Electrode catalyst material, material search program, material search method, and material search device

Abstract

To provide a novel electrode catalyst material, a material search program, a material search method, and a material search device.SOLUTION: A material search program causes a computer to execute a process comprising: inputting a descriptor representing compounds into each of a plurality of machine learning models having different learning algorithms to predict a prediction value of an evaluation value of each of the plurality of compounds; creating, for each machine learning model, a plurality of lists in which the plurality of compounds are arranged in descending order of the prediction values; and presenting compounds that appear repeatedly in upper ranks of the plurality of lists.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an electrode catalyst material, a material search program, a material search method, and a material search apparatus. [Background technology]

[0002] Green hydrogen, a clean energy source that can replace fossil fuels, is attracting attention as a way to address global environmental issues such as global warming. Green hydrogen is an environmentally friendly energy source that does not emit carbon dioxide during its production or use.

[0003] One method for producing green hydrogen is through water electrolysis. In water electrolysis, green hydrogen is produced from a cathode electrode and oxygen is produced from an anode electrode. The anode electrode for oxygen production uses an electrode with an electrocatalyst on the surface of a current collector. Furthermore, various useful substances can be synthesized by mixing various gases and molecules with water. For example, mixing carbon dioxide with water produces green hydrogen, but the carbon dioxide can be reduced to produce methanol or acetate. The former is a fuel, and the latter is a starting material for producing various useful proteins. Electrocatalyst materials are preferably highly durable and resistant to degradation even during prolonged electrolysis to produce the desired substance. Therefore, platinum group elements such as platinum, ruthenium, and iridium, which are highly durable, are commonly used as electrocatalyst materials. These elements have the advantage of low overvoltage, allowing electrolysis of water and substrates at low voltage, and producing green hydrogen, methanol, acetate, and other substances with low power consumption.

[0004] However, platinum group elements are scarce and expensive, making it difficult to mass-produce useful substances such as green hydrogen and to supply the large quantities of electrode materials required. Therefore, active efforts are being made to develop electrocatalyst materials that do not contain platinum group elements but have overpotentials comparable to those of platinum group elements (Non-Patent Documents 1 and 2). When electrocatalyst materials are compounds composed of multiple elements, the number of possible element combinations and composition ratios is enormous. It is extremely difficult for a person to design an experiment based on their own experience and knowledge and discover an electrocatalyst material that exhibits the desired properties within a finite time frame. Therefore, active research is being conducted into methods for discovering novel electrocatalyst materials using data-driven methods such as machine learning (Non-Patent Document 3). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Ning Wang, et al., "Hydration-Effect-Promoting Ni-Fe Oxyhydroxide Catalysts for Neutral Water Oxidation," Advanced Materials, Vol. 32, pp. 1906-806, 2020. [Non-patent document 2] Hiroki Komiya and 2 others, "Electrolyte Engineering for Oxygen Evolution Reaction Over Non-Noble Metal Electrodes Achieving High Current Density in the Presence of Chloride Ion", ChemSusChem, Vol. 15, No. 202201088, 2022 [Non-patent document 3] Ken Sakaushi and two others, "Human-Machine Collaboration for Accelerated Discovery of Promising Oxygen Evolution Electrocatalysts with On-Demand Elements," ACS Central Science, Vol. 9, p. 2216, 2023 Summary of the Invention [Problem to be solved by the invention]

[0006] In one aspect, the present invention aims to provide a novel electrode catalyst material, a material search program, a material search method, and a material search device. [Means for solving the problem]

[0007] According to one aspect, the materials exploration program causes a computer to execute processes including predicting a predicted value of an evaluation value of each of a plurality of compounds by inputting descriptors representing the compound into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models in which the plurality of compounds are arranged in descending order of the predicted value; and presenting the compounds that appear repeatedly at the top of the plurality of lists.

[0008] In the above-described material exploration program, the descriptor may include a sum of the composition ratios of elements contained in the compound for each group, and an average value of any of the electronegativity, atomic number, atomic weight, electron affinity, ionization energy, neutron scattering cross section or absorption cross section, and van der Waals radius of the elements in the group.

[0009] In the material search program, the process may further include selecting, as the descriptor, a descriptor that has a correlation with the evaluation value.

[0010] In the above-described materials search program, the process may further include selecting, as the evaluation value, an evaluation value that is correlated with another evaluation value of the compound.

[0011] In the above-described materials exploration program, the evaluation value may be any one of overpotential, current density, and Tafel slope.

[0012] In the above-described material search program, the process may further include presenting the compound for each number of times that the compound appears at the top of the multiple lists.

[0013] In the material search program, the process may present an average value across the multiple lists of the predicted values of the compounds that appear at the top of the multiple lists.

[0014] In the above-mentioned material exploration program, the learning algorithm may be any one of Gaussian process regression using a radial basis function as a kernel function, Gaussian process regression using a rational quadratic kernel as a kernel function, NGBoost, XGBoost, and random forest.

[0015] In the above-described material search program, the top ranking may be a ranking higher than a predetermined ranking in each of the plurality of lists.

[0016] According to another aspect, the materials exploration method includes a process performed by a computer, the process including: predicting a predicted value of an evaluation value for each of a plurality of compounds by inputting descriptors that describe the compound into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models, in which the plurality of compounds are arranged in descending order of the predicted value; and presenting the compounds that appear repeatedly at the top of the plurality of lists.

[0017] According to another aspect, the materials exploration apparatus has a control unit that executes processes including: predicting a predicted value of an evaluation value of each of a plurality of compounds by inputting descriptors that describe the compound into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models in which the plurality of compounds are arranged in descending order of the predicted value; and presenting the compounds that appear repeatedly at the top of the plurality of lists.

