Material data processing apparatus, material data processing method, program, and method for manufacturing magnet

Abstract

To provide a material data processing apparatus and the like that allow analysis of material considering information on the structure of material.SOLUTION: A material data processing apparatus 1 comprises: a model candidate generation unit 23 that, for each of candidates 70 for the combination of structure variables of material, generates candidates 80 for a structure-characteristic model that predicts the characteristics of the material from the structure variables included in the candidate; and a model selection unit that selects, from the candidates 80 for the structure-characteristic model generated by the model candidate generation unit 23, a candidate 80 based on the combination of structure variables determined to have the highest prediction accuracy, and employs the selected candidate as a structure-characteristic model 8 to be used for analysis.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a materials data processing device, a materials data processing method, a program, and a magnet manufacturing method. [Background technology]

[0002] In recent years, the technological field of materials informatics, which effectively utilizes information science, particularly data science, to develop new materials, has been attracting attention. In materials informatics, data on various experimental conditions and results is associated and stored in a database, and information useful for developing new materials is extracted using statistical analysis, machine learning, simulation, etc.

[0003] For example, Patent Document 1 describes a technology for learning a regression model by relating process variables such as material composition conditions and heat treatment conditions to material properties, and predicting material properties from the material process variables using the regression model. The method for learning and generating a regression model as disclosed in Patent Document 1 is highly accurate and is considered to be easily applicable in the field of material design, which is based on numerical data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6950119 Summary of the Invention [Problem to be solved by the invention]

[0005] However, there is a limit to how much the prediction accuracy can be improved by the method of making predictions using a regression model trained based only on the process variables and properties of the material, as in Patent Document 1. Therefore, in order to achieve more accurate predictions, it is conceivable to take into account, for example, information about the structure of the material.

[0006] The present invention has been made in view of the above-mentioned problems, and has as its object to provide a materials data processing device and the like that enables analysis of materials taking into account the structural information of the materials. [Means for solving the problem]

[0007] A first invention for solving the above-mentioned problems is a materials data processing device comprising: a model candidate generation unit that learns and generates, for each candidate combination of microstructural variables of a material, candidate structure-property models that predict the properties of the material from each of the microstructural variables included in the candidate; and a model selection unit that selects, from the candidate structure-property models generated by the model candidate generation unit, a candidate based on the combination of microstructural variables with the highest prediction accuracy, and adopts this candidate as the structure-property model to be used for analysis.

[0008] The first invention may further comprise a structure data generation unit that generates virtual structure data, which is virtual data for each structure variable; a property prediction unit that inputs the virtual structure data generated by the structure data generation unit into the structure-property model and predicts the properties of the material; and a structure target setting unit that sets targets for each structure variable that satisfy desired properties based on the predictions by the property prediction unit.

[0009] The system may further include a process data generation unit that generates virtual process data, which is virtual data for each process variable, and a process selection unit that selects, from the virtual process data generated by the process data generation unit, a process variable that satisfies the goal set by the organizational goal setting unit.

[0010] The system may further include a process-organization model generation unit that learns and generates a process-organization model that predicts each organizational variable from each process variable, and constructs a model that links the process-organization model and the organization-characteristic model.

[0011] The system may further include a process data generation unit that generates virtual process data, which is virtual data for each process variable; a texture prediction unit that inputs the virtual process data generated by the process data generation unit into the process-texture model and predicts texture variables of the material; a property prediction unit that inputs predicted data of the texture variables predicted by the texture prediction unit into the texture-property model and predicts material properties; a texture selection unit that selects texture variables that satisfy desired properties from the predicted data of the texture variables based on the predictions by the property prediction unit; and a process selection unit that selects process variables that correspond to the texture variables selected by the texture selection unit from the virtual process data.

[0012] In the first invention, for example, the material is a magnet, and at least one of the candidate combinations of structural variables is the main phase proportion of the magnet, the Curie temperature of the main phase, and the ratio c / a of the lattice constants in the c-axis direction and the a-axis direction of the main phase.Alternatively, the material is a ferrite magnet, and at least one of the candidate combinations of structural variables may be the compound phase proportion having a magnetoplumbite structure of the ferrite magnet, the Curie temperature of the compound phase, and the ratio c / a of the lattice constants in the c-axis direction and the a-axis direction of the compound phase.

[0013] A second invention is a materials data processing method including: a model candidate generation step in which a computer learns and generates, for each candidate combination of microstructural variables of a material, candidate structure-property models that predict the properties of the material from each of the microstructural variables included in the candidate; and a model selection step in which a candidate structure-property model based on the combination of microstructural variables with the highest predictive accuracy is selected from the candidate structure-property models generated by the model candidate generation step, and adopted as the structure-property model to be used in analysis.

[0014] In the second invention, the method may further include the following steps: a process-structure model generation step of learning and generating a process-structure model that predicts each of the structure variables included in the combination of structure variables with the highest prediction accuracy from each process variable; a process data generation step of generating virtual process data that is virtual data for each process variable; a structure prediction step of inputting the generated virtual process data into the process-structure model and predicting the structure variables of the material; a property prediction step of inputting the predicted data of the structure variables into the structure-property model and predicting the properties of the material; a structure selection step of selecting, from the predicted data of the structure variables, a structure variable that satisfies a desired property based on the predicted data of the property; and a process selection step of selecting, from the virtual process data, a process variable that corresponds to the selected structure variable, and determining and outputting the process variable as an optimal process.

[0015] The method may further include a process data generation step of generating virtual process data, which is virtual data for each process variable; a texture data generation step of generating virtual texture data, which is virtual data for each texture variable; a property prediction step of inputting the generated virtual texture data into the texture-property model to predict material properties; a texture target setting step of setting targets for each texture variable that satisfy desired properties based on the predicted property data; and a process selection step of selecting process variables that satisfy the set targets from the generated virtual process data, and determining and outputting the process variables as an optimal process.

[0016] A third invention is a program that causes a computer to function as: a model candidate generation unit that learns and generates, for each candidate combination of microstructural variables of a material, candidate structure-property models that predict the properties of the material from each of the microstructural variables included in the candidate combination; and a model selection unit that selects, from the candidate structure-property models generated by the model candidate generation unit, a candidate based on the combination of microstructural variables with the highest prediction accuracy, and adopts the candidate as the structure-property model to be used in analysis.

[0017] The fourth invention is a method for manufacturing a magnet, which manufactures a magnet based on the optimal process output by the second invention. For example, the magnet is a ferrite magnet. [Effects of the Invention]

[0018] The present invention makes it possible to analyze a material taking into account its structural information. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is a diagram illustrating an example of the hardware configuration of a computer used as a material data processing device. [Figure 2] FIG. 1 is a block diagram showing the overall configuration of a material data processing device. [Figure 3] FIG. 2 is a diagram illustrating an example of a functional configuration of a tissue-characteristic model generation unit. [Figure 4] FIG. 10 is a diagram illustrating setting of candidates for combinations of organizational variables. [Figure 5] FIG. 10 is a diagram illustrating a model candidate for a tissue-characteristic model. [Figure 6] FIG. 10 is a diagram illustrating the selection of a texture-property model. [Figure 7] FIG. 1 is a diagram illustrating a process-organization model. [Figure 8] FIG. 1 illustrates an overview of the process for predicting material properties from process variables. [Figure 9] FIG. 2 illustrates an example of a functional configuration of an optimal process determination unit. [Figure 10] 10 is a flowchart illustrating an example of a processing flow of a material data processing method. [Figure 11] FIG. 10 is a diagram showing examples of candidate combinations of structural variables of a ferrite calcined body. [Figure 12] FIG. 10 is a diagram showing examples of candidates for a structure-property model for predicting the properties (saturation magnetization) of a ferrite calcined body. [Figure 13] 1 is a graph showing the predictive accuracy of candidate tissue-property models. [Figure 14]FIG. 10 is a diagram illustrating an example of a generated process-organization model. [Figure 15] 10 is a graph comparing the prediction accuracy of process-organization models using regression methods. [Figure 16] FIG. 10 is a diagram illustrating an example of a predictive model that links a process-organization model and an organization-characteristic model. [Figure 17] FIG. 1 is a diagram showing an outline of data and processing for determining an optimum process for calcined ferrite bodies. [Figure 18] FIG. 10 is a diagram illustrating another example of the functional configuration of the optimal process determination unit. [Figure 19] 10 is a flowchart showing another example of the processing flow of the material data processing method. [Figure 20] FIG. 1 is a diagram showing an outline of data and processing for determining an optimum process for calcined ferrite bodies. [Figure 21] This is a graph comparing the search performance of each method for the optimal process based on data from the initial stage of the experiment. [Figure 22] 1 is a flowchart showing an example of a manufacturing process for a sintered ferrite magnet. [Figure 23] FIG. 1 is a diagram illustrating an overview of a conventional method for predicting material properties from process variables. DETAILED DESCRIPTION OF THE INVENTION

