System and method for assisting in material creation

Abstract

According to the present invention, a search for a production recipe is divided into a preceding stage and a subsequent stage. In the preceding stage, the range of material compositions that can be included in a production recipe is narrowed down. A preceding-stage model constructed using preceding-stage training data, which represents a set of material compositions and material characteristics for each of a plurality of material compositions, is used in narrowing down the range of material compositions. The preceding-stage model represents the relationship between the material compositions and the material characteristics. One or more of the material compositions for the preceding-stage model is designated, whereby a predicted value related to the material composition and / or the material characteristic is obtained for each designated material composition. In the subsequent stage, one or more material compositions for which the resulting predicted value satisfies a requirement and a target material characteristic are designated by a subsequent-stage model that represents the relationship between production method recipes and the material characteristics, whereby a production method recipe is searched from the range of material compositions designated by the subsequent-stage model for a material having the target material characteristic.

Description

System and Method for Supporting Material Creation

[0001] The present invention generally relates to technology for supporting material creation.

[0002] As one method for supporting material creation, materials informatics (MI) is known. MI is a method for supporting material creation using informatics techniques, and generally refers to a method of aggregating data related to material properties and using machine learning to search for materials having new material properties. Regarding MI, for example, the technology disclosed in Patent Document 1 is known.

[0003] Japanese Unexamined Patent Application Publication No. 2024-015482

[0004] In material creation, it is desirable to efficiently perform target searches, for example, searching for a manufacturing recipe for a material having target properties, or searching for properties of a material produced according to a given manufacturing recipe.

[0005] The search for a manufacturing recipe is divided into a first stage and a second stage. In the first stage, the range of material compositions that can be included in the manufacturing recipe is narrowed down. For narrowing down the range of material compositions, a first-stage model constructed using first-stage training data, which is data representing a set of material composition and material property for each of a plurality of material compositions, is used. The first-stage model is a model representing the relationship between material composition and material property. By specifying one or more material compositions for the first-stage model, predicted values related to material composition and / or material property can be obtained for each specified material composition. In the second stage, one or more material compositions for which the obtained predicted values satisfy requirements and the target material property are specified to a second-stage model, which is a model representing the relationship between manufacturing recipe and material property. Thus, a manufacturing recipe for a material having the target material property is searched from the range of material compositions specified to the second-stage model.

[0006] According to the present invention, a target search in material creation can be efficiently performed.

[0007] This document shows an example of the system configuration in the first embodiment of the present invention. This document shows an example of the configuration of the MI platform system 100. This document shows an example of the functional blocks of the AI ​​unit 108. This document shows an example of the processing flow performed by the learning unit 351. This document shows an example of the processing flow performed by the inference unit 352. This document schematically shows recipe provision. This document shows one modified example of learning the first model 301. This document shows an example of the functional blocks of the AI ​​unit 108X according to the second embodiment of the present invention. This document shows an example of learning according to the second embodiment of the present invention. This document shows an example of pre-learning data. This document shows an example of post-learning data. This document shows an example of composition item refinement. This is an explanatory diagram illustrating an example of how composition item refinement contributes to post-learning. This document shows an example of inference according to the second embodiment of the present invention. This document shows an example of a composition formula list. This document shows an example of composition information. This document shows an example of pre-inference input information. This document shows an example of pre-inference output information.

[0008] In the following description, "interface device" may refer to one or more interface devices. These one or more interface devices may be at least one of one or more I / O (Input / Output) interface devices and one or more communication interface devices. The one or more I / O interface devices may be interface devices to at least one of the following: an I / O device and a remote display computer. The I / O interface device to the display computer may be a communication interface device. The at least one I / O device may be either a user interface device, such as an input device like a keyboard and a pointing device, or an output device like a display device. The one or more communication interface devices may be one or more identical communication interface devices (e.g., one or more NICs (Network Interface Cards)) or two or more heterogeneous communication interface devices (e.g., a NIC and an HBA (Host Bus Adapter)).

[0009] Furthermore, in the following explanation, "memory" refers to one or more memory devices, which are examples of one or more storage devices, and may typically be main memory devices. At least one memory device in memory may be a volatile memory device or a non-volatile memory device.

[0010] Furthermore, in the following explanation, "persistent storage device" may refer to one or more persistent storage devices, which are examples of one or more storage devices. Persistent storage devices are typically non-volatile storage devices (e.g., auxiliary storage devices), and specifically may be, for example, HDDs (Hard Disk Drives), SSDs (Solid State Drives), NVME (Non-Volatile Memory Express) drives, or SCMs (Storage Class Memory).

[0011] Furthermore, in the following explanation, "storage device" may refer to at least memory, including both memory and persistent storage.

[0012] Furthermore, in the following explanation, "processor" may refer to one or more processor devices. At least one processor device may typically be a microprocessor device such as a CPU (Central Processing Unit), but may also be other types of processor devices such as a GPU (Graphics Processing Unit). At least one processor device may be single-core or multi-core. At least one processor device may be a processor core. At least one processor device may be a broad-sense processor device such as a circuit that is a collection of gate arrays (e.g., FPGA (Field-Programmable Gate Array), CPLD (Complex Programmable Logic Device), or ASIC (Application Specific Integrated Circuit)) which performs some or all of the processing using a hardware description language.

[0013] Furthermore, in the following explanation, functions may be described using the expression "yyy section," but a function may be realized by the execution of one or more computer programs by a processor, by one or more hardware circuits (e.g., FPGA or ASIC), or by a combination thereof. When a function is realized by the execution of a program by a processor, the defined processing is carried out using a memory device and / or interface device as appropriate, so the function may be at least a part of the processor. The processing described with a function as the subject may be processing performed by the processor or a device having that processor. Programs may be installed from program source. Program source may be, for example, a program distribution computer or a computer-readable recording medium (e.g., a non-temporary recording medium). The description of each function is an example, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

[0014] Hereinafter, several embodiments of the present invention will be described with reference to the drawings. In the following description, "MI" is an abbreviation for materials informatics, and "DB" is an abbreviation for database.

[0015] Figure 1 shows an example of a system configuration in the first embodiment of the present invention.

[0016] A materials creation support system 10 is constructed to assist in the creation of new materials. The materials creation support system 10 may be a physical computer system, a logical computer system based on a physical computer system, or a combination of at least a part of a physical computer system and at least a part of a logical computer system. The physical computer system may consist of one or more physical computers and may include interface devices, memory devices, and processors connected thereto. The logical computer system may include a virtual machine or a system as a cloud computing service.

[0017] The material creation support system 10 includes a data presentation unit 50, a data acquisition unit 60, and a data conversion unit 70. These functions 50, 60, and 70 may be realized by a computer program being executed on a processor.

[0018] The data presentation unit 50 retrieves data from the data mart 107 periodically or each time an inquiry is received and presents the target data. Hereinafter, the data in the data mart 107 will be referred to as "organized data". The target data is at least one of the retrieved organized data and the data based on the inference results obtained by inputting the organized data into a machine learning model. The data mart 107 is an example of a first data source. Preferably, the data includes data relating to ceramic materials.

[0019] The data acquisition unit 60 stores experimental data, which is data obtained from experiments based on the target data, in the electronic laboratory notebook 131, the creation database 121, and the evaluation database 122. The electronic laboratory notebook 131, the creation database 121, and the evaluation database 122 are examples of second data sources. In this embodiment, "experimental data" may refer to data related to the experiment, such as a summary, details, results, or evaluation of the experiment.