[0018] According to yet another aspect, the electrocatalyst material comprises Mn 0.1 Fe 0.1 Co 0.1 Ag 0.1 W 0.6 , Mn 0.1 Fe 0.1 Co 0.1 Mo 0.6 Ag 0.1 , Mn 0.1 Fe 0.1 Co 0.1 Cu 0.1 W 0.6 , Fe 0.1 Co 0.1 Ag 0.1 W 0.7 , Fe 0.1 Co 0.1 Ag 0.2 W 0.6 , Fe 0.1 Co 0.2 Ag 0.1 W 0.6 , Fe 0.2 Co 0.1 Ag 0.1 W 0.6 , Fe 0.1 Co 0.1 Mo 0.1 Ag 0.1 W 0.6 , Fe 0.1 Co 0.1 Mo 0.1 Ag 0.2 W 0.5 , Fe 0.1 Co 0.2 Mo 0.1 Ag 0.1 W 0.5 , Fe 0.2 Co 0.1 Mo 0.1 Ag 0.1W 0.5 、Fe 0.1 Co 0.1 Mo 0.2 Ag 0.1 W 0.5 、Fe 0.1 Co 0.1 Mo 0.2 Ag 0.2 W 0.4 、Fe 0.1 Co 0.2 Mo 0.2 Ag 0.1 W 0.4 、Fe 0.2 Co 0.1 Mo 0.2 Ag 0.1 W 0.4 、Fe 0.1 Co 0.1 Mo 0.3 Ag 0.1 W 0.4 、Fe 0.1 Co 0.1 Mo 0.3 Ag 0.2 W 0.3 、Fe 0.1 Co 0.2 Mo 0.3 Ag 0.1 W 0.3 、Fe 0.2 Co 0.1 Mo 0.3 Ag 0.1 W 0.3 、Fe 0.1 Co 0.1 Mo 0.4 Ag 0.1 W 0.3 、Fe 0.1 Co 0.2 Mo 0.4 Ag 0.1 W 0.2 、Fe 0.2 Co 0.1 Mo 0.4 Ag 0.1 W 0.2 、Fe 0.1 Co 0.1 Mo 0.4 Ag 0.2 W 0.2 、Fe 0.1 Co 0.1 Mo 0.5 Ag 0.1 W 0.2 、Fe0.1 Co 0.1 Mo 0.5 Ag 0.2 W 0.1 、Fe 0.1 Co 0.2 Mo 0.5 Ag 0.1 W 0.1 、Fe 0.2 Co 0.1 Mo 0.5 Ag 0.1 W 0.1 、Fe 0.1 Co 0.1 Mo 0.6 Ag 0.1 W 0.1 、Mn 0.1 Fe 0.1 Co 0.1 Cu 0.1 Mo 0.6 、Fe 0.1 Co 0.1 Mo 0.6 Ag 0.2 、Fe 0.1 Co 0.1 Mo 0.7 Ag 0.1 、Fe 0.1 Co 0.2 Mo 0.6 Ag 0.1 、Fe 0.2 Co 0.1 Mo 0.6 Ag 0.1 、Cr 0.1 Fe 0.1 Co 0.1 Ag 0.1 W 0.6 、Cr 0.1 Fe 0.1 Co 0.1 Ag 0.2 W 0.5 、Cr 0.1 Fe 0.1 Co 0.2 Ag 0.1 W 0.5 、Cr 0.1 Fe 0.2 Co 0.1 Ag 0.1 W 0.5 、及びFe 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6The compound includes a compound selected from the group consisting of: [Effects of the Invention]

[0019] According to the present invention, it is possible to provide a novel electrode catalyst material, a material search program, a material search method, and a material search device. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is an example of a functional configuration diagram of a material exploration apparatus according to the first embodiment. [Figure 2] FIG. 2 is a schematic diagram illustrating an example of a periodicity descriptor according to the first embodiment. [Figure 3] FIG. 3 is a schematic diagram of an example of experimental data according to the first embodiment. [Figure 4] FIG. 4 is a schematic diagram of an example of training data according to the first embodiment. [Figure 5] FIG. 5 is a schematic diagram of an example of a candidate compound list according to the first embodiment. [Figure 6] FIG. 6 is a schematic diagram of an example of an overlapping compound list according to the first embodiment. [Figure 7] FIG. 7 is a schematic diagram showing an example of processing performed by the overlap checking unit when generating a plurality of overlapping compound lists in the first embodiment. [Figure 8] FIG. 8 is a flowchart showing an example of processing performed by the material exploration apparatus according to the first embodiment. [Figure 9] FIG. 9 shows a list of overlapping compounds obtained in the examples. [Figure 10] FIG. 10 is a graph showing the measurement results of the characteristics of the water splitting apparatus in the example. [Figure 11] FIG. 11 is a diagram showing an example of the hardware configuration of the material exploration apparatus according to the first embodiment. [Figure 12] FIG. 12 is a diagram showing an example of the configuration of a water decomposing apparatus according to the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0021] (First embodiment) In the first embodiment, a data-driven method using multiple machine learning models with different learning algorithms is used to search for novel compounds for electrode catalyst materials that have low overpotentials and can promote electrolysis reactions at low voltages, as described below. Furthermore, to realize low-cost electrode catalyst materials, novel compounds that do not contain platinum group elements are searched for.

[0022] 1 is a diagram showing an example of the functional configuration of a materials exploration apparatus according to this embodiment. The materials exploration apparatus 10 is a computer such as a personal computer or a server, and performs a materials search using periodic descriptors.

[0023] 2 is a schematic diagram for explaining an example of a periodic table descriptor according to this embodiment. The periodic table descriptor 19 is an explanatory variable that expresses a compound consisting of multiple elements as an array, and is created based on the periodic table 18.

[0024] In this example, the periodic table descriptor 19 includes a first array 19a and a second array 19b. Each of the arrays 19a and 19b has a plurality of elements corresponding to each group of the periodic table 18. For example, the ith (1≦i≦18) element of each of the arrays 19a and 19b corresponds to the ith group.

[0025] In this case, the sum of the composition ratios of the elements of the i-th group contained in the compound is stored in the i-th element of the first array 19a. For example, if the composition formula is "A 0.2 B 0.3 D 0.2 E 0.1 G 0.2 In this compound, elements A and B belong to group 1, and element D belongs to group 3. Elements E and G belong to groups 4 and 17, respectively.

[0026] In this case, the first element of the first array 19a stores "0.5", which is the sum of the composition ratios of the elements A and B belonging to the first group, "0.2" and "0.3". 0.2 B0.3 D 0.2 E 0.1 G 0.2 Since group 2, the fifth to sixth, and the eighteenth elements are not included in "," "0" is stored in the second, fifth to sixth, and eighteenth elements of first array 19a. Then, "0.2," "0.1," and "0.2," which are the composition ratios of elements D, E, and G, are stored in the third, fourth, and seventeenth elements of first array 19a. In this way, first array 19a stores the sum of the composition ratios of elements included in the compound for each group.

[0027] On the other hand, the i-th element of the second array 19b stores the average value of the feature quantities that indicate the characteristics of the i-th group elements contained in the compound, such as the atomic radius, electronegativity, atomic number, atomic weight, electron affinity, ionization energy, neutron scattering cross section or absorption cross section, and van der Waals radius.

[0028] In the following, we will explain the feature quantity by taking the atomic radius as an example. 0.2 B 0.3 D 0.2 E 0.1 G 0.2 In the case of the compound "A", the first element of the second array 19b stores "0.59", which is the average value of the atomic radii (Å) of the elements A and B belonging to the first group. 0.2 B 0.3 D 0.2 E 0.1 G 0.2 " does not include the 2nd group, 5th to 16th, and 18th elements. In this case, "0" is stored in the 2nd, 5th to 16th, and 18th elements of the second array 19b. Then, "1.39", "1.09", and "1.35", which are the atomic radii (Å) of the elements D, E, and G, are stored in the 3rd, 4th, and 17th elements of the second array 19b.