[0020] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the embodiments of the present invention, a calcined body (ferrite calcined body) of a ferrite magnet (ferrite sintered magnet), which is a type of sintered magnet, will be used as an example of the material, but the material is not limited to a sintered magnet. For example, various materials including alloys, ceramics, resins, etc. can also be used. The term "calcined ferrite sintered magnet" refers to a compound (a ferrite compound with a hexagonal magnetoplumbite (M-type) structure) obtained by a solid-phase reaction when raw material powders (CaCO3 powder, La(OH)3 powder, Fe2O3 powder, ZnO powder, etc.) are mixed and then heated at a predetermined temperature. Furthermore, "having a hexagonal magnetoplumbite (M-type) structure" means that when X-ray diffraction of the ferrite calcined body is measured under general conditions, the X-ray diffraction pattern observed is primarily that of a hexagonal magnetoplumbite (M-type) structure.

[0021] [First embodiment] (Hardware configuration of materials data processing device 1) FIG. 1 is a diagram showing an example of the hardware configuration of a materials data processing apparatus 1 according to an embodiment of the present invention. For example, when a general personal computer is used as the materials data processing apparatus 1, the materials data processing apparatus 1 includes a control unit 101, a memory unit 102, a communication unit 103, an input unit 104, a display unit 105, a peripheral device I / F unit 106, and other components connected via a bus 108, as shown in FIG. 1. Note that the configuration shown in FIG. 1 is merely an example, and the materials data processing apparatus 1 can have various configurations depending on the application and purpose. Furthermore, the materials data processing apparatus 1 may be a multi-computer including multiple computers. It may also be a server-client system, a cloud server, or a virtual machine virtually constructed using software. In the following explanation, the materials data processing apparatus 1 will be described as a single computer.

[0022] The control unit 101 has a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The CPU loads programs stored in the storage unit 102 or a recording medium such as the ROM into a work memory area on the RAM and executes them, and drives and controls each unit connected via a bus 108 to realize each process of the material data processing device 1, which will be described later.

[0023] The ROM is a non-volatile memory that permanently stores programs such as the computer's boot program and BIOS, data, etc. The RAM is a volatile memory that temporarily stores programs and data loaded from the storage unit 102 or a recording medium such as a ROM, and also has a work area that the control unit 101 uses to perform various processes.

[0024] The storage unit 102 is a hard disk drive (HDD), a solid state drive (SSD), a flash memory, or the like, and stores programs executed by the control unit 101, data required for executing the programs, an operating system (OS), etc. For example, the storage unit 102 stores application programs for causing the materials data processing device 1 to execute the processes described below, learning data for learning a prediction model, learned machine learning models (process-structure model, structure-property model), etc.

[0025] The communication unit 103 includes a communication interface and a communication control circuit that mediate communication of the material data processing device 1, and controls communication via a network. The network may be a local area network (LAN), a wide area network (WAN), the Internet, or the like, and may be wired or wireless.

[0026] The input unit 104 includes input devices such as a keyboard, a mouse, or a touch panel, and various operation buttons, etc. The input unit 104 transmits input data and operation instructions to the control unit 101.

[0027] The display unit 105 includes a display such as a liquid crystal panel, and displays data such as images and text on the display in accordance with instructions from the control unit 101.

[0028] The peripheral device I / F unit 106 is a port for connecting peripheral devices, and includes short-range wireless communication such as USB and Bluetooth (registered trademark). The control unit 101 transmits and receives data to and from peripheral devices via the peripheral device I / F unit 106. A printer or the like is connected to the peripheral device I / F unit 106.

[0029] (Functional configuration of materials data processing device 1) Next, the functional configuration of the materials data processing device 1 will be described with reference to FIGS. 2 to 9. FIG. 2 is a block diagram showing the overall configuration of the materials data processing device 1 according to this embodiment. As shown in the figure, the materials data processing device 1 includes a structure-property model generation unit 20 that generates a structure-property model 8 that predicts material properties from the structure variables of the material, a process-structure model generation unit 30 that generates a process-structure model 5 that predicts the structure variables from the process variables of the material, and an optimal process determination unit 40 that determines an optimal process 10 for the material. These functions are realized by the control unit 101 of the materials data processing device 1 reading and executing programs stored in the storage unit 102 or ROM. Below, the functions of the structure-property model generation unit 20, the process-structure model generation unit 30, and the optimal process determination unit 40 will be described in detail.

[0030] 3 is a diagram showing an example of the functional configuration of the tissue-property model generation unit 20. As shown in FIG. 3, the tissue-property model generation unit 20 is made up of a tissue combination candidate setting unit 21, a model candidate generation unit 23, and a model selection unit 25.

[0031] The organizational combination candidate setting unit 21 sets organizational variable combination candidates 70 based on the organizational variable candidates 60 (see FIG. 4). The organizational variable combination candidates 70 are configured by any combination of variables selected from the organizational variable candidates 60 given in advance.

[0032] FIG. 4 is a diagram illustrating the setting of candidate combinations 70 of organizational variables. In the example of FIG. 4, A, B, C, and D are assumed to be given in advance as candidate combinations 60 of organizational variables. Then, {A, B} (candidate 71), {A, B, C} (candidate 72), and {A, B, C, D} (candidate 73) are set as candidate combinations 70 of organizational variables. The user can arbitrarily set the combinations of variables to be used as candidate combinations 70 of organizational variables. For example, although not shown in the example of FIG. 4, combinations such as {A, C}, {B, C}, and {B, C, D} can also be set.

[0033] For each candidate 70 of the combination of tissue variables set by the tissue combination candidate setting unit 21, the model candidate generation unit 23 uses the training data 2 to train and generate a candidate structure-property model 80 that predicts the properties of a material from each of the tissue variables included in the candidate 70. The training data 2 is a data set consisting of pairs of each tissue variable and material property obtained through past experiments and simulations, and is prepared for each candidate structure-property model 80 to be generated.

[0034] FIG. 5 illustrates the generated structure-property model candidates 80. FIG. 5 shows an example of the generation of structure-property model candidates 80 when the material property 9 ({X, Y, Z}) is the prediction target. Specifically, for the set of candidate structure variable combinations 70, i.e., {A, B} (candidate 71), {A, B, C} (candidate 72), and {A, B, C, D} (candidate 73), structure-property model candidates 80 are generated to predict X, Y, and Z, respectively. That is, as shown in FIG. 5 (bottom), model candidates 81 to 83 are generated for predicting X, Y, and Z from {A, B}, respectively; model candidates 84 to 86 are generated for predicting X, Y, and Z from {A, B, C}, respectively; and model candidates 87 to 89 are generated for predicting X, Y, and Z from {A, B, C, D}, respectively. Model candidates 81 to 89 are trained using a dataset consisting of pairs of structure variables and material properties corresponding to each model candidate.