[0020] The data conversion unit 70 acquires experimental data from the electronic laboratory notebook 131, the creation database 121, and the evaluation database 122, and converts the experimental data into data included in an existing or new data mart 107.

[0021] After the organized data based on experimental data is stored in the existing data mart 107, or after a new data mart 107 containing the organized data based on experimental data is created, the target data is based on the updated or newly created data mart 107, i.e., the organized experimental data from experiments conducted based on previously presented target data. Therefore, regardless of whether the type of material to be created is a type of material from which highly accurate simulation data regarding material properties can be obtained, the likelihood of presenting data useful for material creation can be increased. As a result, the types of materials that can be supported by informatics methods for material creation are expanded. Specifically, for example, even if the type of material to be created is an inorganic material (ceramics), data useful for material creation can be presented.

[0022] This embodiment will now be described in detail.

[0023] Regarding the material creation support system 10, the interface device may be connected to communicate with at least one of the experimental system 110, the data acquisition unit 60, and the researcher terminal 11A. The storage device may store at least some of the data from the electronic experimental notebook 131, the fabrication DB 121, the evaluation DB 122, various data 150, the data lake 102, and the data mart 107. The processor may realize the above-described functions 50, 60, and 70 by executing a computer program.

[0024] A data presentation unit 50 and a data conversion unit 70 are provided in the MI platform system 100. A data acquisition unit 60 is provided outside (or inside) the MI platform system 100. The MI platform system 100 may be a physical computer system, a logical computer system, or a combination of at least a part of a physical computer system and at least a part of a logical computer system. In this embodiment, as shown in Figure 2, the MI platform system 100 is a physical computer system and has an interface device 201, a storage device 202, and a processor 203 connected thereto. Communication with a researcher terminal 11 (at least a researcher terminal 11A) may be performed through the interface device 201. A data lake 102 or a data mart 107 may be provided in the storage device 202 and data may be stored there. A program may also be stored in the storage device 202. The processor 203 may read and execute a program from the storage device 202 to realize at least some of the functions of the data presentation unit 50 and the data conversion unit 70 in the MI platform system 100.

[0025] The data presentation unit 50 includes an AI (Artificial Intelligence) unit 108 and an IF (Interface) unit 109.

[0026] The AI ​​unit 108 performs training on a machine learning model and inference using the machine learning model. For example, in response to instructions from the IF unit 109, the AI ​​unit 108 outputs the inference results obtained by inputting the organized data acquired from the data mart 107 into the machine learning model to the IF unit 109.

[0027] The IF unit 109 receives an inquiry from the researcher terminal 11A and, in response to the inquiry, presents the target data to the researcher terminal 11A. The researcher terminal 11A is an information processing terminal (e.g., a personal computer or smartphone) of a materials researcher 5A (an example of a user). The researcher terminal 11A is an example of the source of the inquiry for the target data, and also an example of the destination to which the target data is presented. When the IF unit 109 receives an inquiry from the researcher terminal 11A (or periodically), it presents the object represented by the target data (organized data obtained from the data mart 107, and / or data based on inference results obtained by instructing the AI ​​unit 108). The object presented may be, for example, a manufacturing recipe or material properties. In this embodiment, "manufacturing recipe" means a method for creating a material, and typically includes material composition and / or synthesis process. "Material composition" may be, for example, a blending composition, and "synthesis process" may be raw material types (e.g., different particle sizes) or process conditions. The manufacturing recipe may also include a material development strategy.

[0028] Materials researcher 5A creates new materials or conducts experiments based on the presented target data. The destination (transmission destination) for the target data may be a system such as the experimental system 110, instead of or in addition to the researcher terminal 11A.

[0029] Materials researcher 5A conducts experiments using the experimental system 110 based on the presented subject. For example, the experiment using the experimental system 110 may be a combinatorial experiment. The experimental system 110 may include one or more devices, for example, a ceramics fabrication device 111 (for example, an atmospheric firing furnace) and a ceramics evaluation device 112 (for example, a device for evaluating the coefficient of thermal expansion). Experimental data is output from experimental devices such as the fabrication device 111 and the evaluation device 112. The outputted experimental data is transmitted to and stored in the data server 120. Furthermore, the combinatorial experiment is just one example of an experiment, and at least one of the following may be adopted instead of or in addition to the combinatorial experiment: a high-throughput experiment, an automated experiment using a robot, and an experiment mainly involving manual work.

[0030] The data acquisition unit 60 includes a data server 120 and an experiment notebook unit 130.

[0031] The data server 120 manages the databases (DBs) where experimental data from the experimental system 110 is stored. The databases managed by the data server 120 may exist for each type of experimental apparatus. For example, a fabrication DB 121 storing experimental data from the fabrication apparatus 111 and an evaluation DB 122 storing experimental data from the evaluation apparatus 112 may be managed. The experimental data from the experimental system 110 is structured as data by the data server 120 and stored in the databases.

[0032] The experimental notebook unit 130 manages experimental data obtained from the database managed by the data server 120 as an electronic experimental notebook 131. The electronic experimental notebook 131 is structured data, such as a database.

[0033] The data conversion unit 70 includes a data collection unit 101, a feature calculation unit 103, an image analysis unit 104, a natural language analysis unit 105, and a data organization unit 106.

[0034] The collection unit 101 collects experimental data from the electronic lab notebook 131 and the database of the data server 120, formats the collected experimental data, and stores the formatted experimental data in the data lake 102. Here, "formatting" means arranging the structure of the experimental data into a predetermined structure.

[0035] Furthermore, the collection unit 101 may collect various types of data 150. These various types of data 150 may include experimental data. Specifically, for example, the various types of data 150 may include at least one of the following: numerical data from past experiments, a paid or free materials database, linguistic data representing past experiments (e.g., experimental content and results), materials data obtained by calculations such as simulations, and linguistic data from patent documents and papers. There may be one or more data sources for the various types of data 150. For example, the data source for the numerical data from past experiments and the data source for the linguistic data from patent documents and papers may be different. The various types of data 150 may also include sensor data, which is data containing measurements from sensors. At least a portion of the various types of data 150 may be stored in the data lake 102.

[0036] The feature calculation unit 103 calculates the feature quantities of one or more predetermined types of data in the formatted experimental data. As a result, each of the one or more types of data is converted into a numerical value, which in turn enables processing by the AI ​​unit 108. For example, the one or more types of data may be image data and text data. The image data is analyzed by the image analysis unit 104, and based on the results of the analysis, the feature calculation unit 103 calculates the feature quantities of the image data. The text data is text-mined by the natural language processing unit 105, and based on the results of the text mining, the feature calculation unit 103 calculates the feature quantities of the text data. The above-mentioned neural network may also be used for feature calculation. A copy of the experimental data is generated in the data lake 102, and each of the one or more predetermined types of data in the copy may be converted into the calculated feature quantities (numerical values).

[0037] The sorting unit 106 sorts the data lake 102 into one or more data marts 107 that meet predetermined conditions. For example, a data mart 107 containing sorted data that can be presented as at least part of the target data without going through the AI ​​unit 108 may be generated, and a data mart 107 containing sorted data that can be presented as at least part of the target data via the AI ​​unit 108 may be generated.

