to the first embodiment, the structure determination unit 25 of the first computer 11a notifies the calculation unit 21 of the second computer 11b of the structure determined in step S9, including the structure in an instruction to calculate the Curie temperature obtained from the structure. The calculation unit 21 of the second computer 11b, which has received the notification, calculates the calculated value of the Curie temperature obtained from the notified structure by first-principles calculation.

[0118] Next, the evaluation value acquisition unit 26 of the first calculator 11a acquires the evaluation value calculated by the calculation unit 21 of the second calculator 11b (step S13).

[0119] Next, similarly to the first embodiment, the evaluation value acquisition unit 26 of the first computer 11a determines whether to end the material search (step S15).

[0120] If it is determined that the material search should not be terminated (NO), the process proceeds to step S17, where the evaluation value acquisition unit 26 of the first computer 11a adds the material descriptor 32 corresponding to the structure determined by the structure determination unit 25 and the evaluation value acquired in step S13 to the second training data 27 of the second computer 11b. Then, the process returns to step S1 again.

[0121] On the other hand, if it is determined in step S15 that the material search is to be ended (YES), the process ends.

[0122] According to the present embodiment described above, the first, third, and fourth learning models 28 are made to learn the spin polarizability and magnetic moment in advance, and then the learning models 28 are made to learn the Curie temperature by transfer learning (steps S5, S27, and S29). Therefore, the first, third, and fourth learning models 28 are more likely to improve the prediction accuracy when predicting the Curie temperature due to the trend of the Curie temperature inherent in the trend of the spin polarizability and magnetic moment. This increases the possibility that a structure exhibiting a high Curie temperature will be determined in step S9. As a result, even if the number of times steps S1 to S17 are repeatedly executed is reduced, a material having a high Curie temperature exceeding the standard set by the user can be found, and the search time required for material search can be shortened.

[0123] Fig. 12 is a diagram obtained by investigating how much the search time is shortened in this embodiment. The horizontal axis of Fig. 12 indicates the number of times steps S1 to S17 are repeated (the number of loops). Moreover, the vertical axis of Fig. 12 indicates the evaluation value of the Curie temperature calculated in step S11.

[0124] In this investigation, AkaiKKR was used as the calculation code executed by the calculation unit 21 (see FIG. 2), and the calculation unit 21 executed the density functional theory as the first-principles calculation. In addition, A, a ternary magnetic alloy material having a B2 structure, was used as the material. x B y C z A plurality of structure data 30 corresponding to different composition ratios x, y, and z were used. The element types A, B, and C are not particularly limited, and may be any of Li, Be, B, Mg, Al, Si, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Ge, As, Y, Zr, Nb, Mo, Ru, Rh, Pd, Ag, Cd, In, Sn, Sb, Hf, Ta, W, Ir, Pt, and Au. Then, the composition ratios x, y, and z whose Curie temperatures exceed 900° C. were searched for. The search time of a comparative example in which the first to third transfer learning steps in steps S5, S27, and S29 were not performed was also investigated.

[0125] 12, in the comparative example, the number of loops required until a material with a Curie temperature exceeding 900° C. was found was 34. In contrast, in the present embodiment, a material with a Curie temperature exceeding 900° C. was found after 14 loops, and it was confirmed that the search time was shorter than that of the comparative example. From this result, it was confirmed that performing the first to third transfer learnings is effective in shortening the search time.

[0126] Moreover, since each of the first to third transfer learning is performed by the transfer learning unit 23 of the first computer 11a and the third computer 11c using the second training data 27 of the second computer 11b, it is possible to utilize the computational resources of each of the computers 11a to 11c.

[0127] In this embodiment, the execution subjects of steps S1 to S17 are merely examples, and each step may be distributed and executed by each of the computers 11a to 11c. For example, different processes among the plurality of steps S1 to S17 may be executed by different computers among the first to third computers 11a to 11c. This allows a material search to be performed while effectively utilizing the computational resources of each of the computers 11a to 11c in the same manner as described above.

[0128] Furthermore, although the number of computers is three in this example, similar processing may be performed using four or more computers.

[0129] Third embodiment In the first and second embodiments, the evaluation value in step S11 is calculated by first-principles calculation. In contrast, in the present embodiment, an experimental value obtained by an experiment is used as the evaluation value.