[0029] Descriptors are explanatory variables in machine learning that predict evaluation values such as overvoltage, and have a significant impact on the prediction results. The periodic table descriptor 19 according to this embodiment includes a first array 19a that indicates the compound composition, as well as a second array 19b that stores feature quantities specific to each group. Therefore, compared to a descriptor that only expresses the compound composition, the periodic table descriptor 19 contains more information about the compound's features, which can improve the prediction accuracy by machine learning.

[0030] Furthermore, the number of elements in each of arrays 19a and 19b is equal to 18, which is the number of groups in the periodic table, and the total number of elements in periodic descriptor 19 is 36 (= 2 × 18). However, if there is a group that is clearly not included in the compound being searched for, the element in each of arrays 19a and 19b corresponding to that group may be deleted. For example, since the 18th element of the noble gases is unlikely to form a compound with other elements, the 18th element in each of arrays 19a and 19b may be deleted. This reduces the total number of elements in periodic descriptor 19 to 34, thereby reducing the computational cost in machine learning.

[0031] Referring again to Fig. 1, the material exploration device 10 includes a communication unit 11, a control unit 12, and a storage unit 13.

[0032] The storage unit 13 is a processing unit that stores various information necessary for material search. In this example, the storage unit 13 stores experimental data 14, training data 15, multiple machine learning models 16, multiple candidate compound lists 51, and overlapping compound lists 52.

[0033] FIG. 3 is a schematic diagram of an example of experimental data 14 according to this embodiment. As shown in FIG. 3, the experimental data 14 is information that associates a "compound" with the experimental value of its evaluation value. In this example, an oxide of the "compound" is synthesized as a candidate material for the electrode catalyst material. The composition formula of the "compound" is the composition formula of the portion obtained by removing oxygen from the oxide actually synthesized in this manner. The number of elements contained in each "compound" is not particularly limited. In this example, mono- to penta-compound compounds are included in the "compound." The elements contained in the "compound" are schematically represented in FIG. 3 by capital letters "A," "B," "C," "D," and "E." As mentioned above, oxygen is removed from the composition formula of the "compound," so "A," "B," "C," "D," and "E" all represent elements other than oxygen.

[0034] The electrode catalyst material is not limited to oxides of "compounds," but may be a group of substances containing metals and various anions, such as sulfides, nitrides, fluorides, chlorides, carbides, boron compounds, and oxychlorides of "compounds." The group of substances can be obtained by synthesizing the electrode catalyst material in the presence of these anions.

[0035] The evaluation value is a value that evaluates the performance of an electrode catalyst material that uses an oxide of a "compound." Examples of evaluation values include "overvoltage," "current density," and "Tafel gradient." For example, the smaller the "overvoltage," the smaller the voltage required for water electrolysis, and the better the electrode catalyst material is evaluated. Furthermore, "current density" is the density of the current that flows between the electrodes when a predetermined voltage is applied between the anode and cathode electrodes, and the higher this value, the better the electrode catalyst material is evaluated as being capable of efficient electrolysis. The same is true for the Tafel gradient.

[0036] It is not necessary to store all of the "overvoltage," "current density," and "Tafel slope" in the experimental data 14; the user may store at least one of these in the experimental data 14. Furthermore, evaluation values other than these may be stored in the experimental data 14.

[0037] 4 is a schematic diagram of an example of training data 15 according to this embodiment. The training data 15 is data used for learning the machine learning model 16, and is generated by the materials exploration apparatus 10 from the experimental data 14.

[0038] The "descriptors" of the training data 15 are, among multiple types of descriptors that describe each compound included in the experimental data 14, descriptors that are correlated with the experimental value of "overpotential," which is an example of an evaluation value. In this example, a periodic descriptor 19 that is correlated with the experimental value of "overpotential" is used as the descriptor of the training data 15. In the following, a first array 19a and a second array 19b of the periodic descriptor 19 are separated by a semicolon ";."

[0039] Furthermore, if there are multiple types of descriptors that are correlated with the experimental evaluation value, a combination of an evaluation value and a descriptor that is correlated with another evaluation value may be used in the training data 15.

[0040] Referring again to FIG. 1 , the machine learning model 16 is a model that predicts the evaluation value of each compound, and multiple models are stored in the storage unit 13. It is preferable that the user selects multiple machine learning models 16 that have high prediction accuracy, each with a different algorithm, and that output the prediction results as a probability distribution. High prediction accuracy is preferable because the closer the predicted evaluation value of each compound is to the experimental value and the more reliable it is, the less effort is required to synthesize the compound experimentally and obtain its evaluation value. Furthermore, by employing machine learning models 16 with different algorithms, the reliability of the prediction can be improved when multiple machine learning models 16 predict that the predicted evaluation value of a certain compound is good. Furthermore, by outputting the prediction results as a probability distribution, the user can grasp the statistical reliability of the prediction results. Hereinafter, the term "predicted value" simply refers to an average value based on a probability distribution.

[0041] In this example, the user stores five machine learning models 16 using the learning algorithms Gaussian process regression (rbf), Gaussian process regression (rq), NGBoost, XGBoost, and Random Forest in the storage unit 13 in advance. Note that Gaussian process regression (rbf) is a Gaussian process regression that uses a radial basis function as a kernel function. Also, Gaussian process regression (rq) is a Gaussian process regression that uses a rational quadratic kernel as a kernel function.

[0042] 5 is a schematic diagram of an example of a candidate compound list 51 according to this embodiment. The candidate compound list 51 is a list in which compounds are arranged in descending order of the magnitude of the predicted evaluation value, and is stored in the storage unit 13 for each machine learning model 16. Note that, to avoid cluttering the diagram, FIG. 5 only shows the candidate compound lists 51 generated by two machine learning models 16, "A" and "B."

[0043] The compounds may be arranged in the candidate compound list 51 in ascending or descending order of predicted evaluation values. For example, if a smaller evaluation value is preferred, the materials exploration apparatus 10 arranges the compounds in descending order of predicted value so that compounds with smaller predicted values are ranked higher. On the other hand, if a larger evaluation value is preferred, the materials exploration apparatus 10 arranges the compounds in descending order of predicted value so that compounds with larger predicted values are ranked higher.

[0044] In this example, "overvoltage" is used as the evaluation value, but in the case of electrode catalyst materials, a small overvoltage is desired in order to reduce the voltage required for water electrolysis. Therefore, in the example in Figure 5, the "compounds" are arranged in ascending order of "overvoltage."