[0038] In this way, organized data based on experimental data collected and formatted by the MI platform system 100 is prepared, and the target data is presented to the materials researcher 5A by the MI platform system 100 based on the organized data.

[0039] Furthermore, the presentation of data to materials researcher 5A may be done without going through the MI platform system 100. For example, another materials researcher 5B may use researcher terminal 11B to acquire experimental data from the electronic lab notebook 131, organize the experimental data, and input it into the AI ​​unit 140. The AI ​​unit 140 may be implemented outside (or inside) the researcher terminal 11B. The AI ​​unit 140, like the AI ​​unit 108, may perform inference by inputting the organized data into a machine learning model. Data based on the results of this inference may be presented to materials researcher 5A. Materials researcher 5A may conduct experiments based on this presented data. The data from these experiments may also be stored in the data server 120 from the experimental system 110 and collected by the MI platform system 100.

[0040] A predetermined function implemented within or outside the researcher terminal 11A may facilitate experiments based on the presented target data. For example, this function may determine what kind of experiment to recommend to the materials researcher 5A and present the decision to the materials researcher 5A. Alternatively, the target data may include data indicating what kind of experiment is recommended.

[0041] According to this embodiment, it is expected that the cycle of presenting target data → conducting experiments → collecting experimental data → converting experimental data into organized data → presenting target data based on the organized data can be performed with high efficiency and speed. In this way, MI (Material Inspection) based primarily on experimental data is realized. Therefore, it is possible to support the creation of materials of material types for which it is difficult to obtain highly accurate simulation data regarding material properties, such as materials with complex higher-order structures in ceramics.

[0042] Furthermore, according to this embodiment, various internal and external data of the organization (e.g., a company) providing the MI platform system 100 can be aggregated, structured, and managed in the MI platform system 100 and reflected in the data mart 107. As a result, the likelihood of the presented target data being useful for material creation is increased.

[0043] Now, the AI unit 108 according to the present embodiment will be described in detail below. In the following description, it is assumed that the material to be created is ceramics. In addition, in the following description, the definitions of terms are as follows. ・The term "data set" refers to one logical mass of electronic data viewed from a program such as an application program, and may be any of, for example, a record, a file, a key-value pair, a tuple, or the like. ・The term "researcher input information" refers to information input from the researcher 5A. The researcher input information is input to the IF unit 109. In the present embodiment, the researcher input information is typically information representing a manufacturing recipe or material properties. ・The term "researcher output information" refers to information output (provided) to the researcher 5A. The researcher output information is output from the IF unit 109. In the present embodiment, the researcher output information is typically information representing material properties or a manufacturing recipe. ・The term "manufacturing recipe data set" refers to a data set representing a manufacturing recipe for a material. When the researcher input information is information representing a manufacturing recipe, the manufacturing recipe data set may be a data set based on said researcher input information. ・The term "material property data set" refers to a data set representing material properties. When the researcher output information is information representing material properties, said researcher output information may be information based on the material property data set.

[0044] Fig. 3 shows an example of functional blocks of the AI unit 108.

[0045] The AI unit 108 may receive a data set from the IF unit 109 (that is, the IF unit 109 may generate a data set based on researcher input information), or may receive researcher input information from the IF unit 109 and generate a data set based on said researcher input information. Furthermore, the AI unit 108 may output a data set to the IF unit 109 (that is, the IF unit 109 may generate researcher output information based on the data set), or may output researcher output information based on the data set to the IF unit 109. Furthermore, the IF unit 109 may be included in the AI unit 108.

[0046] The AI unit 108 inputs the production recipe data set 311 and outputs (predicts) the material property data set 313 in inference using a machine learning model. In addition, in the present embodiment, when the researcher input information is information representing material properties, the AI unit 108 specifies the production recipe data set 311 of a production recipe that is expected to produce a material having the material properties, and outputs the production recipe data set 311 (or researcher output information based on the production recipe data set 311) to the IF unit 109. That is, when the researcher 5A inputs information representing a production recipe to the MI platform system 100, the researcher 5A can receive, from the MI platform system 100, provision of information representing predicted material properties of a material assumed to be produced by the production recipe. In addition, when the researcher 5A inputs information representing material properties to the MI platform system 100, the researcher 5A can receive, from the MI platform system 100, provision of information representing a production recipe of a material that is expected to produce a material having the material properties.

[0047] The AI unit 108 includes a learning unit 351 that performs learning of a machine learning model, and an inference unit 352 that performs inference using the learned machine learning model (that is, the inference described above). Further, as the machine learning model, there are first to third models 301 to 303. The first to third models 301 to 303 are stored in the storage device 202.

[0048] The first model 301 is a model that receives the production recipe data set 311 as an input and outputs the material property data set 313 as an output. Ceramics have a complex higher-order structure in which a large number of crystal particles are aggregated (for example, particle size, particle size distribution, grain boundary phase, etc.), and material features including such a higher-order structure affect the material properties of ceramics, so it is generally difficult to predict material properties with high accuracy from a production recipe.

[0049] Therefore, in this embodiment, based on the fact that material characteristics influence material properties, or in other words, that there is a correlation between material characteristics and material properties, a second model 302 is prepared that takes a material characteristics dataset 312 (a dataset representing material characteristics) as input and outputs a material characteristics dataset 313. Then, in order to make the material characteristics dataset 313 predictable from the manufacturing recipe dataset 311 via the intermediate dataset of material characteristics dataset 312, a third model 303 is prepared that takes the manufacturing recipe dataset 311 as input and outputs the material characteristics dataset 312.

[0050] The "manufacturing recipe" represented by the manufacturing recipe dataset 311 may include, as described above, material composition and / or synthesis process, and specifically, may include, for example, one or more recipe items (e.g., synthesis temperature) and values ​​(typically numerical) for each of the one or more recipe items.

[0051] The "material properties" represented by the material property dataset 313 refer to the properties exhibited by a substance (e.g., a structure) as a material, and specifically include, for example, one or more property items (e.g., strength, thermal expansion coefficient, etc.) and the values ​​(typically numerical) for each of those one or more property items.

[0052] The "material features" represented by the material feature dataset 312 refer to the physical and / or chemical state exhibited by a substance as a material, and specifically include, for example, one or more feature items (e.g., grain size, grain size distribution, etc.) and values ​​(typically numerical) for each of those one or more feature items. Feature quantities may be included in the material feature dataset 312 for each feature item. For example, using the second model 302, it is possible to predict material properties including values ​​for one or more characteristic items from material features including values ​​for multiple feature items, or to predict material properties including values ​​for one or more characteristic items from material features including a value for one feature item.

[0053] The material feature dataset 312 may be a dataset representing features calculated by the feature calculation unit 103, and may be stored in the data mart 107 and retrieved from the data mart 107. One or more features represented by the material feature dataset 312 are extracted by the feature calculation unit 103 using data processing (for example, processing using CNN) from image data (e.g., SEM image) or spectral data. SEM stands for "Scanning Electron Microscope," and CNN stands for "Convolutional Neural Network."