[0130] Fig. 13 is a system configuration diagram of a material exploration system according to this embodiment. In Fig. 13, the same elements as those described in the first embodiment are given the same reference numerals as those in the first embodiment, and their description will be omitted below.

[0131] 13, a material exploration system 10 according to this embodiment includes first to k-th robots 41a to 41k. The robots 41a to 41k are robots that measure the physical property values ​​of materials through experiments, and are connected to computers 11a to 11n via a network 12.

[0132] In this example, it is assumed that the physical property values ​​of the measurement targets are different for each of the first to k-th robots 41a to 41k. In the following, it is assumed that the measurement target of the second robot 41b is the Curie temperature. In this case, the measurement target of the first robot 41a is, for example, the spin polarizability.

[0133] In the following, it is assumed that at least two of the multiple robots 41a to 41k are managed by different managers. The locations where the robots 41a to 41k are installed are not particularly limited. For example, all of the robots 41a to 41k may be installed in Japan or outside Japan. Also, some of the robots 41a to 41k may be installed in Japan, and the remaining robots may be installed outside Japan.

[0134] Fig. 14 is a flow chart showing the processing performed by each computer in this embodiment. In Fig. 14, the same steps as those explained in the first and second embodiments are given the same reference numerals as those in these embodiments, and their explanation will be omitted below. The following processing can be applied to both the case where two computers 11a and 11b perform processing as in the first embodiment (see Fig. 6) and the case where three computers 11a to 11c perform processing as in the second embodiment (see Fig. 8).

[0135] In this example, the next structure is determined in step S9 (see FIGS. 7 and 11), and then the following processing is performed.

[0136] First, the second robot 41b performs an experiment on a material having a structure determined by the structure determination unit 25, and measures the Curie temperature of the material as an evaluation value (step S31). For example, the structure determination unit 25 (see FIG. 2) of the first computer 11a notifies the second robot 41b of the structure determined in step S9, including the structure in an instruction to measure the Curie temperature obtained from the structure. Then, the second robot 41b that has received the notification synthesizes a material of the notified structure, and measures the Curie temperature of the material.

[0137] The measured values ​​do not include calculation errors in the first-principles calculations, which increases the reliability of the evaluation values.

[0138] Next, the evaluation value acquisition unit 26 of the first computer 11a acquires the measurement value of the Curie temperature measured by the second robot 41b as an evaluation value (step S13).

[0139] Thereafter, the processes from step S15 onward are carried out in accordance with the first and second embodiments.

[0140] According to the present embodiment described above, the evaluation value of a material is measured by experiment, so that it is possible to determine in step S15 (see FIGS. 7 and 11) whether the evaluation value satisfies the criteria based on the evaluation value that is free of calculation error.

[0141] (Fourth embodiment) In the first to third embodiments, the first to n-th computers 11a to 11n execute the processes shown in Fig. 7, Fig. 10, Fig. 11, and Fig. 14 in a distributed manner. In this case, each of the computers 11a to 11n executes a process that is determined in advance in a material exploration program included in the computer. In contrast, in this embodiment, each of the computers 11a to 11n executes a process according to an instruction from the material exploration device as follows.

[0142] Fig. 15 is a system configuration diagram of a material exploration system according to this embodiment. In Fig. 15, the same elements as those described in the first and second embodiments are given the same reference numerals as those in these embodiments, and their description will be omitted below.

[0143] As shown in FIG. 15, a material exploration system 10 according to this embodiment includes first to n-th computers 11n and a material exploration device 43, which are connected to each other via a network 12.

[0144] The material search device 43 is a computer such as a PC or a server that issues instructions to each of the computers 11a to 11n to perform processing.

[0145] 16 is a functional configuration diagram of a material exploration device 43 according to this embodiment. As shown in FIG. 16, the material exploration device 43 includes a communication unit 44 and a control unit 45.

[0146] The communication unit 44 is an interface for connecting the communication unit 44 to the network 12 .

[0147] The control unit 45 is a processing unit that controls each unit of the material exploration device 43, and includes an instruction unit 46. The instruction unit 46 is a processing unit that issues instructions to each of the computers 11a to 11n to distribute and execute each process of Fig. 7, Fig. 10, Fig. 11, and Fig. 14.