[0035] The model selection unit 25 is a functional unit that selects a candidate 80 based on a combination of tissue variables that is determined to have the highest prediction accuracy from the tissue-property model candidates 80 generated by the model candidate generation unit 23, and adopts it as the tissue-property model 8 to be used in actual analysis. The model selection unit 25 includes a prediction accuracy calculation unit 251 and a tissue combination determination unit 252.

[0036] The prediction accuracy calculation unit 251 calculates the prediction accuracy for each of the structure-property model candidates 80 generated by the model candidate generation unit 23. The structure combination determination unit 252 selects and outputs the optimal structure variable combination 7 and structure-property model 8 with the highest prediction accuracy for each property 9 of the material from the structure variable combination candidate 70 and structure-property model candidate 80.

[0037] FIG. 6 is a diagram illustrating the selection of a structure-property model 8. FIG. 6 displays the results of the prediction accuracy of each material property 9 ({X, Y, Z}) for each of the structure-property model candidates 80 in FIG. 5 (the best one is displayed). For example, the optimal structure variable combination 7 that showed the highest prediction accuracy for property X is {A, B, C}. Therefore, the structure-property model candidate under these conditions (candidate 84 in FIG. 6) is selected as the optimal structure-property model 8 for predicting property X. Similarly, the optimal structure variable combination 7 that showed the highest prediction accuracy for property Y is {A, B}, and the structure-property model candidate in this case (candidate 82 in FIG. 6) is selected as the optimal structure-property model 8 for predicting property Y. Furthermore, the optimal structure variable combination 7 that showed the highest prediction accuracy for property Z is {A, B, C, D}, and the structure-property model candidate in this case (candidate 89 in FIG. 6) is selected as the optimal structure-property model 8 for predicting property Z.

[0038] In the example in Figure 6, the optimal combination of tissue variables and the optimal tissue-property model are selected for each of the properties X, Y, and Z. However, if you wish to reduce the number of models, you can consider selecting just one tissue-property model 8 that has the highest average prediction accuracy for each of the properties X, Y, and Z.

[0039] As described above, the structure-property model generation unit 20 determines and outputs the optimal combination of structure variables 7 (structure variables) for predicting material properties 9, and the optimal structure-property model 8 based on that combination. By using this structure-property model 8, it becomes possible to predict and analyze material properties 9 with high accuracy from data on any structure variables 7.

[0040] For simplicity of explanation, the following description will be limited to the case where one characteristic X is the target of prediction. In this case, the texture-characteristic model generation unit 20 outputs {A, B, C} as the optimal combination of texture variables 7 (textural variables) for predicting the characteristic X, and outputs a candidate 84 model as the optimal texture-characteristic model 8 (see FIG. 6).

[0041] Next, we will explain the function of the process-structure model generation unit 30 in Figure 2. The process-structure model generation unit 30 is a functional unit that learns and generates a process-structure model 5 that predicts each structure variable 7 from each process variable 4 of a material (see Figure 7). Here, we will explain the process variables 4 of a material as three variables {a, b, c}.

[0042] 2, the process-organization model generation unit 30 receives as input the optimal combination of organizational variables 7 (here, organizational variables {A, B, C}) output by the organization-characteristic model generation unit 20, and uses the learning data 3 to learn and generate a process-organization model 5 that predicts each organizational variable 7 ({A, B, C}) from each process variable 4 ({a, b, c}). The learning data 3 is a data set consisting of pairs of each process variable 4 ({a, b, c}) and each organizational variable 7 ({A, B, C}), obtained through past experiments and simulations.

[0043] FIG. 7 is a diagram illustrating the process-structure model 5. As shown in FIG. 7 (upper diagram), the process variables 4 of the material are {a, b, c}, and the structure variables 7 to be predicted are {A, B, C} (optimal structure variable combination 7) output by the structure-property model generation unit 20. As shown in FIG. 7 (lower diagram), the process-structure model generation unit 30 learns and generates the process-structure model 5 that predicts each of the structure variables 7 ({A, B, C}) from the process variables 4 ({a, b, c}).

[0044] FIG. 8 is a diagram showing an overview of the process for predicting a material property 9 from a process variable 4. First, a process-structure model 5 is used to predict a structure variable 7 from an arbitrary process variable 4. Then, a structure-property model 8 is used to predict the material property 9 from the predicted structure variable 7. In this way, in this embodiment, a model can be constructed that links the process-structure model 5 and the structure-property model 8. Conventionally, the material property 9 was predicted directly from the process variable 4, but in this embodiment, by incorporating the structure variable 7, the structure information of the material is reflected in the prediction process, which is expected to improve prediction accuracy.

[0045] Next, a description will be given of the function of the optimum process determination unit 40 in Fig. 2. The optimum process determination unit 40 is a functional unit that determines and outputs an optimum process 10, which is the optimum process condition for obtaining desired characteristics.

[0046] 9 is a diagram showing an example of the functional configuration of the optimum process determination unit 40. The optimum process determination unit 40 includes a process data generation unit 41, a texture prediction unit 42, a characteristic prediction unit 43, a texture selection unit 44, and a process selection unit 45.

[0047] The process data generator 41 generates virtual process data 11, which is virtual data for each process variable 4. The virtual process data 11 is data on process variables that are candidates for the optimal process 10 for obtaining desired characteristics. The virtual process data 11 is not actually measured data, but a large amount of data that is artificially generated using random values ​​or statistical techniques.

[0048] The structure prediction unit 42 inputs the virtual process data 11 generated by the process data generation unit 41 into the process-structure model 5, predicts the structure variables of each virtual process data 11, and outputs predicted data 13 of the structure variables.

[0049] The property prediction unit 43 inputs the predicted data 13 of the structure variables predicted and output by the structure prediction unit 42 into the structure-property model 8, predicts the properties of the material, and outputs predicted data 14 of the properties.

[0050] The tissue selection unit 44 uses Bayesian optimization based on the prediction data 14 predicted by the property prediction unit 43 and the like to select tissue variables that satisfy desired properties from the prediction data 13 of tissue variables.

[0051] The process selection unit 45 selects process variables corresponding to the organizational variables selected by the organizational selection unit 44 from the virtual process data 11 and outputs the selected process variables as the optimal process 10. In this way, an optimal process 10, which is the optimum process conditions for obtaining the desired properties, is determined and output. The optimal process 10 can be used as a candidate for the next experiment or adopted as the process conditions for actual material manufacturing.

[0052] (Processing of materials data processing device 1) Next, with reference to Figs. 10 to 17, a material data processing method executed by the material data processing device 1 will be described. Here, a specific processing flow will be described using as an example a calcined body of a ferrite magnet (ferrite sintered magnet), which is a type of sintered magnet (hereinafter, sometimes referred to as a "ferrite calcined body") as a material. In this embodiment, the target is a ferrite calcined body containing calcium (Ca), lanthanum (La), iron (Fe), and zinc (Zn). The composition of the ferrite calcined body is represented by the general formula: (Ca 1-y La y ) x Fe 12-z Zn z The composition ratios x, y, and z are the process variables 4. The saturation magnetization σ of the ferrite calcined body is s Let be the target characteristic 9. The general formula is expressed by the atomic ratio of the metal elements, but the composition containing oxygen (O) is expressed by the general formula: (Ca 1-y La y ) x Fe 2n-z Zn z O αThe number of moles of oxygen, α, is basically α = 19, but it varies depending on the Fe valence, the values ​​of x, y, and z, etc. Furthermore, the ratio of oxygen to metal elements changes due to oxygen vacancies (vacancies) when fired in a reducing atmosphere, changes in the Fe valence in the ferrite phase, etc. Therefore, the actual number of moles of oxygen, α, may deviate from 19. Therefore, in the present invention, the composition is expressed in terms of the atomic ratio of metal elements.