[0045] As described above, since the learning algorithms of the machine learning models 16 are different, the candidate compound lists 51 generated by the machine learning models 16 do not completely match. However, there are cases where a compound appears at the top of all candidate compound lists 51. In the example of Figure 5, compounds 51a to 51f appear in both candidate compound lists 51.

[0046] 6 is a schematic diagram of an example of the overlapping compound list 52 according to this embodiment. The overlapping compound list 52 is a list of compounds 51a to 51f that appear overlappingly in all of the candidate compound lists 51 and their predicted evaluation values. The predicted values of the compounds 51a to 51f in the candidate compound list 51 are the average values of the predicted values of the compounds across all of the candidate compound lists 51.

[0047] 6, overpotential is used as the evaluation value, and the average value of the overpotential in each of the machine learning models 16 of A and B is stored in the overlapping compound list 52. Then, the compounds 51a to 51f are arranged in ascending order of the average value of the overpotential.

[0048] Referring again to Figure 1, the communication unit 11 is an interface for connecting the material exploration apparatus 10 to a network such as the Internet or a LAN (Local Area Network).

[0049] The control unit 12 is a processing unit that controls each unit of the material exploration apparatus 10, and includes an acquisition unit 21, a correlation analysis unit 22, a selection unit 23, a machine learning unit 24, a prediction unit 25, a list creation unit 26, a duplication confirmation unit 27, a presentation unit 28, and an input acceptance unit 29.

[0050] The acquisition unit 21 acquires the experiment data 14 generated by the user via the communication unit 11 and stores the experiment data 14 in the storage unit 13 .

[0051] The correlation analysis unit 22 calculates a correlation coefficient indicating the strength of correlation between a plurality of types of descriptors representing compounds in the experimental data 14 (see FIG. 3) and a plurality of types of evaluation values included in the experimental data 14 for all combinations. An example of the correlation coefficient is the Pearson product-moment correlation coefficient. Note that the types of descriptors for which the correlation coefficient is calculated are defined, for example, in the code of the materials exploration program, and include the aforementioned periodic descriptor 19 (see FIG. 2), as well as one-hot vectors and various descriptors provided by computational codes such as Matminer and XenonPy.

[0052] Furthermore, the correlation analysis unit 22 may calculate the Pearson product-moment correlation coefficient as a correlation coefficient indicating the strength of correlation between two evaluation values for all two combinations of multiple types of evaluation values, such as overvoltage and current density, included in the experimental data 14.

[0053] The selection unit 23 selects the descriptors and the evaluation values that are correlated with each other based on the correlation coefficient calculated by the correlation analysis unit 22. As an example, the selection unit 23 determines that there is a correlation between the descriptors and the evaluation values when the correlation coefficient is equal to or greater than a predetermined value.

[0054] When there are multiple combinations of correlated descriptors and evaluation values, the selection unit 23 may determine whether each evaluation value in the multiple combinations is correlated with another evaluation value. Whether the evaluation values are correlated with each other is determined based on the correlation coefficient calculated by the correlation analysis unit 22. As an example, the selection unit 23 determines that the evaluation values are correlated with each other if the correlation coefficient is equal to or greater than a predetermined value. The selection unit 23 may then select a descriptor and evaluation value related to a combination determined to be correlated with another evaluation value. For example, when there is a correlation between the periodic descriptor 19 and overvoltage and there is also a correlation between the overvoltage and current density, the selection unit 23 selects the periodic descriptor 19 and the overvoltage.

[0055] The machine learning unit 24 generates training data 15 from the experimental data 14. As an example, the machine learning unit 24 represents each compound in the experimental data 14 using a descriptor selected by the selection unit 23. The selection unit 23 then generates training data 15 by associating the descriptor with the evaluation value selected by the selection unit 23 for the experimental data 14.

[0056] Furthermore, the machine learning unit 24 generates a plurality of machine learning models 16 with different learning algorithms, and further trains each machine learning model 16 by supervised learning using the training data 15. At this time, the machine learning unit 24 adjusts hyperparameters in each machine learning model 16 so as to reduce the error between the evaluation value in the training data 15 and the predicted value of that evaluation value.

[0057] Each machine learning model 16 is a model that outputs a predicted value of the evaluation value selected by the selection unit 23 when a descriptor selected by the selection unit 23 is input. Because the selection unit 23 selects descriptors that are correlated as evaluation values, it is possible to suppress large variations in the predicted value of the evaluation value when the descriptor is input to the machine learning model 16.

[0058] Furthermore, as described above, if the selection unit 23 selects an evaluation value that is correlated with another evaluation value, if the predicted value of that evaluation value is good, the user can determine that the other predicted value is also good.

[0059] The prediction unit 25 predicts the evaluation value of each of the multiple compounds using each machine learning model 16. For example, the prediction unit 25 generates the descriptor selected by the selection unit 23 for each of the multiple compounds to be predicted, and inputs it to each learning model 16, thereby obtaining a predicted value of the evaluation value of each compound from each machine learning model 16.

[0060] A plurality of compounds to be predicted are listed in advance, for example, in a file (not shown) stored in the storage unit 13. The file may be created by the user or automatically by the materials exploration apparatus 10.

[0061] The number of elements contained in the compound to be predicted is preferably the same as the number of elements in the experimental data 14. For example, if the compounds contained in the experimental data 14 are mono- to penta-element compounds as described above, the compound to be predicted is preferably also a mono- to penta-element compound. This ensures that the number of elements used for training and prediction is the same, thereby improving the prediction accuracy of each machine learning model 16.

[0062] Furthermore, the elements contained in the compound to be predicted are preferably metal elements other than platinum group elements in order to reduce the cost of the electrode catalyst material.

[0063] The list creation unit 26 creates, for each machine learning model 16, a plurality of candidate compound lists 51 (see FIG. 5) in which the compounds are arranged in descending order of the magnitude of the predicted evaluation value.

[0064] The overlap checking unit 27 identifies compounds that appear overlappingly at the top of multiple candidate compound lists 51. For example, a compound that appears higher than a predetermined rank in each of the candidate compound lists 51 is considered to be higher. Then, the overlap checking unit 27 generates an overlap compound list 52 (see FIG. 6 ) that associates the identified compounds with predicted values of their evaluation values, and stores the overlap compound list 52 in the storage unit 13. As an example, the overlap checking unit 27 averages the predicted values of the identified compounds across the multiple candidate compound lists 51, and arranges the compounds in descending order of the average value to generate the overlap compound list 52.

[0065] The overlap checking unit 27 may generate a plurality of overlap compound lists 52 as follows, instead of one overlap compound list 52 as shown in FIG.