[0148] According to such a material exploration device 43, since the instruction unit 45 issues instructions to each of the computers 11a to 11n, it is not necessary to install a material exploration program including processing contents and processing procedures in each of the computers 11a to 11n. Therefore, a material exploration can be performed without bothering the administrator of each of the computers 11a to 11n.

[0149] (Hardware configuration diagram) FIG. 17 is a diagram showing the hardware configuration of each of the computers 11a to 11n according to the first to third embodiments and the material exploration apparatus 43 according to the fourth embodiment.

[0150] 17, each of the computers 11a to 11n and the material exploration apparatus 43 includes a storage device 101, a memory 102, a processor 103, a communication interface 104, and a medium reading device 105. These components are connected to each other by a bus 106.

[0151] Of these, the storage device 101 is a non-volatile storage such as a hard disk drive (HDD) or a solid state drive (SSD), and stores a material exploration program 110 according to this embodiment.

[0152] The material search program 110 may be recorded on a computer-readable recording medium 111 , and the processor 103 may read the material search program 110 via the medium reading device 105 .

[0153] Such recording medium 111 may be, for example, a physically portable recording medium such as a CD-ROM (Compact Disc - Read Only Memory), a DVD (Digital Versatile Disc), or a USB (Universal Serial Bus) memory. A semiconductor memory such as a flash memory or a hard disk drive may also be used as the recording medium 111. These recording media 111 are not temporary media such as carrier waves that do not have a physical form.

[0154] Furthermore, the material search program 110 may be stored in a device connected to a public line, the Internet, a LAN, etc. In that case, the processor 103 may read and execute the material search program 110.

[0155] On the other hand, the memory 102 is hardware that temporarily stores data, such as a dynamic random access memory (DRAM).

[0156] The processor 103 is hardware such as a CPU (Central Processing Unit) or a GPU (Graphical Processing Unit) that controls each of the computers 11a to 11n and each part of the material exploration apparatus 43. The processor 103 also executes a material exploration program 110 in cooperation with the memory 102.

[0157] In this way, the memory 102 and the processor 103 cooperate to execute the material exploration program 110, thereby realizing the control unit 15 (see FIG. 2) of each of the computers 11a to 11n and the control unit 45 (see FIG. 16) of the material exploration apparatus 43.

[0158] The storage unit 16 (see FIG. 2) of each of the computers 11a to 11n is realized by the storage device 101 and the memory .

[0159] Furthermore, the communication interface 104 is hardware such as a network interface card (NIC) for connecting each of the computers 11a-11n and the material exploration apparatus 43 to the network 12. The communication interface 104 realizes the communication unit 14 (see FIG. 2) of each of the computers 11a-11n and the communication unit 44 of the material exploration apparatus 43.

[0160] The medium reading device 105 is hardware for reading the recording medium 111, such as a CD drive, a DVD drive, or a USB interface. [Explanation of symbols]

[0161] 10...material exploration system, 11a to 11n...1st to nth computers, 12...network, 14, 44...communication unit, 15, 45...control unit, 16...memory unit, 21...calculation unit, 22...machine learning unit, 23...transfer learning unit, 24...model selection unit, 25...structure determination unit, 26...evaluation value acquisition unit, 27...training data, 28...learning model, 29...physical property value, 30...structure data, 32...material descriptor, 33...set, 35...neural network, 36...input layer, 37...hidden layer, 38...output layer, 41a to 41k...1st to kth robots, 43...material exploration device, 46...instruction unit, 101...storage device, 102...memory, 103...processor, 104...communication interface, 105...medium reading device, 106...bus, 111...recording medium.

Claims

1. generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; performing a first transfer learning on the first learning model using the second training data so as to output the second physical property value corresponding to the structure; selecting a learning model having characteristics that satisfy a criterion from among a plurality of learning models including the first learning model on which the first transfer learning has been performed and the second learning model; determining a target structure for obtaining an evaluation value of the second physical property value based on the selected learning model; obtaining the evaluation value in the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data; A materials exploration program for distributing and executing the above-mentioned processing among a plurality of computers connected to each other via a network.

2. 2. The material exploration program according to claim 1, wherein the first machine learning is machine learning using an ensemble neural network.

3. the evaluation value is a calculated value of the second physical property value, 3. The material exploration program according to claim 1, wherein the processing includes calculating the calculated value.