[0054] For each of the first to third models 301 to 303, the model may be linear regression (e.g., Ridge regression, Lasso regression, Elastic Net regression), logistic regression, SVM (Support Vector Machine), decision tree models (e.g., Random Forest, XGBoost (Xtreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine)), neural networks (e.g., CNN, RNN (Recurrent Neural Network), ResNet (Residual Network)), Bayesian optimization, k-NN (k-Nearest Neighbor), or any combination of two or more of these models. Preferably, for example, for each of the first to third models 301 to 303, the model may be a random forest, XGBooster, LightGBM, CNN, Bayesian optimization, or any combination of two or more of these models. For example, the following can be expected: - Although the number of manufacturing recipe datasets 311 used as training data is relatively small, Bayesian optimization is capable of prediction including prediction deviation, is excellent at finding candidate points for adding data, and finds the optimal value while increasing the data. - Decision tree models (e.g., random forest, XGBooster, LightGBM) are combinations of simple binomial distribution problems, so they have low computational cost and high accuracy. For example, XGBooster or LightGBM create multiple decision trees and weight them to address targets with prediction errors, thereby preventing overfitting and increasing robustness. CNNs offer high degrees of freedom, can handle complex systems, are useful when dealing with a large number of features, and allow for adjustment of the degree of overfitting by appropriately controlling the number of training iterations (typically the number of epochs).

[0055] In this embodiment, each of the first to third models 301 to 303 is a regression equation.

[0056] Figure 4 shows an example of the processing flow performed by the learning unit 351.

[0057] The learning unit 351 acquires training data for the first model 301, specifically, one or more manufacturing recipe datasets 311 and one or more material property datasets 313 corresponding to the one or more manufacturing recipe datasets 311 (S401). These datasets 311 and / or 313 may be input from the researcher 5A during the training process, may be prepared in advance as training data, may be acquired from experimental data and / or various data 150, or may be a manufacturing recipe dataset 311 (a newly generated manufacturing recipe dataset 311) that represents a manufacturing recipe with narrowed recipe items or value ranges in a manufacturing recipe represented by a certain manufacturing recipe dataset 311 (or a manufacturing recipe with expanded recipe items or value ranges).

[0058] The learning unit 351 learns the first model 301 using the acquired datasets 311 and 313 (S402). Specifically, for example, the learning unit 351 learns the first model 301 using the manufacturing recipe dataset 311 and the material properties dataset 313 corresponding to the manufacturing recipe dataset 311 (manufacturing recipes and the material properties of the materials created by those manufacturing recipes).

[0059] The learning unit 351 determines whether or not to terminate the learning of the first model 301 (S403). For example, if the accuracy (prediction accuracy) of the first model 301 is equal to or greater than the first accuracy threshold (or if the accuracy of the first model 301 is higher than the accuracy of the second model 302 and the third model 303), the determination result in S403 is true (S403: Yes). In this case, the processing of the learning unit 351 ends.

[0060] If the judgment result in S403 is false (S403: No), the learning unit 351 acquires the learning data for the second model 302, specifically, one or more material feature datasets 312 and one or more material property datasets 313 corresponding to the one or more material feature datasets 312 (S404). These datasets 312 and / or 313 may be input from the researcher 5A during the learning process, may be prepared in advance as learning data, may be acquired from experimental data and / or various data 150, or may be a material feature dataset 312 (a material feature dataset 312 newly generated by the learning unit 351) that represents material features with narrowed feature items or value ranges in the material features represented by a certain material feature dataset 312 (or material features with expanded feature items or value ranges).

[0061] The learning unit 351 learns a second model 302 using the acquired datasets 312 and 313 (S405). Specifically, for example, the learning unit 351 learns a second model 302 using the material feature dataset 312 and the material property dataset 313 corresponding to the material feature dataset 312 (material features and the material properties of materials possessing those material features). The purpose of this learning is to find material features that have a strong correlation with material properties, or in other words, to identify material features that correspond to material properties.

[0062] The learning unit 351 determines whether or not to terminate the learning of the second model 302 (S406). For example, if the accuracy (prediction accuracy) of the second model 302 is equal to or greater than the second accuracy threshold, the determination result in S406 is true (S406: Yes). The accuracy of the second model 302 being equal to or greater than the threshold means that important feature items and / or values ​​(features) have been identified. In other words, if important feature items and / or values ​​(features) are not identified, typically the accuracy of the second model 302 will not be equal to or greater than the threshold.

[0063] If the determination result in S406 is true, the learning unit 351 acquires the learning data for the third model 303, specifically, one or more material feature datasets 312 identified in the learning of the second model 302, and one or more manufacturing recipe datasets 311 corresponding to the one or more material feature datasets 312 (S407). The manufacturing recipe datasets 311 corresponding to the material feature datasets 312 may be, for example, a manufacturing recipe dataset 311 corresponding to a material property dataset 313 corresponding to the material feature dataset 312. These datasets 312 and / or 311 may be datasets acquired from the datasets used in this learning process, or they may be a manufacturing recipe dataset 311 (a newly generated manufacturing recipe dataset 311) that represents a manufacturing recipe with narrowed recipe items or value ranges in a manufacturing recipe represented by a certain manufacturing recipe dataset 311 (or a manufacturing recipe with expanded recipe items or value ranges).

[0064] The learning unit 351 learns a third model 303 using the acquired datasets 312 and 311 (S408). Specifically, for example, the learning unit 351 learns a third model 303 using the material feature dataset 312 and the manufacturing recipe dataset 311 (material features and manufacturing recipes for creating materials with those material features) that correspond to the material feature dataset 312.

[0065] The training of the third model 303 (S407 and S408) may be carried out in parallel with the training of the second model 302 (S404 and S405). The material feature dataset 312 acquired in S407 may be a dataset representing the features identified (narrowed down) during the training of the second model 302, or it may be the material feature dataset 312 before such identification.

[0066] The learning unit 351 determines whether or not to terminate the learning of the third model 303 (S409). If the accuracy (prediction accuracy) of the third model 303 is equal to or greater than the third accuracy threshold, the determination result in S409 is true (S409: Yes). The first to third accuracy thresholds mentioned above may be the same value or different values.

[0067] If the result of the determination in S409 is true, the learning unit 351 performs S401 again. In S401, the manufacturing recipe dataset 311 identified (or newly generated) in the learning of the third model 303 and the material properties dataset 313 corresponding to the manufacturing recipe dataset 311 may be acquired.

[0068] According to the process shown in Figure 4, as described above, if the accuracy of the first model 301 becomes equal to or greater than the first accuracy threshold and learning is completed, the learning unit 351 terminates. However, even if the determination in S403 is performed N times (where N is an integer of 2 or more), if the accuracy of the first model 301 is less than the first accuracy threshold (i.e., the determination result in S403 is not true), the learning unit 351 may also terminate.

[0069] In this embodiment, "retraining" means that the processing of the learning unit 351 is completed once and then the processing of the learning unit 351 is performed again. Retraining may be started at a predetermined trigger. The predetermined trigger may be at least one of the following, for example: - For at least one of the models 301 to 303, the amount of training data has increased sufficiently since the previous training (for example, the amount of training data has increased by more than a certain amount). - The accuracy of at least one of the models 301 to 303 (for example, the first model 301) has deteriorated to below the first accuracy threshold.