[0066] 7 is a schematic diagram showing an example of processing performed by the overlap checking unit 27 when generating multiple overlapping compound lists 52. In the example of Fig. 7, the overlap checking unit 27 generates multiple overlapping compound lists 52 for each number of times a compound appears at the top of the candidate compound lists 51 obtained from each of the five machine learning models 16.

[0067] For example, the "five times" overlapping compound list 52 is a list of compounds that overlap and appear in the top positions of all five candidate compound lists 51. The "twice" to "four times" overlapping compound lists 52 are lists of compounds that overlap and appear in the top positions of two to four candidate compound lists 51, respectively. The "single" candidate compound list 51 is a list of compounds that appear in the top positions of only one of the five candidate compound lists 51.

[0068] Even though the learning algorithms of the machine learning models 16 are different, compounds that appear at the top of multiple candidate compound lists 51 can be trusted to have excellent evaluation values for overpotential, etc. Furthermore, the reliability of the predicted value increases as the number of times a compound appears repeatedly increases.

[0069] 1 again, the presentation unit 28 is a processing unit that presents to the user the overlapping compound list 52. For example, the presentation unit 28 displays the overlapping compound list 52 on a display device 30 such as a liquid crystal display.

[0070] As described above, the compounds included in the overlapping compound list 52 can be trusted to have excellent predicted evaluation values for overpotential, etc. Therefore, by actually synthesizing the compounds in the overlapping compound list 52, it is highly likely that an electrode catalyst material capable of electrolyzing water at low voltage can be obtained.

[0071] 7, the presentation unit 28 may present a list of overlapping compounds 52 based on the number of times that compounds overlap and appear at the top of multiple candidate compound lists 51. This allows the user to understand the reliability of the predicted value according to the number of times. As a result, by actually synthesizing compounds with high reliability and a large number of times, and not synthesizing compounds with low reliability and a small number of times, it is possible to reduce unnecessary experiments and efficiently search for new materials.

[0072] The input receiving unit 29 receives various instructions input from an input device 40 such as a keyboard or a mouse. For example, the input receiving unit 29 receives from the input device 40 an instruction to start a material search.

[0073] Next, a material search method according to this embodiment will be described. Fig. 8 is a flowchart showing an example of processing performed by the material search device 10 according to this embodiment. This processing is started when, for example, the user issues an instruction to start the processing via the input device 20.

[0074] First, the acquisition unit 21 acquires the experimental data 14 and stores it in the storage unit 13 (step S1).

[0075] Next, the correlation analysis unit 22 calculates (step S2) a correlation coefficient between the descriptor and the evaluation value for all combinations of the multiple types of descriptors representing the compounds in the experimental data 14 (see FIG. 3) and the multiple types of evaluation values included in the experimental data 14. Furthermore, the correlation analysis unit 22 may calculate a correlation coefficient indicating the strength of the correlation between the two evaluation values for all combinations of two types of evaluation values included in the experimental data 14.

[0076] Next, the selection unit 23 selects a descriptor and an evaluation value that are correlated with each other based on the correlation coefficient between the descriptor and the evaluation value (step S3). For example, the selection unit 23 selects the periodicity descriptor 19 and the overvoltage as described above. Note that if there are multiple combinations of descriptors and evaluation values that are correlated with each other, the selection unit 23 may select an evaluation value and a descriptor that are correlated with another evaluation value based on the correlation coefficient between the evaluation values.

[0077] Next, the machine learning unit 24 generates training data 15 from the experimental data 14 by associating the selected descriptors with evaluation values (step S4).

[0078] Next, the machine learning unit 24 generates a plurality of machine learning models 16 (step S5), and further trains each machine learning model 16 by supervised learning using the training data 15 (step S6).

[0079] Next, the prediction unit 25 generates a plurality of descriptors that represent each of the plurality of compounds to be predicted (step S7). Then, the prediction unit 25 inputs the descriptors to each machine learning model 16, thereby predicting the predicted value of the evaluation value of each compound for each machine learning model 16 (step S8).

[0080] Next, the list creation unit 26 creates a plurality of candidate compound lists 51 (see FIG. 5) for each machine learning model 16, in which the compounds are arranged in descending order of the magnitude of the predicted evaluation value (step S9).

[0081] Furthermore, the overlap checking unit 27 identifies compounds that appear in the top positions of multiple candidate compound lists 51 (step S10), and generates an overlap compound list 52 in which the overlapping compounds are arranged in order of the average value of the predicted value (step S11). Note that the overlap checking unit 27 may generate multiple overlap compound lists 52 for each number of times a compound appears in the top positions of the candidate compound lists 51, as shown in FIG.

[0082] Next, the presentation unit 28 presents the overlapping compound list 52 to the user (step S12).

[0083] This completes the basic processing performed by the material exploration apparatus 10 according to this embodiment.

[0084] According to the present embodiment described above, the materials exploration apparatus 10 generates a candidate compound list 51, in which a plurality of compounds are arranged in descending order of predicted evaluation values, for each of a plurality of machine learning models 16 with different learning algorithms (step S9). Then, the materials exploration apparatus 10 presents to the user a duplicate compound list 52, in which compounds that appear duplicated at the top of each candidate compound list 51 are arranged (step S12). Compounds that appear duplicated at the top of each candidate compound list 51 have highly reliable predicted evaluation values. Therefore, the experimental evaluation values obtained by actually synthesizing the compounds are very close to the predicted values, making it possible to quickly discover and synthesize compounds that are advantageous for improving the performance of electrode catalyst materials.

[0085] Next, an example of this embodiment will be described. The number of compounds included in the experimental data 14 (FIG. 3) in this example was 120. Furthermore, the compounds to be predicted were mono- to penta-nary compounds. The elements constituting each compound were 18: Sc (scandium), Ti (titanium), V (vanadium), Cr (chromium), Mn (manganese), Fe (iron), Co (cobalt), Ni (nickel), Cu (copper), Zn (zinc), Y (yttrium), Zr (zirconium), Nb (niobium), Mo (molybdenum), Ag (silver), Hf (hafnium), Ta (tantalum), and W (tungsten). In this case, the total number of mono- to penta-nary compounds to be searched was 1,358,816. Furthermore, because none of the 18 elements are platinum group elements, the cost of the electrode catalyst material can be reduced. As mentioned above, the electrode catalyst material actually used is an oxide of a compound composed of these elements.

[0086] The number of machine learning models16 was set to five, and the learning models used were Gaussian process regression (rbf), Gaussian process regression (rq), NGBoost, XGBoost, and random forest. Furthermore, the aforementioned periodicity descriptor19 was used as the descriptor, and the overpotential was used as the evaluation value for each compound. The overpotentials of all 1,358,816 search target compounds were predicted.