[0053] 10 is a flowchart showing an example of the processing flow of the materials data processing method. First, the control unit 101 (structure combination candidate setting unit 21) of the materials data processing device 1 sets structure variable combination candidates 70, which are arbitrary combinations of the structure variable candidates 60 of the ferrite magnet (step S1).

[0054] 11 is a diagram showing examples of candidate combinations 70 of structural variables of the ferrite calcined body set in step S1. In the example of FIG. 11, five candidates 60 of structural variables are given: "M-phase ratio," "Tc," "c," "a," and "c / a." The candidate combinations 70 of these variables are set as follows: {M-phase ratio, Tc} (candidate 70a), {M-phase ratio, Tc, c / a} (candidate 70b), and {M-phase ratio, Tc, c, a} (candidate 70c). The "M-phase ratio" refers to the ratio of a compound phase (M phase) having a magnetoplumbite structure in the ferrite calcined body (sintered ferrite magnet), "Tc" refers to the Curie temperature of the compound phase, "c" refers to the lattice constant of the c-axis of the crystal structure of the compound phase, "a" refers to the lattice constant of the a-axis of the crystal structure of the compound phase, and "c / a" refers to the axial ratio of the crystal structure.

[0055] Next, for each candidate 70 (candidates 70a to 70c) of the combination of structural variables set in step S1, the control unit 101 (model candidate generation unit 23) of the material data processing device 1 uses the learning data 2 to calculate the properties of the ferrite calcined body (saturation magnetization σ s) is learned and generated (step S2). The learning data 2 is a data set in which each of the structure variables corresponding to the candidates 70a to 70c is associated with actual data on the properties of the material.

[0056] As a method for learning the process-organization model, various regression methods can be used, such as Gaussian Process Regression (GPR), LightGBM (Light Gradient Boosting Machine: LGBM), Neural Network (NN), Random Forest (RF), etc. In this embodiment, Gaussian Process Regression (GPR) is used.

[0057] Figure 12 shows the properties of the calcined ferrite body (saturation magnetization σ s 12 shows an example of a candidate structure-property model 80 for predicting the saturation magnetization σ from each candidate structure-property model 70 (70a, 70b, 70c) set in FIG. s Specifically, a structure-property model candidate 80 is generated to predict the saturation magnetization σ from {M phase ratio, Tc} (candidate 70a). s Candidate 80a of the structure-property model predicts saturation magnetization σ from {M phase fraction, Tc, c / a} (candidate 70b). s Candidate 80b of the structure-property model predicts saturation magnetization σ from {M phase fraction, Tc, c, a} (candidate 70c). s A candidate tissue-property model 80c is generated that predicts

[0058] Next, the control unit 101 (model selection unit 25) of the materials data processing device 1 selects the structure-property model candidate 80 based on the combination of structure variables that is determined to have the highest prediction accuracy from the structure-property model candidate 80 generated in step S2, and adopts it as the structure-property model 8 to be used in the actual analysis (step S3). Specifically, the control unit 101 (prediction accuracy calculation unit 251) calculates the mean absolute error (MAE), which is the average of the absolute errors of the difference between the predicted value and the actual value, and the coefficient of determination (R 2 ) or other evaluation index for the prediction accuracy of the prediction model. Then, the control unit 101 (tissue combination determination unit 252) calculates an evaluation index for the prediction accuracy of the prediction model, such as the one that minimizes the mean absolute error (MAE) or the coefficient of determination (R 2 ) are selected and output as the optimal organizational variable combination 7 and organizational-characteristic model 8.

[0059] 13 is a graph showing the prediction accuracy of the candidate structure-property model 80 generated in step S2. Using 83 data sets obtained by carrying out four calcination experiments on ferrite calcined bodies, Gaussian process regression (GPR) was used to calculate the saturation magnetization σ under a 9-tesla magnetic field. s The candidate tissue-property models 80 (80a to 80b) for predicting the above were created. From the graph in Figure 13, the candidate tissue-property model 80b has the smallest mean absolute error (MAE) and the highest coefficient of determination (R 2 ) is the largest, it can be seen that this is the model with the highest prediction accuracy. Next, the structure-property model candidate 80c and the structure-property model candidate 80a have the highest prediction accuracy in this order. On the other hand, the results on the left side of Figure 13 show that the prediction accuracy is high in the case of the direct process variable 4 to property 9 (saturation magnetization σ s 23 shows the prediction accuracy of the conventional process-property model 16 (see FIG. 23) that predicts the structure-property model 80 (80a to 80c) of this embodiment, and it can be seen that the prediction accuracy is lower than that of any of the candidate structure-property models 80 (80a to 80c) of this embodiment. This suggests that higher prediction accuracy can be achieved by predicting properties from structure variables.

[0060] Here, based on the prediction accuracy results of FIG. 13, in step S3, the control unit 101 (structure combination determination unit 252) outputs the structure-property model candidate 80b with the highest prediction accuracy as the structure-property model 8 to be used in the analysis, and outputs the structure variable combination candidate 70b ({M-phase ratio, Tc, c / a}) at that time as the optimal structure variable combination 7.

[0061] By the above-described processing of steps S1 to S3, the properties of the ferrite calcined body (saturation magnetization σ s The optimal combination of structural variables 7 (structural variables) for predicting the structural property 9 is determined, and the optimal structural-property model 8 based on this combination is then determined. By using this structural-property model 8, it becomes possible to predict and analyze the material properties 9 with high accuracy from data on any structural variables 7.

[0062] Next, the control unit 101 (process-structure model generation unit 30) of the materials data processing device 1 receives as input the optimal combination of structure variables 7 (here, structure variables {M-phase ratio, Tc, c / a}) output in step S3, and uses the training data 3 to learn and generate a process-structure model 5 that predicts each structure variable 7 ({M-phase ratio, Tc, c / a}) from each process variable 4 ({x, y, z}) (step S4). The training data 3 is a data set in which actual data for each process variable 4 ({x, y, z}) and each structure variable 7 ({M-phase ratio, Tc, c / a}) are associated with each other.

[0063] FIG. 14 is a diagram showing the process-structure model 5 generated in step S4. As shown in FIG. 14, the ferrite calcined body (Ca 1-y La y ) x Fe 12-z Zn z A process-structure model 5 is trained and generated to predict structure variables 7 ({M phase ratio, Tc, c / a}) from process variables 4 ({x, y, z}), which are the composition ratios of the alloys. Various regression methods can be used to train the process-structure model 5, including Gaussian process regression (GPR), LightGBM (LGBM), neural networks (NN), and random forests (RF).

[0064] FIG. 15 is a graph comparing the prediction accuracy of the process-structure model 5 using regression methods. The graph shows the prediction accuracy of the process-structure model 5 trained using Gaussian process regression (GPR), LightGBM (LGBM), neural network (NN), and random forest (RF). FIG. 15(a) is a graph showing the prediction accuracy of the structure variable 7 (“M-phase ratio”), FIG. 15(b) is a graph showing the prediction accuracy of the structure variable 7 (“Tc”), and FIG. 15(c) is a graph showing the prediction accuracy of the structure variable 7 (“c / a”). It can be seen from FIG. 15 that relatively high prediction accuracy was achieved when Gaussian process regression (GPR) was used as the regression method. In this embodiment, Gaussian process regression (GPR), which had high prediction accuracy, was used to train the process-structure model 5.

[0065] By the above processing of step S4, a process-structure model 5 is generated that predicts the structural variable 7 ({M-phase ratio, Tc, c / a}) from the process variable 4 ({x, y, z}), which is the composition ratio of the ferrite calcined body. As a result, as shown in FIG. 16 , the process-structure model 5 predicts the structural variable 7 ({M-phase ratio, Tc, c / a}) from the process variable 4 ({x, y, z}), and the property 9 (saturation magnetization σ) of the ferrite calcined body is generated from the structural variable 7 ({M-phase ratio, Tc, c / a}). s ) can be constructed by linking the structure-property model 8 to predict the structure-property model 8.