[0070] The inference performed by the inference unit 352 may be carried out in parallel with the processing of the learning unit 351. Furthermore, the dataset obtained in the inference performed by the inference unit 352 may be used to train any of the models 301 to 303. Specifically, for example, at least one of the following may be adopted: ・If the judgment result in S403 is true (the training of the first model 301 is completed), inference using the first model 301 may be performed, and a manufacturing recipe may be provided as appropriate. ・If the judgment result in S406 is false (the training of the second model 302 is not yet completed), inference using the second model 302 may be performed, and material characteristics may be provided as appropriate. Inference using the second model 302 may be possible, for example, when the difference between the accuracy of the second model 302 and the second accuracy threshold is less than or equal to a predetermined difference. - If the judgment result in S409 is false (the training of the third model 303 is not yet complete), inference using the third model 303 may be performed, and a manufacturing recipe may be provided as appropriate. Inference using the third model 303 may be possible, for example, when the difference between the accuracy of the third model 303 and the third accuracy threshold is less than or equal to a predetermined difference. - If the judgment result in S409 is true (the training of the third model 303 is complete), and the judgment result in S403 is false (the training of the first model 301 is not yet complete), inference using the second model 302 and the third model 303 may be performed, and a manufacturing recipe may be provided as appropriate. - Based on the manufacturing recipe provided after inference, or based on the manufacturing recipe estimated by researcher 5A based on the material characteristics provided after inference, an experiment may be conducted under new conditions determined by researcher 5A, and experimental data representing the manufacturing recipe, material properties, and material characteristics related to that experiment may be accumulated. The manufacturing recipe dataset 311, material feature dataset 312, and material properties dataset 313 obtained from the experimental data may be used in training or retraining to make any of the judgment results in S403, S406, and S409 true. The new conditions may include conditions based on the inference results, or they may include conditions determined independently of the inference results based on the knowledge of researcher 5A.

[0071] Figure 5 shows an example of the processing flow performed by the inference unit 352.

[0072] The inference unit 352 determines whether or not relearning has been performed (S501). If the result of the determination in S501 is true (S501: Yes), the inference unit 352 invalidates the list (S502). The "list" here refers to data generated or updated during inference, which is data recording the relationship between material properties and manufacturing recipes and / or material characteristics. The list is stored in the storage device 202 (see Figure 2) by the inference unit 352 and is accessible to the IF unit 109. "Invalidating" the list means making the list inaccessible to the IF unit 109. Specifically, this may involve, for example, deleting the list from the storage device 202, associating data that indicates invalidity with the list, or moving the list from one storage area to another. S502 is also optional. For example, each time a new list is generated or updated, the latest list may be made valid (a list referenced for providing manufacturing recipes or material properties).

[0073] The inference unit 352 determines whether the training of the first model 301 is complete (S503). If the determination result in S503 is true (S503: Yes), the inference unit 352 performs inference using the first model 301 (S504). Specifically, for example, the inference unit 352 may acquire multiple manufacturing recipe datasets 311 and, for each acquired manufacturing recipe dataset 311, acquire a material properties dataset 313 by inputting the manufacturing recipe dataset 311 into the first model 301. Each of the multiple manufacturing recipe datasets 311 may be a dataset acquired from newly added experimental data or various data 150, or it may be a dataset generated by changing the combination of recipe items or the values ​​for each recipe item (for example, the inference unit 352 may increase the number of manufacturing recipe datasets 311 by repeatedly changing the manufacturing recipe, such as increasing or decreasing the value of the synthesis temperature by a fixed amount). The inference unit 352 records in a list, for each pair of input manufacturing recipe dataset 311 and acquired material property dataset 313, the manufacturing recipe represented by the manufacturing recipe dataset 311 and the material property represented by the material property dataset 313. In addition, material properties may be recorded in the list if the material property meets a predetermined condition (for example, if the value of a certain property item falls within a predetermined value range).

[0074] If the result of the determination in S503 is false (S503: No), the inference unit 352 determines whether or not the training of the third model 303 has been completed (S505). If the result of the determination in S505 is true (S505: Yes), the inference unit 352 performs inference using the second and third models 302 and 303 (S506). Specifically, for example, the inference unit 352 obtains a material properties dataset 313 by inputting the material feature dataset 312 into the second model 302. Towards creating a material having the material properties represented by the obtained material properties dataset 313 (i.e., to identify a manufacturing recipe dataset 311 whose output is the material feature dataset 312 input for obtaining the material properties dataset 313), the inference unit 352 may obtain the material feature dataset 312 by inputting the manufacturing recipe dataset 311 into the third model 303. The inference unit 352 records in a list, for each set of input manufacturing recipe dataset 311, acquired material feature dataset 312, and acquired material property dataset 313, the manufacturing recipe represented by the manufacturing recipe dataset 311, the material feature represented by the material feature dataset 312, and the material property represented by the material property dataset 313. The list may also record material features corresponding to the manufacturing recipe and material properties. Furthermore, material properties may be recorded in the list if the material property meets a predetermined condition (for example, if the value of a certain property item falls within a predetermined value range). In addition, each of the multiple manufacturing recipe datasets 311 or material feature datasets 312 may be a dataset acquired from newly added experimental data or various data 150, or it may be a dataset generated by changing the combination of each item or the values ​​of each item (for example, the inference unit 352 may increase the number of manufacturing recipe datasets 311 by repeatedly changing the manufacturing recipe, such as by increasing or decreasing the value of the synthesis temperature by a fixed amount).

[0075] In addition, in S505, in addition to determining whether the training of the third model 303 has been completed, the inference unit 352 may also determine whether the accuracy of the inference using the second and third models 302 and 303 is above a threshold. If the result of this determination is also true, the inference unit 352 may perform S506 (inference using the second and third models 302 and 303).

[0076] If the result of the judgment in S505 is false (S505: No), the inference unit 352 determines whether the training of the second model 302 has been completed (S507). If the result of the judgment in S507 is true (S507: Yes), the inference unit 352 performs inference using the second model 302 and the third model 303 (S508). Specifically, for example, the inference unit 352 first uses the second model 302 to identify the material features necessary to satisfy the target material properties (a material feature dataset 312 as input such that the output is a material feature dataset 313 representing the target material properties). Then, the inference unit 352 uses the third model 303 to infer a manufacturing recipe to obtain the identified material features (specifically, it identifies a manufacturing recipe dataset 311 as input such that the output is a material feature dataset 312 representing the identified material features). The inference unit 352 records the manufacturing recipe, material features, and material properties in a list for each set of the input manufacturing recipe dataset 311, acquired material feature dataset 312, and acquired material property dataset 313. Each of the multiple manufacturing recipe datasets 311 or material feature datasets 312 may be a dataset acquired from newly added experimental data or various data 150, or it may be a dataset generated by changing the combination of recipe items or the values ​​for each recipe item (for example, the inference unit 352 may increase the number of manufacturing recipe datasets 311 by repeatedly changing the manufacturing recipe, such as by increasing or decreasing the synthesis temperature value by a fixed amount).