4. Obtaining the evaluation value includes: notifying any one of the plurality of computers of the determined structure and an instruction to calculate the calculated value obtained from the structure; obtaining the calculated value from the one computer; 4. The material exploration program according to claim 3, further comprising:

5. the evaluation value is an experimental value of the second physical property value, 3. The material exploration program according to claim 1, wherein the processing includes an instruction to a robot to measure the experimental value.

6. 6. The material exploration program according to claim 1, wherein the characteristic is a generalization performance of each of the plurality of learning models.

7. generating the first learning model is executed by a first computer among the plurality of computers; generating the second learning model is executed by a second computer of the plurality of computers; The first computer inquires of the second computer whether the second learning model has been generated; A material exploration program as described in any one of claims 1 to 6, characterized in that when the second computer responds to the query that the second learning model has been generated, the program selects the learning model.

8. 8. The material exploration program according to claim 1, further comprising: repeating generating the first learning model, generating the second learning model, performing the first transfer learning, selecting the learning model, determining the structure, obtaining the evaluation value, and adding the material descriptor and the evaluation value to the second training data.

9. 9. The material exploration program according to claim 1, wherein different processes among generating the first learning model, generating the second learning model, performing the first transfer learning, selecting the learning model, determining the structure, obtaining the evaluation value, and adding the material descriptor and the evaluation value to the second training data are executed by different computers among the plurality of computers.

10. The process comprises: generating a third learning model that outputs the third physical property value corresponding to the structure by a third machine learning using a plurality of third material descriptors corresponding to each of the plurality of structures and a third physical property value of the material for each of the third material descriptors as third training data; generating a fourth learning model that outputs the first physical property value and the third physical property value corresponding to the structure by a fourth machine learning using the first training data and the third training data; performing second transfer learning on the third learning model using the second training data so as to output the second physical property value corresponding to the structure; performing a third transfer learning on the fourth learning model using the second training data to output the second physical property value corresponding to the structure; The plurality of learning models include the third learning model on which the second transfer learning has been performed, and the fourth learning model on which the third transfer learning has been performed. The material exploration program according to any one of claims 1 to 9.

11. Determining the target structure comprises: calculating a predicted value and a variance of the second physical property value using the selected learning model; generating a gain function based on the predicted value and the variance; determining a structure for which the acquisition function is maximum or minimum as the target structure; 11. The material exploration program according to claim 1, further comprising:

12. generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; performing a first transfer learning on the first learning model using the second training data so as to output the second physical property value corresponding to the structure; selecting a learning model having characteristics that satisfy a criterion from among a plurality of learning models including the first learning model on which the first transfer learning has been performed and the second learning model; determining a target structure for obtaining an evaluation value of the second physical property value based on the selected learning model; obtaining the evaluation value in the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data; The materials exploration method includes distributing and executing the above-mentioned processing among a plurality of computers connected to each other via a network.

13. generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; performing a first transfer learning on the first learning model using the second training data so as to output the second physical property value corresponding to the structure; selecting a learning model having characteristics that satisfy a criterion from among a plurality of learning models including the first learning model on which the first transfer learning has been performed and the second learning model; determining a target structure for obtaining an evaluation value of the second physical property value based on the selected learning model; obtaining the evaluation value in the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data; A materials exploration system in which the above-mentioned processes are carried out in a distributed manner by multiple computers connected to each other via a network.

14. generating a first learning model that outputs the first physical property values ​​corresponding to a plurality of structures of a material by a first machine learning using a plurality of first material descriptors corresponding to each of the plurality of structures of a material and a first physical property value of the material for each of the first material descriptors as first training data; generating a second learning model that outputs the second physical property values ​​corresponding to the structures by a second machine learning using a plurality of second material descriptors corresponding to each of the plurality of structures and a second physical property value of the material for each of the second material descriptors as second training data; performing a first transfer learning on the first learning model using the second training data so as to output the second physical property value corresponding to the structure; selecting a learning model having characteristics that satisfy a criterion from among a plurality of learning models including the first learning model on which the first transfer learning has been performed and the second learning model; determining a target structure for obtaining an evaluation value of the second physical property value based on the selected learning model; obtaining the evaluation value in the determined structure; and adding a material descriptor corresponding to the determined structure and the evaluation value to the second training data; The material exploration apparatus distributes the processing including the above to a plurality of computers connected to each other via a network and executes the processing.

Images