[0066] Next, the control unit 101 (optimal process determination unit 40) of the material data processing device 1 uses the generated structure-property model 8 and process-structure model 5 to determine the optimal process 10 for the ferrite calcined body through the processing of the subsequent steps S5 to S9.

[0067] First, the control unit 101 (process data generation unit 41) of the material data processing apparatus 1 generates virtual process data 11, which is virtual data for each process variable 4 ({x, y, z}) (step S5). The virtual process data 11 is data on process variables that are candidates for the optimal process 10 for obtaining desired characteristics. The virtual process data 11 is not actually measured data, but a large amount of data that is artificially generated using random values ​​or statistical techniques. As shown in FIG. 17, the generated virtual process data 11 holds data on the process variables 4 ({x, y, z}) linked to an ID 111 that uniquely identifies the data.

[0068] Next, the control unit 101 (structure prediction unit 42) of the materials data processing device 1 inputs the virtual process data 11 generated in step S5 into the process-structure model 5, predicts the structure variables 7 ({M-phase ratio, Tc, c / a}) of each virtual process data 11, and outputs predicted data 13 of the structure variables (step S6, (1) in FIG. 17). As shown in FIG. 17, the predicted data 13 of the structure variables holds the data of the predicted structure variables 7 ({M-phase ratio, Tc, c / a}) linked to the same ID 111 as the virtual process data 11.

[0069] Next, the control unit 101 (property prediction unit 43) of the material data processing device 1 inputs the predicted data 13 of the structural variables predicted in step S6 into the structure-property model 8, and calculates the properties of the ferrite calcined body (saturation magnetization σ s ) and output the predicted data 14 of the characteristics (step S7, (2) of FIG. 17). As shown in FIG. 17, the predicted data 14 of the characteristics is linked to the same ID 111 as the virtual process data 11 and includes the predicted characteristics (saturation magnetization σ s ) data is retained.

[0070] Next, the control unit 101 (structure selection unit 44) of the materials data processing device 1 uses Bayesian optimization based on the prediction data 14 predicted in step S7, etc., to select structure variables that satisfy the desired properties from the prediction data 13 of structure variables (step S8, (3) in Figure 17).

[0071] Specifically, the control unit 101 (tissue selection unit 44) selects tissue variables that maximize the acquisition function calculated by Bayesian optimization. The acquisition function here may be, for example, a characteristic (saturation magnetization σ s ) is maximized. This allows the characteristics (saturation magnetization σ s ) are searched for and selected. For example, as shown in FIG. 17, an organizational variable 7i (organization variable 7 with ID "i") is selected at this stage from the prediction data 13 output in step S7 as an organizational variable that satisfies the desired characteristics.

[0072] In this embodiment, Bayesian optimization is used in step S8 of Fig. 10, but the use of Bayesian optimization is not essential. For example, the desired characteristics (for example, saturation magnetization σ) can be calculated by referring to the predicted data 14 of the characteristics obtained in step S7. s It is also possible to select one or more pieces of data that satisfy the condition (is equal to or greater than a predetermined threshold value), and select the organizational variables corresponding to the selected data from the predicted data 13 of the organizational variables.

[0073] Then, the control unit 101 (process selection unit 45) of the material data processing device 1 selects a process variable corresponding to the organizational variable selected in step S8 from the virtual process data 11 generated in step S5, and determines and outputs the process variable as the optimal process 10 (step S9, (4) in FIG. 17). In the example in FIG. 17, the process variable 4i (process variable 4 with ID "i") corresponding to the organizational variable 7i (organization variable 7 with ID "i") selected in step S8 is selected from the virtual process data 11, and this process variable 4i becomes the optimal process 10. By the above processing of steps S5 to S9, the optimum process 10 for obtaining the desired characteristics is determined.

[0074] As described above, the materials data processing apparatus 1 according to the first embodiment generates, for each candidate combination 70 of microstructural variables of a material, candidate structure-property models 80 that predict the material properties from the respective microstructural variables included in each candidate. Then, from the generated candidate structure-property models 80, a candidate 80 based on a combination of microstructural variables that is determined to have the highest prediction accuracy is selected and adopted as the structure-property model 8 to be used in the analysis. This makes it possible to determine the optimal combination 7 of microstructural variables (microstructural variables) and structure-property model 8 for predicting the material properties 9. Furthermore, it becomes possible to predict and analyze the material properties 9 with high accuracy from arbitrary data on the microstructural variables 7 using the structure-property model 8.

[0075] The materials data processing device 1 also learns and generates a process-structure model 5 that predicts each microstructural variable from each process variable. As a result, a prediction model is constructed that links the process-structure model 5 and the microstructural-property model 8, as shown in Figure 8. In conventional methods, material properties 9 are predicted directly from process variables 4 (see Figure 23), but in this embodiment, by incorporating a microstructural variable 7, the microstructural information of the material is reflected in the prediction process, which is expected to improve prediction accuracy.

[0076] Furthermore, the materials data processing device 1 determines an optimal process 10 for the material using the generated structure-property model 8 and process-structure model 5. Specifically, the materials data processing device 1 generates virtual process data 11, inputs it into the process-structure model 5, and outputs prediction data 13 that predicts the structure variables for each virtual process data 11. The predicted data 13 for the structure variables is also input into the structure-property model 8 to predict the material properties. Based on the prediction results, the materials data processing device 1 then selects a structure variable that satisfies the desired properties from the predicted data 13 for the structure variables, and further selects a process variable corresponding to the selected structure variable from the virtual process data 11, and determines and outputs the process variable as the optimal process 10. This determines the optimal process 10 for obtaining the desired properties, which can be used as a candidate for the next experiment or adopted as a process condition for an actual material.

[0077] [Second embodiment] Next, a second embodiment will be described. In the first embodiment, predicted data 13 of organizational variables was generated using a process-organization model 5 based on virtual process data 11, and optimal organizational variables 7 that satisfy desired characteristics were selected from the predicted data 13. However, the predicted data 13 depends on the virtual process data 11 and the process-organization model 5, and depending on conditions such as the data range of the generated virtual process data 11, prediction of organizational variables that achieve high characteristics may be insufficient. Therefore, the organizational variables selected based on the predicted data 13 may not necessarily be optimal, and the corresponding process variables 4 in the virtual process data 11 may not necessarily be optimal either.

[0078] Therefore, in the second embodiment, a large amount of virtual data (virtual organization data 12) of organizational variables is generated, organizational variable targets 15 that satisfy desired characteristics are set based on the virtual organization data 12, and process variables 4 that satisfy the organizational variable targets 15 are selected from the virtual process data 11. Because the virtual organization data 12 is data that is generated independently of the virtual process data 11 and the process-organization model 5, it is also possible to intentionally include organizational variables that achieve high characteristics. As a result, it becomes possible to evaluate a wider range of organizational variables, increasing the possibility of discovering optimal process variables 4 that are often overlooked due to the constraints of the virtual process data 11 and the process-organization model 5.

[0079] (Functional configuration of material data processing device 1a) First, the functional configuration of a material data processing device 1a according to the second embodiment will be described. The overall functional configuration of the material data processing device 1a is the same as that of the first embodiment (FIG. 2), but in the second embodiment, the optimal process determination unit 40 in FIG. 2 is replaced with an optimal process determination unit 50. The hardware configuration of the material data processing device 1a is the same as that of the first embodiment (FIG. 1).

[0080] 18 is a diagram illustrating an example of the functional configuration of the optimal process determination unit 50 in the second embodiment. The optimal process determination unit 50 includes a process data generation unit 51, an organizational data generation unit 52, a characteristic prediction unit 53, an organizational goal setting unit 54, and a process selection unit 55.