[0077] If the result of the determination in S507 is false (S507: No), the inference unit 352 performs inference using two models, the second and third models 302 and 303, or the first model 301 (S509). Specifically, for example, the inference unit 352 performs (S509-1) or (S509-2) below. More specifically, for example, if the prediction accuracy of the second model 302 and the third model 303 is higher than the prediction accuracy of the first model 301, the inference unit 352 performs (S509-1), and if the prediction accuracy of the second model 302 and the third model 303 is the same as or lower than the prediction accuracy of the first model 301, it may perform (S509-2). In the following, each of the multiple manufacturing recipe datasets 311 or material characteristic datasets 312 may be a dataset obtained from newly added experimental data or various data 150, or it may be a dataset generated by changing the combination of each item or the values ​​of each item (for example, the inference unit 352 may increase the number of manufacturing recipe datasets 311 by repeatedly changing the manufacturing recipe, such as by increasing or decreasing the value of the synthesis temperature by a fixed amount). (S509-1) The inference unit 352 obtains a material property dataset 313 by inputting the material characteristic dataset 312 into the second model 302. Towards creating a material having the material properties represented by the obtained material property dataset 313 (i.e., to identify the manufacturing recipe dataset 311 whose output is the material characteristic dataset 312 input for obtaining the said material property dataset 313), the inference unit 352 obtains a material characteristic dataset 312 by inputting the manufacturing recipe dataset 311 into the third model 303. The inference unit 352 records in a list the manufacturing recipe represented by the manufacturing recipe dataset 311, the material features represented by the material features dataset 312, and the material properties represented by the material properties dataset 313 for each set of input manufacturing recipe dataset 311, acquired material feature dataset 312, and acquired material property dataset 313. (S509-2) The inference unit 352 acquires the material property dataset 313 by inputting the manufacturing recipe dataset 311 into the first model 301 (in actual experiments, the material feature dataset 312 may also be collected for evaluation of material features).The inference unit 352 records in a list, for each pair of input manufacturing recipe dataset 311 and acquired material property dataset 313, the manufacturing recipe represented by the manufacturing recipe dataset 311 and the material property represented by the material property dataset 313.

[0078] The inference unit 352 determines whether or not to terminate the inference process (S510). If the result of the determination in S510 is true (S510: Yes), the process terminates. If the result of the determination in S510 is false (S510: No), the process returns to S501.

[0079] In the process shown in Figure 5, the list 600 shown in Figure 6 is generated or updated. When the IF unit 109 receives an inquiry from researcher 5A specifying material properties, it identifies a manufacturing recipe for creating a material with those material properties based on the list 600 and provides the identified manufacturing recipe to researcher 5A.

[0080] In the first embodiment, the manufacturing recipe dataset may include information about the material composition in addition to information about the synthesis process, such as manufacturing conditions and material conditions. Information about the material composition may include, for example, information representing the empirical formula, elemental characteristics (e.g., atomic weight, primitive radius), crystal structure, effective mass, carrier density, formation energy, and theoretical density.

[0081] Therefore, a recipe can contain many variables. If there are many variables in a recipe, the processing load on the recipe search can be high, which can make the search take a long time, and it may also be difficult to perform a highly accurate search.

[0082] Furthermore, while information regarding material composition can be obtained from publicly available information or calculated using generally known calculation methods, information regarding synthesis processes and material characteristics must be obtained through experiments, etc., and is less readily available than information regarding material composition. Therefore, for each of the first model 301 and the third model 303, which require a manufacturing recipe dataset as training data, and each of the second model 302 and the third model 303, which require a material characteristic dataset as training data, it may take a long time to train them until a certain level of accuracy is achieved.

[0083] Therefore, there is a second embodiment. In the second embodiment, it is expected that the time required to search for manufacturing recipes will be reduced. Furthermore, in the second embodiment, it is expected that the accuracy of the searched manufacturing recipes will be improved. Moreover, in the second embodiment, it is expected that the time required to train at least one of the first to third models 301 to 303 will be reduced.

[0084] The second embodiment will be described below. In this description, the differences from the first embodiment will be explained primarily, and the similarities with the first embodiment will be omitted or simplified.

[0085] Figure 8 shows an example of a functional block of the AI ​​unit 108X according to a second embodiment of the present invention.

[0086] The AI ​​unit 108X has a pre-processing unit 801 and a post-processing unit 811.

[0087] The subsequent processing unit 811 substantially corresponds to the learning unit 351 and the inference unit 352 in the AI ​​unit 108 of the first embodiment. In the second embodiment, the learning unit and the inference unit are referred to as the "learning / inference unit," but the learning unit and the inference unit may be separate. The subsequent processing unit 811 learns the subsequent model 850. The subsequent model 850 includes the first to third models 301 to 303. The subsequent processing unit 811 performs inference using at least one model in the subsequent model 850.

[0088] The pre-processing unit 801 performs training of the pre-model 800 and inference using the pre-model 800. In other words, in this embodiment, training consists of pre-training, which is training of the pre-model 800, and post-training, which is training of the post-model 850, and inference consists of pre-inference, which is inference using the pre-model 800, and post-inference, which is inference using the post-model 850.

[0089] The preliminary model 800 is a model of the relationship between material composition and material properties, and this relationship is learned during the preliminary learning phase. The data input to the preliminary model 800 is data related to material composition, and the data output from the preliminary model 800 is data related to material properties. Preliminary learning contributes to subsequent learning, and preliminary inference contributes to subsequent inference. Specifically, for example, the learning results of the preliminary model 800 contribute to shortening the time required to train the subsequent model 850 and improving the accuracy of the subsequent model 850 by making the training data used for training the subsequent model 850 more suitable for that training. In addition, inference using the preliminary model 800 narrows down the composition as the search range for manufacturing recipes, thus contributing to shortening the time required for inference (search for manufacturing recipes) using the subsequent model 850 and improving the accuracy of inference using the subsequent model 850.

[0090] The type of the preceding model 800 may be the same as any of the subsequent models 850, or it may be a different type from any of the subsequent models 850. The preceding model 800 may be linear regression (e.g., Ridge regression, Lasso regression, Elastic Net regression), logistic regression, SVM (Support Vector Machine), decision tree models (e.g., Random Forest, XGBoost (Xtreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine)), neural networks (e.g., CNN, RNN (Recurrent Neural Network), ResNet (Residual Network)), Bayesian optimization, k-NN (k-Nearest Neighbor), or any combination of two or more of these models. Preferably, for example, the preceding model 800 may be a random forest, XGBooster, LightGBM, CNN, Bayesian optimization, or any combination of two or more of these models. Both the preceding model 800 and the succeeding model 850 may be machine learning models, statistical models, or models that include both.

[0091] The following describes an example of learning and inference performed in this embodiment.

[0092] Figure 9 shows an example of learning according to this embodiment.

[0093] The pre-processing unit 801 trains the pre-model 800 using the pre-training data 900. The pre-training data 900 includes data representing the material composition and data representing the material properties for each material composition. Specifically, for each material composition, the pre-training data 900 includes values ​​for one or more composition items related to the material composition and values ​​for one or more property items related to the material properties for that material composition.

[0094] Multiple composition items represented by the preliminary learning data 900 are narrowed down to a subset of composition items. A contribution determination unit 910 is used for this narrowing down. The contribution determination unit 910 determines the degree to which each of the multiple composition items (e.g., multiple explanatory variables) contributes to the material properties (e.g., the objective variable) (e.g., contribution rate). The contribution determination unit 910 may be included in the preliminary processing unit 801, or it may be an external function of the preliminary processing unit 801 (e.g., a service provided by an external website, or a function in the MI platform system 100).