[0087] FIG. 9 shows the overlapping compound list 52 obtained in this example. The "compounds" in this overlapping compound list 52 are compounds that appear in all of the top positions in the five candidate compound lists 51 corresponding to the five machine learning models 16. The "rank" indicates the rank of each compound when the compounds are sorted in ascending order of predicted overpotential. Compounds with the same predicted overpotential were assigned the same rank. Of the 1,358,816 compounds, the compounds ranked 1st to 20th were included in the overlapping compound list 52. The reason for limiting the ranks to be included in the overlapping compound list 52 to 1st to 20th is that compounds ranked 21st and below cannot be expected to have a sufficiently low overpotential to improve the performance of the electrode catalyst material.

[0088] In this example, the "predicted value of overvoltage" is 1 mA / cm 2 The predicted voltage between the anode and cathode electrodes required to obtain a current density of 0 mA / cm was used. 2 and 1mA / cm 2 This is because it is difficult to precisely calculate the voltage at which current begins to flow.

[0089] As shown in Figure 9, the compound at position 20 is Fe 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 The predicted overvoltage was 335.38 mV. 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 If the compound is ranked 1st to 19th above Fe 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 Therefore, the inventors of the present invention have developed a method for producing a material having an overvoltage smaller than that of Fe. 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 The oxide was actually synthesized and the characteristics of a water splitting device using the oxide as an electrocatalyst material for the anode electrode were measured.

[0090] Fe 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 The oxide synthesis was carried out by first dissolving the chlorides of Fe, Co, Cu, Ag, and W in ethanol to adjust the composition ratio of each element to 0.1:0.1:0.1:0.1:0.6, then dropping the solution onto a titanium metal substrate and firing it. The firing was carried out in air at atmospheric pressure for 1 hour at 450°C.

[0091] Figure 10 is a graph showing the measurement results. The horizontal axis of the graph represents the voltage between the cathode electrode and the anode electrode, and the vertical axis represents the density of the current flowing between the cathode electrode and the anode electrode.

[0092] 10 also shows the measurement results of a first comparative example in which RuO2 (ruthenium oxide), an oxide of a platinum group element, was used as the electrode catalyst material. 0.5 The measurement results when Ag was used as the electrode catalyst material are also shown. 0.5 Ag is a compound predicted by Non-Patent Document 3 as a material with a high oxygen evolution reaction (OER).

[0093] As shown in Figure 10, the graph of this example has a steeper slope than the graphs of the first and second comparative examples. Therefore, in this example, the voltage required to obtain a certain current density can be made smaller than in the first and second comparative examples. For example, 40 mA / cm, which is close to the practical condition, can be obtained. 2 The voltage required to obtain a current density of 1.8 V in the present example is lower than that in the first and second comparative examples, which is approximately 1.9 V and 2.0 V. As a result, the present example can efficiently electrolyze water at a lower voltage than the comparative examples.

[0094] In addition, the actual operating conditions are close to 40mA / cm 2 The voltage obtained by subtracting the theoretical water decomposition voltage (1.23 V) from the voltage required to obtain this current density is the overvoltage at that current density. The experimental value of this overvoltage in this example was 0.53 V, which was confirmed to be smaller than the experimental values in the first and second comparative examples (0.56 V and 0.59 V, respectively).

[0095] As described above, the Fe 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6It has been revealed that this oxide is an excellent material as an electrocatalyst material for the anode used in water electrolysis, even though it does not contain platinum group elements.

[0096] Furthermore, Fe according to this embodiment 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 As described above, the predicted value of the overpotential of 1 was 335.38 mV, while the experimental value was 353.40 mV. As a result, it was confirmed that the prediction error, defined as the difference between the predicted value and the experimental value, was 18.2 mV, which is sufficiently small. This proves that the compounds that appeared at the top of multiple candidate compound lists 51 actually have overpotentials close to the experimental value, and that the prediction accuracy of this embodiment is sufficiently high.

[0097] The compounds ranked 1 to 19 in Figure 9, which were discovered using this highly accurate material search method, also have the same properties as the 20th compound, Fe. 0.1 Co 0.1 Cu 0.1 Ag 0.1 W 0.6 Similar to the above, it is expected that such compounds will have small overpotentials despite not containing any platinum group elements. 0.1 Fe 0.1 Co 0.1 Ag 0.1 W 0.6 , Mn 0.1 Fe 0.1 Co 0.1 Mo 0.6 Ag 0.1 , Mn 0.1 Fe 0.1 Co 0.1 Cu 0.1 W 0.6 , Fe 0.1 Co 0.1 Ag 0.1 W 0.7 , Fe 0.1 Co 0.1 Ag 0.2 W 0.6 , Fe 0.1 Co 0.2 Ag 0.1 W 0.6 , Fe0.2 Co 0.1 Ag 0.1 W 0.6 、Fe 0.1 Co 0.1 Mo 0.1 Ag 0.1 W 0.6 、Fe 0.1 Co 0.1 Mo 0.1 Ag 0.2 W 0.5 、Fe 0.1 Co 0.2 Mo 0.1 Ag 0.1 W 0.5 、Fe 0.2 Co 0.1 Mo 0.1 Ag 0.1 W 0.5 、Fe 0.1 Co 0.1 Mo 0.2 Ag 0.1 W 0.5 、Fe 0.1 Co 0.1 Mo 0.2 Ag 0.2 W 0.4 、Fe 0.1 Co 0.2 Mo 0.2 Ag 0.1 W 0.4 、Fe 0.2 Co 0.1 Mo 0.2 Ag 0.1 W 0.4 、Fe 0.1 Co 0.1 Mo 0.3 Ag 0.1 W 0.4 、Fe 0.1 Co 0.1 Mo 0.3 Ag 0.2 W 0.3 、Fe 0.1 Co 0.2 Mo 0.3 Ag 0.1 W 0.3 、Fe 0.2 Co 0.1 Mo 0.3 Ag 0.1 W 0.3 、Fe 0.1 Co 0.1 Mo 0.4Ag 0.1 W 0.3 、Fe 0.1 Co 0.2 Mo 0.4 Ag 0.1 W 0.2 、Fe 0.2 Co 0.1 Mo 0.4 Ag 0.1 W 0.2 、Fe 0.1 Co 0.1 Mo 0.4 Ag 0.2 W 0.2 、Fe 0.1 Co 0.1 Mo 0.5 Ag 0.1 W 0.2 、Fe 0.1 Co 0.1 Mo 0.5 Ag 0.2 W 0.1 、Fe 0.1 Co 0.2 Mo 0.5 Ag 0.1 W 0.1 、Fe 0.2 Co 0.1 Mo 0.5 Ag 0.1 W 0.1 、Fe 0.1 Co 0.1 Mo 0.6 Ag 0.1 W 0.1 、 Mr 0.1 Feb 0.1 Co 0.1 Cu 0.1 Mo 0.6 、Fe 0.1 Co 0.1 Mo 0.6 Ag 0.2 、Fe 0.1 Co 0.1 Mo 0.7 Ag 0.1 、Fe 0.1 Co 0.2 Mo 0.6 Ag 0.1 、Fe 0.2 Co 0.1 Mo 0.6 Ag 0.1 、Cr 0.1 Feb 0.1 Co 0.1 Ag0.1 W 0.6 , Cr 0.1 Fe 0.1 Co 0.1 Ag 0.2 W 0.5 , Cr 0.1 Fe 0.1 Co 0.2 Ag 0.1 W 0.5 , and Cr 0.1 Fe 0.2 Co 0.1 Ag 0.1 W 0.5 is.