[0081] As in the first embodiment, the process data generation unit 51 is a functional unit that generates virtual process data 11, which is virtual data for each process variable 4. The virtual process data 11 is data on process variables that are candidates for the optimal process 10 for obtaining desired characteristics. The virtual process data 11 is not actually measured data, but a large amount of data that is artificially generated using random values ​​or statistical techniques.

[0082] The tissue data generator 52 generates virtual tissue data 12, which is virtual data for each tissue variable. Like the virtual process data 11, the virtual tissue data 12 is not actually measured data, but a large amount of data artificially generated using random values ​​or statistical techniques. In order to generate a wider range of tissue variables that cannot be covered by predictions based on the virtual process data 11, it is desirable to generate a larger amount of virtual tissue data 12 than the number of pieces of data in the virtual process data 11.

[0083] The property prediction unit 53 inputs the virtual structure data 12 generated by the structure data generation unit 52 into the structure-property model 8, predicts the material properties of each virtual structure data 12, and outputs predicted property data 14.

[0084] The organizational target setting unit 54 sets targets 15 of organizational variables that satisfy desired characteristics, using Bayesian optimization (first time) based on the prediction data 14 predicted by the characteristic prediction unit 53 and the like.

[0085] The process selection unit 55 uses Bayesian optimization (second time) to select process variables that satisfy the organizational variable targets 15 set by the organizational target setting unit 54 from the virtual process data 11 generated by the process data generation unit 51, and determines and outputs the process variables as the optimal process 10. In this way, an optimal process 10, which is the optimum process conditions for obtaining the desired properties, is determined and output. The optimal process 10 can be used as a candidate for the next experiment or adopted as the process conditions for actual material manufacturing.

[0086] (Processing of material data processing device 1a) Next, a material data processing method executed by the material data processing device 1a will be described with reference to FIGS. 19 and 20. As in the first embodiment, a ferrite calcined body is used as an example of a material. That is, a ferrite calcined body containing calcium (Ca), lanthanum (La), iron (Fe), and zinc (Zn) is used as the target. 1-y La y ) x Fe 12-z Zn z The composition ratios x, y, and z are defined as process variables 4. The saturation magnetization σ of the ferrite calcined body is s Let be the target characteristic 9.

[0087] FIG. 19 is a flowchart showing an example of the processing flow of the materials data processing method. The processing of steps S11 to S14 in FIG. 19 is the same as steps S1 to S4 in the first embodiment (FIG. 10). That is, based on the candidates 60 of structural variables, the control unit 101 (structure combination candidate setting unit 21) of the materials data processing device 1a sets candidate structural variable combinations 70, which are arbitrary combinations of these variables (step S11). Next, for each candidate structural variable combination 70 set in step S11, the control unit 101 (model candidate generation unit 23) of the materials data processing device 1a uses the learning data 2 to calculate the properties of the ferrite calcined body (saturation magnetization σ ) from each structural variable included in the candidate 70. s ) is trained and generated (step S12).

[0088] The control unit 101 (model selection unit 25) of the materials data processing device 1a then selects the candidate structure-property model 80 based on the combination of structure variables determined to have the highest prediction accuracy from the candidate structure-property models 80 generated in step S12, and adopts it as the structure-property model 8 to be used in actual analysis (step S13). Next, the control unit 101 (process-structure model generation unit 30) of the materials data processing device 1a uses the learning data 3 to learn and generate a process-structure model 5 that predicts each structure variable 7 ({M-phase fraction, Tc, c / a}) from each process variable 4 ({x, y, z}) based on the optimal combination of structure variables 7 (here, structure variables {M-phase fraction, Tc, c / a}) output in step S13 (step S14).

[0089] In the second embodiment, the control unit 101 (optimum process determination unit 50) of the material data processing device 1a determines the optimum process 10 for the ferrite calcined body by the processes of the following steps S15 to S19.

[0090] First, the control unit 101 (process data generation unit 51) of the material data processing device 1a generates virtual process data 11, which is virtual data for each process variable 4 ({x, y, z}) (step S15). The virtual process data 11 is data on process variables that are candidates for the optimal process 10 for obtaining desired characteristics. The virtual process data 11 is not actually measured data, but a large amount of data that is artificially generated using random values ​​or statistical techniques. As shown in FIG. 20, the generated virtual process data 11 holds data on the process variables 4 ({x, y, z}) linked to an ID 111 that uniquely identifies the data.

[0091] Next, the control unit 101 (structure data generating unit 52) ​​of the material data processing device 1a generates virtual structure data 12, which is virtual data for each structure variable 7 ({M-phase ratio, Tc, c / a}) (step S16). Like the virtual process data 11, the virtual structure data 12 is not actually measured data, but is a large amount of data artificially generated using random values ​​or statistical techniques. As shown in FIG. 20, the generated virtual structure data 12 holds data for the structure variables 7 ({M-phase ratio, Tc, c / a}) linked to an ID 121 that uniquely identifies the data.

[0092] Next, the control unit 101 (property prediction unit 53) of the material data processing device 1a inputs the virtual structure data 12 generated in step S16 into the structure-property model 8, predicts the material properties of each virtual structure data 12, and outputs predicted property data 14 (step S17, (1) in FIG. 20). As shown in FIG. 20, the predicted property data 14 is linked to the same ID 121 as the virtual structure data 12 and stores the predicted property (saturation magnetization σ s ) data is retained.

[0093] Next, the control unit 101 (structure target setting unit 54) of the material data processing device 1a uses Bayesian optimization (first time) based on the predicted data 14 predicted in step S17, etc., to set a structure variable target 15 that satisfies the desired characteristics (step S18, (2) in Figure 20).

[0094] Specifically, the control unit 101 (tissue target setting unit 54) sets the target 15 of the tissue variable that maximizes the acquisition function calculated by Bayesian optimization. The acquisition function here may be, for example, a characteristic (saturation magnetization σ s ) is maximized. This allows the characteristics (saturation magnetization σ s) are searched for and selected, and the targets 15 of the organizational variables are determined. For example, as shown in FIG. 20, from the virtual organizational data 12 generated in step S16, the top K (K is, for example, about 10) virtual organizational data 12K of organizational variables that satisfy the desired characteristics are searched for and selected, and the targets 15 of the organizational variables are set based on this. For example, the targets 15 of the organizational variables, i.e., the target values ​​of the "M-phase ratio," "Tc," and "c / a," are set from the ranges or average values ​​of the "M-phase ratio," "Tc," and "c / a" of the virtual organizational data 12K.

[0095] Then, the control unit 101 (process selection unit 55) of the material data processing device 1a uses Bayesian optimization (second time) to select one or more process variables that satisfy the target 15 of the organizational variables set in step S18 from the virtual process data 11 generated in step S15, and determines and outputs the process variables as the optimal process 10 (step S19, (3) in Figure 20).

[0096] Specifically, the control unit 101 (process selection unit 55) selects process variables that maximize the acquisition function calculated by Bayesian optimization. For example, the acquisition function used here is PTR (Probability in Target Range), which increases the probability that the value of the organizational variable falls within the target 15 of the organizational variable. This allows efficient search and selection of process variables that have a high probability that the value of the organizational variable falls within the target 15 of the organizational variable. In the example of FIG. 20, process variable 4j (process variable 4 with ID "j") is selected from the virtual process data 11 generated in step S15 as a process variable that satisfies the desired characteristics, and this process variable 4j becomes the optimal process 10. By the above processing of steps S15 to S19, the optimum process 10 for obtaining the desired characteristics is determined.