[0095] The pre-processing unit 801 instructs the contribution determination unit 910 to determine the degree of contribution for each of the multiple composition items represented by the pre-learning data 900. In other words, a contribution determination instruction is given to the contribution determination unit 910. The pre-learning data 900 is associated with this instruction, and as a result, the pre-learning data 900 may be input to the contribution determination unit 910. In response to this instruction, the contribution determination unit 910 determines (calculates) the degree of contribution to the material properties for each of the multiple composition items represented by the pre-learning data 900, based on the relationship represented by the pre-learning data 900 (a pair of values ​​for each of the multiple composition items and each of the values ​​for one or more characteristic items for each material composition). Based on the degree of contribution of each composition item, the composition items are narrowed down. This narrowing down may be performed by the pre-processing unit 801 based on the output data from the contribution determination unit 910 (degree of contribution for each composition item), or it may be performed by the contribution determination unit 910. In response to the instruction to determine contribution, the pre-processing unit 801 identifies the degree of contribution (or narrowed-down composition item) to the material properties for each of the multiple composition items represented by the pre-learning data 900 from the output data of the contribution determination unit 910.

[0096] Based on the preliminary training data 900 and the narrowed-down composition items, the subsequent training data 950 is generated. Specifically, for example, the subsequent training data 950 is data that includes at least a portion of the preliminary training data 900 and additional data. "At least a portion of the preliminary training data 900" is data from which data (values) for composition items other than the narrowed-down composition items have been removed. "Additional data" is data that includes values ​​for each item of one or more condition items and / or one or more feature items based on one or more composition items selected based on their contribution.

[0097] The subsequent training data 950 is input to the subsequent processing unit 811. The subsequent training data 950 includes a manufacturing recipe dataset 311, a material feature dataset 312, and a material property dataset 313 for each material composition. The subsequent processing unit 811 uses the subsequent training data 950 to train the first to third models 301 to 303.

[0098] The elements related to learning shown in Figure 9 will be explained in detail below.

[0099] Figure 10 shows an example of the preliminary training data 900.

[0100] The preliminary learning data 900 has an entry for each material composition. The entry includes information such as reference information 1001, composition information 1002, and property information 1003.

[0101] Reference information 1001 may include information that forms the basis for determining the importance of values ​​for at least one composition item (e.g., a coefficient of a variable) and the importance of values ​​for at least one characteristic item (e.g., a coefficient of a variable) in learning and / or inference.

[0102] For example, reference information 1001 includes information representing the data source and information representing the data type.

[0103] In Figure 10, "data source" refers to the source of composition information 1002 and characteristic information 1003 in an entry. Examples of data sources include "external" and "internal." If information obtained from "internal" sources is more reliable than information obtained from "external" sources, then in learning and / or inference, the importance of the value corresponding to at least one composition item (the value in composition information 1002) and / or the value corresponding to at least one characteristic item (the value in characteristic information 1003) may be higher for "internal" sources than for "external" sources. Thus, depending on the data source, the importance of the values ​​corresponding to composition items and / or the values ​​corresponding to characteristic items may be adjusted by the pre-processing unit 801 in at least one of the pre-learning and pre-inference stages, or by the post-processing unit 811 in at least one of the post-learning and post-inference stages. Note that the data source corresponding to "external" sources may be an external server of the MI platform system 100. For example, the data corresponding to "external" information (composition information 1002 and characteristic information 1003) can be the data included in the various data 150 (see Figure 1).

[0104] In Figure 10, "data type" refers to the type of composition information 1002 and characteristic information 1003 in the entry. For example, if all (or some) of the values ​​in the composition information 1002 and characteristic information 1003 in the entry are experimentally obtained values, the data type is "experimental value". Also, for example, if all (or some) of the values ​​in the composition information 1002 and characteristic information 1003 in the entry are calculated values, the data type is "calculated value". If the "experimental value" is more reliable than the "calculated value", then in learning and / or inference, the importance of the value corresponding to at least one composition item and / or the importance of the value corresponding to at least one characteristic item may be higher for the "experimental value" than for the "calculated value". In other words, depending on the value as the data type, the importance of the value corresponding to the composition item and / or the importance of the value corresponding to the characteristic item may be adjusted by the pre-processing unit 801 in at least one of the pre-learning and pre-inference stages, or by the post-processing unit 811 in at least one of the post-learning and post-inference stages.

[0105] The composition information 1002 includes values ​​for each of several composition items. Examples of composition items other than the composition formula, as shown in Figure 10, include elemental characteristics (e.g., atomic weight, primitive radius), crystal structure, effective mass, carrier density, and theoretical density.

[0106] The characteristic information 1003 includes values ​​for each of one or more characteristic items relating to the material properties. An example of a characteristic item, as shown in Figure 10, is thermal conductivity.

[0107] Figure 11 shows an example of the subsequent training data 950.

[0108] The subsequent learning data 950 has an entry for each material composition. In addition to reference information 1101, composition information 1102, and property information 1103, each entry has condition information 1104 and feature information 1105. In each entry, the reference information 1101, composition information 1102, and property information 1103 may be the same as the reference information 1001, composition information 1002, and property information 1003 corresponding to that entry.

[0109] Each entry includes information about the manufacturing recipe, such as composition information 1102 and condition information 1104.

[0110] Condition information 1104 is information regarding conditions as at least one element of the manufacturing recipe other than the material composition. Here, "conditions" may be conditions relating to the synthesis process, for example, conditions relating to at least one of the manufacturing method (manufacturing) and materials. Condition information 1104 includes values ​​for each of one or more condition items, which are one or more items relating to the conditions.

[0111] The feature information 1105 includes values ​​for each of one or more feature items related to material characteristics.

[0112] For each entry, the characteristic information 1103 may be included in the material characteristic dataset 313, the composition information 1102 and condition information 1104 may be included in the manufacturing recipe dataset 311, and the feature information 1105 may be included in the material feature dataset 312.

[0113] Figure 12 shows an example of composition refinement.

[0114] As described above, the preliminary learning data 900 represents the composition / property relationship, which is the relationship between material composition and material properties. Since the preliminary learning data 900 can include values ​​for multiple composition items regarding material composition, it also represents the relationships between composition items. Furthermore, since the preliminary learning data 900 can include values ​​for multiple property items regarding material properties, it also represents the relationships between property items. Therefore, the "composition / property relationship" may represent not only the relationship between at least one composition item and at least one property item, but also the relationships between composition items and / or between property items. In response to a contribution determination instruction, the contribution determination unit 910 calculates the contribution of each of the multiple composition items (for example, composition items other than the composition formula) represented by the preliminary learning data 900, based on the composition / property relationship represented by the preliminary learning data 900 associated with the instruction, to the material properties. Existing methods may be used to calculate the contribution of each composition item to the material properties based on the composition / property relationship. By using the contribution determination unit 910, the multiple composition items represented by the prior learning data 900 are narrowed down to one or more composition items with a large contribution. "One or more composition items with a large contribution" may be composition items whose contribution is greater than a predetermined threshold, composition items whose contribution is greater than the statistical value (e.g., mean) of the multiple contributions obtained, or composition items whose contribution falls into the top N (e.g., N=3).

[0115] Figure 13 is an explanatory diagram illustrating an example where composition refinement contributes to subsequent learning.

[0116] As shown in the example in Figure 13, as a result of the compositional refinement illustrated in Figure 12, atomic weight, crystal structure, effective mass, and carrier density are selected from among multiple compositional items as some of the compositional items with a large contribution. At least some of the conditional items adopted in the subsequent training data 950 are conditional items that have a high correlation (e.g., relationship) with the selected compositional items from among multiple conditional items. Similarly, at least some of the feature items adopted in the subsequent training data 950 are feature items that have a high correlation (e.g., relationship) with the selected compositional items from among multiple feature items. In other words, the conditional items and / or feature items adopted in the subsequent training data 950 may include conditional items and / or feature items that are adopted regardless of the selected compositional items. This is because conditional items and feature items (typically items related to higher-order structure) that are not present in the preceding training are adopted, and it is difficult to determine them based on the compositional items. Note that in both the first and second embodiments, there may be different types of items even if they have the same item name. For example, an item with the item name "effective mass" may be adopted as both a compositional item and a feature item.