[0098] <Hardware configuration> Next, the hardware configuration of the material exploration apparatus 10 will be described.

[0099] Fig. 11 is a diagram showing an example of the hardware configuration of a material exploration apparatus 10 according to this embodiment. As shown in Fig. 11, the material exploration apparatus 10 includes a storage device 101, a memory 102, a processor 103, a communication interface 104, and a medium reader 105. These components are connected to each other via a bus 106.

[0100] Of these, the storage device 101 is a non-volatile storage such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores a material exploration program 110 according to this embodiment.

[0101] 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 reader 105 .

[0102] Such recording media 111 include physically portable recording media such as CD-ROMs (Compact Disc - Read Only Memory), DVDs (Digital Versatile Discs), and USB (Universal Serial Bus) memories. Also, semiconductor memories such as flash memories and hard disk drives may be used as the recording media 111. These recording media 111 are not temporary media such as carrier waves that do not have a physical form.

[0103] Furthermore, the material search program 110 may be stored in a device connected to a public line, the Internet, a LAN, etc. In this case, the processor 103 may read and execute the material search program 110.

[0104] On the other hand, the memory 102 is hardware that temporarily stores data, such as a DRAM (Dynamic Random Access Memory).

[0105] The processor 103 is hardware such as a CPU (Central Processing Unit) or a GPU (Graphical Processing Unit) that controls each part of the material exploration apparatus 10. The processor 103 also executes the material exploration program 110 in cooperation with the memory 102.

[0106] In this way, the memory 102 and the processor 103 cooperate to execute the material search program 110, thereby realizing the control unit 12 (see FIG. 1).

[0107] The storage unit 13 (see FIG. 1) is realized by the storage device 101 and the memory 102.

[0108] Furthermore, the communication interface 104 is hardware such as a network interface card (NIC) for connecting the material exploration apparatus 10 to a network. The communication interface 104 realizes the communication unit 11 (see FIG. 1).

[0109] The medium reading device 105 is hardware such as a CD drive, a DVD drive, or a USB interface for reading the recording medium 111 .

[0110] (Second embodiment) In the second embodiment, a water splitting device will be described that uses sulfides, nitrides, fluorides, chlorides, carbides, boron compounds, oxychlorides, and the like of the compounds discovered in the first embodiment as electrode catalyst materials.

[0111] 12 is a diagram showing an example of the configuration of a water decomposition apparatus according to this embodiment. As shown in FIG. 12, a water decomposition apparatus 60 includes dehumidifiers 61 and 62, an oxygen gas separation column 63, a hydrogen gas separation column 64, an electrolysis cell 65, a water tank 66, a cathode electrode 67, an anode electrode 68, and a power supply 69.

[0112] The electrolytic cell 65 is a container containing water 71 in which a cathode electrode 67 and an anode electrode 68 are immersed. Hydrogen is produced from the surface of the cathode electrode 67 and oxygen is produced from the surface of the anode electrode 68 by electrolysis of water.

[0113] The anode 68 has a current collecting electrode 68a, such as a titanium plate, and an electrode catalyst 68b formed on the surface thereof. The electrode catalyst material used for the electrode catalyst 68b is any one of sulfides, nitrides, fluorides, chlorides, carbides, boron compounds, and oxychlorides, which are compounds newly discovered in the first embodiment as having low overvoltages and the like. As an example, the compound is any one of the compounds ranked 1 to 20 in FIG. 9.

[0114] The power supply 69 is a DC power supply for applying a voltage between the cathode electrode 67 and the anode electrode 68 .

[0115] The oxygen gas separation tower 63 is a tower that separates the oxygen gas generated in the electrolytic cell 65 from the water vapor. Methods for separating oxygen gas from water vapor include a density difference separation method that separates water, which is a liquid with a high density, from oxygen gas, which has a low density, by using gravity or a collision force, and a gas-liquid separation method that utilizes the surface tension of the liquid.

[0116] The hydrogen gas separation column 64 is a column that separates the hydrogen gas generated in the electrolysis cell 65 from the water vapor by the aforementioned density difference separation method or gas-liquid separation method.

[0117] The water tank 66 recovers the water vapor separated in the oxygen gas separation tower 63 and the hydrogen gas separation tower 64, and supplies external water supplied through a pipe 66a to the electrolysis cell 65.

[0118] The dehumidifier 61 dehumidifies the oxygen gas (O2) separated in the oxygen gas separation tower 63 and supplies it to the outside. In addition, the dehumidifier 62 dehumidifies the hydrogen gas (H2) separated in the hydrogen gas separation tower 64 and supplies it to the outside as green hydrogen.

[0119] According to the water splitting device 60 described above, various compounds newly discovered in the first embodiment as having low overvoltage and the like are used as the electrode catalyst 68b of the anode electrode 68. As a result, a sufficient amount of green hydrogen can be produced even when the voltage of the power source 69 is low, thereby achieving energy savings in the water splitting device 60. Furthermore, as described in the first embodiment, the electrode catalyst 68b does not contain expensive platinum group elements, thereby achieving low costs for the water splitting device 60. This allows for the inexpensive production of large amounts of green hydrogen, a clean energy source that can replace fossil fuels, and makes it possible to address global environmental issues such as global warming.