[0097] In this embodiment, Bayesian optimization is used in steps S18 and S19, but the use of Bayesian optimization is not essential. For example, in step S18, the desired characteristics (for example, saturation magnetization σ) are calculated by referring to the predicted data 14 of the characteristics obtained in step S17.s It is also possible to select one or more pieces of data that satisfy the condition (is equal to or greater than a predetermined threshold) and set a target 15 for the organizational variable based on the value in the predicted data 13 for the organizational variable that corresponds to the selected data.

[0098] Similarly, in step S19, one or more organizational variables that satisfy the organizational variable targets 15 set in step S18 can be selected from the predicted data of organizational variables predicted from the virtual process data 11 using the process-organization model 5, and process variables corresponding to the selected organizational variables can be selected from the virtual process data 11.

[0099] As described above, the materials data processing apparatus 1a according to the second embodiment differs from the first embodiment in the method for determining the optimal process 10. Specifically, the materials data processing apparatus 1a generates virtual structure data 12 in addition to virtual process data 11. The materials data processing apparatus 1a then inputs the generated virtual structure data 12 into the structure-property model 8, predicts the material properties of each virtual structure data 12, and sets targets 15 for structure variables that satisfy the desired properties. The materials data processing apparatus 1a then selects one or more process variables from the virtual process data 11 that satisfy the determined targets 15 for structure variables, and determines and outputs the process variables as the optimal process 10. This determines the optimal process 10 for obtaining the desired properties, which can then be used as a candidate for the next experiment or adopted as process conditions for an actual material.

[0100] In particular, in the second embodiment, by generating virtual structure data 12 and inputting it into the structure-property model 8, it is possible to predict a wider range of material properties. For example, by introducing the virtual structure data 12, it becomes possible to predict regions with higher properties that are difficult to predict in the first embodiment. Furthermore, in the second embodiment, by setting targets for structure variables based on the virtual structure data 12, it is possible to efficiently narrow the search range of the virtual process data 11 and quickly arrive at a process that satisfies the desired properties. Therefore, even in the early stages of experiments when only limited experimental data is available, it is possible to efficiently find a process that achieves high properties.

[0101] Figure 21 is a graph comparing the search performance of each method for the optimum process based on data obtained in the early stages of the experiment. Four calcination experiments were carried out on the ferrite calcined body, and the saturation magnetization σ s The top three data points with the highest values ​​of saturation magnetization σ were selected. s 16 is a graph plotting the processes corresponding to the top three data points with the highest values ​​of (characteristics). These processes serve as evaluation criteria for ferrite calcined bodies to achieve high properties. Next, based on the data (24 data points) from the first calcination experiment of the four experiments, a process-structure model 5, a structure-property model 8 (see FIG. 16), and a process-property model 16 (see FIG. 23; conventional model) were learned and generated. In addition, virtual structure process data 11 and virtual structure data 12 were generated. Then, the method of the first embodiment, the method of the second embodiment, and the conventional method were used to determine the optimal process to be used as a candidate for the next experiment.

[0102] FIG. 21(b) is a graph plotting the optimal process 10 (24 points) that will be the next experiment candidate determined by the method of the first embodiment. FIG. 21(c) is a graph plotting the optimal process 10 (24 points) that will be the next experiment candidate determined by the method of the second embodiment. FIG. 21(d) is a graph plotting the optimal process 10 (24 points) that will be the next experiment candidate determined by the conventional method. The conventional method is a method of determining an optimal process using a process-characteristic model 16 (FIG. 23). In summary, the conventional method first inputs virtual process data 11 into the process-characteristic model 16 to predict the characteristics of each virtual process data 11, and then uses Bayesian optimization to select one or more process variables from the virtual process data 11 that satisfy the desired characteristics, and then determines and outputs the process variables as the optimal process 10. The acquisition function calculated by Bayesian optimization is a characteristic (saturation magnetization σ s ) is maximized by using PI (Probability of Improvement).

[0103] From FIG. 21(c), it can be seen that the optimum process 10 (10-2) determined in the second embodiment has a saturation magnetization σ s It was confirmed that the optimal process 10 (10-1) determined by the first embodiment matches two of the top three processes with the highest (characteristic) values. Furthermore, from FIG. 21(b), it was confirmed that the optimal process 10 (10-1) determined by the first embodiment matches one of the top three processes. On the other hand, from FIG. 21(c), the optimal process determined by the conventional method did not match any of the top three processes. These results demonstrate that by applying the methods of the first and second embodiments, a process that achieves high performance can be selected even from limited data in the early stages of experiments. In particular, in the second embodiment (FIG. 21(c)), the matching rate with the process in FIG. 21(a) is high, demonstrating that a process that achieves high performance is efficiently searched for. In other words, it is expected that an optimal process that achieves high performance can be efficiently found even from limited experimental data.

[0104] (Method of manufacturing sintered ferrite magnets) By analyzing a material using the material data processing device, material data processing method, and program according to the embodiment of the present invention, for example, when the material is a ferrite magnet, a high saturation magnetization σ s When the calcined ferrite body is used for a sintered ferrite magnet, it becomes possible to obtain a sintered ferrite magnet having excellent magnetic properties. An example of using the ferrite calcined body of this embodiment to produce a sintered ferrite magnet will be described below.

[0105] FIG. 22 is a flowchart showing an example of a manufacturing process for a sintered ferrite magnet. In step S31 ("raw material process"), appropriate amounts of raw material powders, such as CaCO3 powder, La(OH)3 powder, Fe2O3 powder, and ZnO powder, are prepared according to the optimal process 10 determined and output by the materials data processor 1 or 1a (the optimal process 10 determined and output in step S9 of FIG. 10 or step S19 of FIG. 19). The raw material powders can be oxides, carbonates, hydroxides, nitrates, chlorides, etc. of the respective metals, regardless of their valence. The prepared raw material powders are mixed to obtain a mixed raw material powder. The raw material powders can be mixed either wet or dry. Mixing with a medium such as steel balls can achieve more uniform mixing of the raw material powders. In the wet process, water is preferably used as a dispersant. To improve the dispersibility of the raw material powders, known dispersants such as ammonium polycarboxylate and calcium gluconate may be used. The wet-mixed raw material slurry can be calcined directly, or the raw material slurry can be calcined after dehydration.

[0106] In step S32 (the "calcination step"), the mixed raw material powder obtained by dry mixing or wet mixing is heat-treated using an electric furnace, gas furnace, or the like to form a ferrite compound having a hexagonal M-type magnetoplumbite structure through a solid-state reaction (ferritization reaction). This process is generally called "calcination," and the resulting heat-treated product is called a "calcined body" or "ferrite calcined body." Calcination is preferably carried out in an atmosphere with an oxygen concentration of 5% by volume or more. An oxygen concentration of less than 5% by volume may result in abnormal grain growth, the formation of heterogeneous phases, etc. A more preferred oxygen concentration is 20% by volume or more. The calcination temperature is preferably about 1100 to 1450°C. The calcination time is preferably about 0.5 to 5 hours.

[0107] In step S33 ("pulverization step"), the ferrite calcined body is pulverized into powder using a hammer mill, vibration mill, ball mill, attritor, or the like. The average particle size of the powder is preferably about 0.4 to 0.8 μm (air permeation method). The pulverization step may be either dry pulverization or wet pulverization. Typically, wet pulverization produces a slurry containing water (dispersion medium) and powder of the calcined body. A known dispersant and / or surfactant may be added to the slurry in an amount of 0.2 to 2 mass % in terms of solid content. After wet pulverization, the slurry may be concentrated.

[0108] In step S34 ("molding step"), the slurry obtained in the pulverization step is typically poured into a mold of a molding machine and press-molded in a magnetic field while the dispersion medium is discharged. In step S35 ("sintering step"), the compact obtained by press molding is degreased as necessary and then fired (sintered) to produce a sintered body. Sintering is carried out using an electric furnace, gas furnace, or the like. Sintering is preferably carried out in an atmosphere with an oxygen concentration of 10% by volume or more, more preferably 20% by volume or more, and most preferably 100% by volume. The firing temperature is preferably about 1150 to 1250°C. The firing time is preferably about 0.5 to 2 hours. Alternatively, a firing step disclosed in International Publication No. 2014 / 021149 may be employed, in which the heating rate in the temperature range from 1100°C to the firing temperature is 1 to 4°C / min, and the cooling rate in the temperature range from the firing temperature to 1100°C is 6°C / min or more.

[0109] In step S36 ("processing step"), the sintered body is ground to the desired shape and dimensions to complete the ferrite magnet. In step S37 ("inspection step"), the completed ferrite magnet is inspected. Inspection items include measuring the magnetic properties of the ferrite magnet and checking for chips or scratches on its exterior. After the processing step, a cleaning step may be performed before the inspection step.

[0110] In this embodiment, the composition ratio of the raw materials mixed in step S31 is shown as an example of a process variable. However, the process variables determined by the material data processing device 1 or 1a include not only the composition ratio but also the oxygen concentration, calcination temperature, and calcination time of the calcination in step S32, the type, amount, and particle size of additives mixed into the ferrite calcined body in step S33, the magnetizing current and press pressure for applying a magnetic field in step S34, and the oxygen concentration, calcination temperature, and calcination time for the calcination in step S35.

[0111] While the preferred embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to these examples. It is clear that those skilled in the art can conceive of various modifications or alterations within the scope of the technical ideas disclosed herein, and it is understood that these modifications also fall within the technical scope of the present invention. [Explanation of symbols]

[0112] 1, 1a: Material data processing device 2, 3: Training data 4: Process variables 4i, 4j: Selected process variables 5: Process-organization model 7: Optimal combination (organizational variables) 7i: Selected organizational variables 8: Organization-Property Model 9: Material characteristics 10: Optimal process 11: Virtual process data 12: Virtual organization data 12K: Virtual tissue data that meets the desired characteristics 13: Prediction data for organizational variables 14: Predicted data of characteristics 15: Organizational variable goals 16: Process-Characteristic Model 20: Characteristic model generation unit 21: Organization combination candidate setting unit 23: Model candidate generation unit 25: Model selection section 30: Organization model generation unit 40: Optimal process determination unit 41: Process data generation unit 42: Tissue Prediction Department 43: Characteristics prediction section 44: Tissue Selection Department 45: Process selection section 50: Optimal process determination unit 51: Process data generation unit 52: Organizational data generation unit 53: Characteristics prediction section 54:Organizational Goal Setting Department 55: Process selection section 60: Candidate organizational variables 70: Candidate combinations of organizational variables 70a-70c: Candidate combinations of organizational variables 71-71: Candidate combinations of organizational variables 80: Candidate organization-property model 80a-80c: Candidate structure-property models 81-89: Candidate tissue-property models 101: Control unit 102: Storage section 103: Communications Department 104: Input section 105:Display section 106: Peripheral device I / F section 108: Bus 251: Prediction accuracy calculation unit 252: Organizational Combination Determination Department

Claims

1. a model candidate generation unit that learns and generates, for each candidate combination of structural variables of a material, a structural-property model candidate that predicts the properties of the material from each structural variable included in the candidate; a model selection unit that selects a candidate based on a combination of tissue variables with the highest prediction accuracy from the candidates for the structure-property model generated by the model candidate generation unit, and adopts the candidate as the structure-property model to be used for analysis; A material data processing device comprising:

2. an organizational data generation unit that generates virtual organizational data, which is virtual data of each organizational variable; a property prediction unit that inputs the virtual structure data generated by the structure data generation unit into the structure-property model and predicts the properties of the material; an organizational target setting unit that sets targets for each organizational variable that satisfy desired characteristics based on the predictions of the characteristic prediction unit; The material data processing apparatus of claim 1 further comprising:

3. a process data generation unit that generates virtual process data, which is virtual data of each process variable; a process selection unit that selects, from the virtual process data generated by the process data generation unit, a process variable that satisfies the goal set by the organizational goal setting unit; The material data processing apparatus of claim 2 further comprising:

4. a process-organization model generation unit that learns and generates a process-organization model that predicts each organizational variable from each process variable, The materials data processing apparatus according to claim 1 , wherein a model is constructed by coupling the process-structure model and the structure-property model.

5. a process data generation unit that generates virtual process data, which is virtual data of each process variable; a structure prediction unit that inputs the virtual process data generated by the process data generation unit into the process-structure model and predicts structure parameters of the material; a property prediction unit that inputs the predicted data of the structural variables predicted by the structural prediction unit into the structural-property model and predicts the properties of the material; a texture selection unit that selects texture variables that satisfy desired properties from the predicted data of the texture variables based on the predictions of the property prediction unit; a process selection unit that selects, from the virtual process data, a process variable corresponding to the tissue variable selected by the tissue selection unit; The material data processing apparatus of claim 4 further comprising:

6. 2. The materials data processing device according to claim 1, wherein the material is a magnet, and at least one of the candidate combinations of structural variables is a main phase ratio of the magnet, a Curie temperature of the main phase, and a ratio c / a of the lattice constants of the main phase in the c-axis direction to the a-axis direction.

7. 2. The materials data processing apparatus according to claim 1, wherein the material is a ferrite magnet, and at least one of the candidate combinations of structural variables is a compound phase ratio having a magnetoplumbite structure of the ferrite magnet, a Curie temperature of the compound phase, and a ratio c / a of the lattice constants in the c-axis direction and the a-axis direction of the compound phase.

8. The computer a model candidate generation step of learning and generating, for each candidate combination of structural variables of the material, a structural-property model candidate that predicts the properties of the material from each structural variable included in the candidate; a model selection step of selecting a candidate based on a combination of tissue variables with the highest prediction accuracy from the candidates for the structure-property model generated by the model candidate generation step, and adopting the candidate as the structure-property model to be used for analysis; A material data processing method comprising:

9. a process-organization model generation step of learning and generating a process-organization model that predicts each organizational variable included in the combination of organizational variables with the highest prediction accuracy from each process variable; a process data generation step of generating virtual process data, which is virtual data of each process variable; a texture prediction step of inputting the generated virtual process data into the process-texture model to predict the texture parameters of the material; a property prediction step of inputting the predicted data of the structural variables into the structure-property model to predict the properties of the material; a texture selection step of selecting texture variables that satisfy desired properties from the texture variable prediction data based on the property prediction data; a process selection step of selecting a process variable corresponding to the selected organizational variable from the virtual process data, determining the process variable as an optimal process, and outputting the process variable; 9. The materials data processing method of claim 8, further comprising:

10. a process data generation step of generating virtual process data, which is virtual data of each process variable; an organizational data generation step of generating virtual organizational data, which is virtual data of each organizational variable; a property prediction step of inputting the generated virtual structure data into the structure-property model to predict the properties of the material; an organizational goal setting step of setting goals for each organizational variable that satisfy desired characteristics based on the predicted data of the characteristics; a process selection step of selecting process variables that satisfy the set target from the generated virtual process data, determining the process variables as an optimal process, and outputting the selected process variables; 9. The materials data processing method of claim 8, further comprising:

11. Computer, a model candidate generation unit that learns and generates, for each candidate combination of structural variables of a material, a structural-property model candidate that predicts the properties of the material from each structural variable included in the candidate; a model selection unit that selects a candidate based on a combination of tissue variables with the highest prediction accuracy from the candidates for the structure-property model generated by the model candidate generation unit, and adopts the candidate as the structure-property model to be used for analysis; A program that functions as a

12. A method for manufacturing a magnet, comprising manufacturing a magnet based on an optimum process output by the material data processing method according to claim 9 or 10.

13. The method for producing a magnet according to claim 12, wherein the magnet is a ferrite magnet.

Images