[0120] In this example, the compound discovered in the first embodiment is applied to the electrode catalyst 68b of the water decomposition device 60, but the application of the compound is not limited to this. For example, the compound discovered in the first embodiment may be applied as an electrode catalyst in a device that mixes carbon dioxide with water 71 to reduce carbon dioxide and produce methanol and acetate. In this case, too, the electrolysis reaction proceeds at a low voltage, making it possible to produce methanol, acetate, and the like with low energy. [Explanation of symbols]

[0121] 10...material exploration device, 11...communication unit, 12...control unit, 13...memory unit, 14...experimental data, 15...training data, 16...machine learning model, 18...periodic table, 19...periodic table descriptor, 19a...first array, 19b...second array, 21...acquisition unit, 22...correlation analysis unit, 23...selection unit, 24...machine learning unit, 25...prediction unit, 26...list creation unit, 27...duplicate confirmation unit, 28...presentation unit, 29...input Reception unit, 30...display device, 40...input device, 51...candidate compound list, 51a to 51f...compounds, 52...duplicate compound list, 60...water decomposition device, 61, 62...dehumidifier, 63...oxygen gas separation tower, 64...hydrogen gas separation tower, 65...electrolysis cell, 66...water tank, 66a...piping, 67...cathode electrode, 68...anode electrode, 68a...collecting electrode, 68b...electrode catalyst, 69...power source, 71...water.

Claims

1. predicting a predicted value of an evaluation value for each of a plurality of compounds by inputting descriptors representing the compounds into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models, in which the plurality of compounds are arranged in order of magnitude of the predicted value; presenting the compounds that appear at the top of the multiple lists; A materials search program for causing a computer to execute a process including the steps of:

2. 2. The material exploration program according to claim 1, wherein the descriptor includes a sum of composition ratios of elements contained in the compound for each group, and an average value of any of the electronegativity, atomic number, atomic weight, electron affinity, ionization energy, neutron scattering cross section or absorption cross section, and van der Waals radius of the elements in the group.

3. 3. The material search program according to claim 1, wherein the processing further comprises selecting, as the descriptor, a descriptor that has a correlation with the evaluation value.

4. 4. The material search program according to claim 3, wherein the process further comprises selecting, as the evaluation value, an evaluation value that is correlated with another evaluation value of the compound.

5. 5. The material exploration program according to claim 1, wherein the evaluation value is any one of overpotential, current density, and Tafel gradient.

6. 6. The material search program according to claim 1, wherein the processing further comprises presenting the compound by the number of times that the compound appears at the top of the plurality of lists.

7. 7. The material exploration program according to claim 1, wherein the processing further comprises presenting an average value across the plurality of lists of the predicted values of the compounds that appear at the top of the plurality of lists.

8. 8. The material exploration program according to claim 1, wherein the learning algorithm is any one of Gaussian process regression using a radial basis function as a kernel function, Gaussian process regression using a rational quadratic kernel as a kernel function, NGBoost, XGBoost, and random forest.

9. 9. The material search program according to claim 1, wherein the higher ranking is a ranking higher than a predetermined ranking in each of the plurality of lists.

10. predicting a predicted value of an evaluation value of each of a plurality of compounds by inputting descriptors representing the compounds into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models, in which the plurality of compounds are arranged in order of magnitude of the predicted value; presenting the compounds that appear at the top of the multiple lists; A material exploration method in which a computer executes processes including the steps of:

11. predicting a predicted value of an evaluation value of each of a plurality of compounds by inputting descriptors representing the compounds into each of a plurality of machine learning models having different learning algorithms; creating a plurality of lists for each of the machine learning models, in which the plurality of compounds are arranged in order of magnitude of the predicted value; presenting the compounds that appear at the top of the multiple lists; A material exploration apparatus having a control unit that executes processing including:

12. Mn ０．１ Fe ０．１ Co ０．１ A0 ０．１ W ０．６ 、 ０．１ Fe ０．１ Co ０．１ Mo ０．６ A0 ０．１ 、 ０．１ Fe ０．１ Co ０．１ Cu ０．１ W ０．６ 、Fe ０．１ Co ０．１ A0 ０．１ W ０．７ 、Fe ０．１ Co ０．１ A0 ０．２ W ０．６ 、Fe ０．１ Co ０．２ A0 ０．１ W ０．６ 、Fe ０．２ Co ０．１ A0 ０．１ W ０．６ 、Fe ０．１ Co ０．１ Mo ０．１ A0 ０．１ W ０．６ 、Fe ０．１ Co ０．１ Mo ０．１ A0 ０．２ W ０．５ 、Fe ０．１ Co ０．２ Mo ０．１ A0 ０．１ W ０．５ 、Fe ０．２ Co ０．１ Mo ０．１ A0 ０．１ W ０．５ 、Fe ０．１ Co ０．１ Mo ０．２ A0 ０．１ W ０．５ 、Fe ０．１ Co ０．１ Mo ０．２ A0 ０．２ W ０．４ 、Fe ０．１ Co ０．２ Mo ０．２ A0 ０．１ W ０．４ 、Fe ０．２ Co ０．１ Mo ０．２ A0 ０．１ W ０．４ 、Fe ０．１ Co ０．１ Mo ０．３ A0 ０．１ W ０．４ 、Fe ０．１ Co ０．１ Mo ０．３ A0 ０．２ W ０．３ 、Fe ０．１ Co ０．２ Mo ０．３ A0 ０．１ W ０．３ 、Fe ０．２ Co ０．１ Mo ０．３ A0 ０．１ W ０．３ 、Fe ０．１ Co ０．１ Mo ０．４ A0 ０．１ W ０．３ 、Fe ０．１ Co ０．２ Mo ０．４ A0 ０．１ W ０．２ 、Fe ０．２ Co ０．１ Mo ０．４ A0 ０．１ W ０．２ 、Fe ０．１ Co ０．１ Mo ０．４ A0 ０．２ W ０．２ 、Fe ０．１ Co ０．１ Mo ０．５ A0 ０．１ W ０．２ 、Fe ０．１ Co ０．１ Mo ０．５ A0 ０．２ W ０．１ 、Fe ０．１ Co ０．２ Mo ０．５ A0 ０．１ W ０．１ 、Fe ０．２ Co ０．１ Mo ０．５ A0 ０．１ W ０．１ 、Fe ０．１ Co ０．１ Mo ０．６ Ag ０．１ W ０．１ , Mn ０．１ Fe ０．１ Co ０．１ Cu ０．１ Mo ０．６ , Fe ０．１ Co ０．１ Mo ０．６ Ag ０．２ , Fe ０．１ Co ０．１ Mo ０．７ Ag ０．１ , Fe ０．１ Co ０．２ Mo ０．６ Ag ０．１ , Fe ０．２ Co ０．１ Mo ０．６ Ag ０．１ , Cr ０．１ Fe ０．１ Co ０．１ Ag ０．１ W ０．６ , Cr ０．１ Fe ０．１ Co ０．１ Ag ０．２ W ０．５ , Cr ０．１ Fe ０．１ Co ０．２ Ag ０．１ W ０．５ , Cr ０．１ Fe ０．２ Co ０．１ Ag ０．１ W ０．５ , and Fe ０．１ Co ０．１ Cu ０．１ Ag ０．１ W ０．６ An electrocatalytic material comprising a compound selected from the group consisting of:

Images