[0117] The characteristics of the composition / property relationship, which is the relationship between material composition and material properties, may differ depending on the preceding training data 900. Therefore, the contribution determined (calculated) for each composition item may differ depending on the preceding training data 900.

[0118] Figure 14 shows an example of reasoning according to this embodiment.

[0119] Data called "composition formula list 1400" (illustrated in Figure 15) is input. Composition formula list 1400 represents multiple composition formulas. At least one of these multiple composition formulas may be data from an external server, for example, data included in various data 150 (see Figure 1). At least one of these multiple composition formulas may be a composition formula that is automatically predicted (created) based on one composition formula.

[0120] The pre-processing unit 801 instructs the composition acquisition unit 1405 to acquire values ​​for each of the multiple composition items represented by the composition formula list 1400. The composition formula list 1400 may be associated with this instruction.

[0121] In response to this instruction, the composition acquisition unit 1405 acquires (e.g., predicts) a value for each of one or more composition items for each composition formula represented by the composition formula list 1400, and outputs composition information 1410 (exemplified in Figure 16) containing the acquired values ​​for each composition item for each composition formula. The composition acquisition unit 1405 may be included in the pre-processing unit 801, or it may be an external function of the pre-processing unit 801 (for example, provided by an external website, or a function in the MI platform system 100). Existing methods may be used to acquire the values ​​for each composition item for each composition formula.

[0122] The pre-processing unit 801 generates pre-inference input data 1420 (exemplified in Figure 17) using the composition formula list 1400 and composition information 1410. The pre-inference input data 1420 includes each composition formula represented by the composition formula list 1400 and the values ​​for each composition item included in the composition information 1410 for each composition formula. The configuration of the pre-inference input data 1420 may be the same as that of the pre-learning data 900, except that the reference information 1001 and characteristic information 1003 are removed from each entry.

[0123] The pre-processing unit 801 inputs the pre-inference input data 1420 to the trained pre-model 800, thereby obtaining predicted values ​​for each of one or more characteristic items for each compositional formula represented by the pre-inference input data 1420. In other words, the pre-model 800 outputs predicted values ​​for each of one or more characteristic items for each compositional formula, based on the input of each compositional formula and the values ​​for each compositional item for each compositional formula. The pre-inference output data 1430 (exemplified in Figure 18), which includes the predicted values ​​obtained for each compositional formula, is generated by the pre-processing unit 801.

[0124] The subsequent processing unit 811 narrows down the preceding inference output data 1430 to subsequent inference input data 1450 that includes predicted values ​​(characteristic values) that satisfy the requirements of the target characteristics and the corresponding composition formulas, and inputs the subsequent inference input data 1450 to the subsequent model 850. In other words, the search range for the manufacturing recipe is narrowed down to the range of composition formulas represented by the subsequent inference input data 1450. Note that "predicted values ​​that satisfy the requirements of the target characteristics" may be predicted values ​​greater than a predetermined threshold, predicted values ​​greater than the statistical value (e.g., mean) of multiple obtained predicted values, or predicted values ​​that fall within the upper M (e.g., M=3). The subsequent inference input data 1450 may include information other than composition formulas and predicted values. When the subsequent inference input data 1450 is input to the subsequent model 850, the processing shown in Figure 5 may be performed. For example, inference using the first model 301, or inference using the second and third models 302 and 303 may be performed. As a result, the manufacturing recipe data 1460 may be obtained as data including the output data of the subsequent model 850. The manufacturing recipe data 1460 may include one or more manufacturing recipe datasets 311.

[0125] Although several embodiments have been described above, these are merely illustrative examples for explaining the present invention and are not intended to limit the scope of the invention to these embodiments only. The present invention can be carried out in various other forms.

[0126] For example, the AI ​​unit 140 shown in Figure 1 may have the same function as the AI ​​unit 108 or AI unit 108X described above.

[0127] Furthermore, for example, the "experimental data" collected by the collection unit 101 may include at least some of the following: data from combinatorial experiments, data obtained from an electronic laboratory notebook 131 where past experimental data is stored, and data obtained from external sites, etc. The "experimental data" may also include computational data (for example, data predicted by first-principles calculations or separate machine learning) and linguistic data. At least one of the datasets 311 to 313 may also include computational data or linguistic data. In addition, images may be generated during training, but in this embodiment, image generation is not required during training. This is expected to result in a lower computational load and a simpler model.

[0128] Furthermore, the above explanation can be summarized as follows. The summary below may include supplementary explanations and variations of the above explanation.

[0129] The material creation support system 10 may include, for example, an interface device 201 connected to a user device (e.g., a researcher terminal 11A) which is a device having an input device and a display device, a storage device 202 in which models 301 to 303 are stored, and a processor 203 connected to the interface device 201 and the storage device 202.

[0130] The processor 203 may implement the AI ​​unit 108 and the IF unit (interface unit) 109 by, for example, executing a computer program. The AI ​​unit 108 performs inference using either the first model 301 or the second model 302 and the third model 303, depending on the accuracy (e.g., prediction accuracy) of at least one of the first to third models 301 to 303, and generates or updates a list 600 (an example of recipe characteristic data, which is data representing the association between manufacturing recipes and material properties). The IF unit 109 receives a query specifying material properties, identifies a manufacturing recipe corresponding to the properties specified in the query based on the list 600, and provides the identified manufacturing recipe.

[0131] Based on the correlation between material features and material properties, a second model 302 is prepared, and a third model 303 is prepared to enable the prediction of a material property dataset 313 from a manufacturing recipe dataset 311 via an intermediate dataset called a material feature dataset 312. The existence of the second and third models 302 and 303 allows for the proposal of manufacturing recipes for creating materials with desired material properties, even if the accuracy of the first model 301 is insufficient due to insufficient experimental data or other reasons.

[0132] The AI ​​unit 108 may generate or update list 600 in inference using the first model 301 if the accuracy of the first model 301 is equal to or greater than the first accuracy threshold, or if the accuracy of the first model 301 is higher than the accuracy of the second and third models 302 and 303. This allows inference for generating or updating list 600, which is used to provide manufacturing recipes, to be performed with less computational load (without input / output of intermediate datasets as material feature dataset 312). Note that "accuracy of the second model 302 and third model 303" refers to the accuracy when both the second model 302 and the third model 303 are used, not the individual accuracy of the second model 302 and the third model 303.

[0133] The AI ​​unit 108 may train the first model 301 using the manufacturing recipe dataset 311 and material properties dataset 313 identified or generated in the training or inference of the second and third models 302 and 303. For example, in the training or inference of the second and third models 302 and 303, it may be identified that it is preferable for a manufacturing recipe to have or not have a condition item called synthesis time. In this way, in the training or inference of the second and third models 302 and 303, the condition items may be expanded or reduced, or the values ​​for the condition items may be expanded or reduced, and as a result, it is expected that the manufacturing recipe dataset 311, which is preferable for training the first model 301, will increase. Therefore, it is expected that the accuracy of the first model 301 will be improved.

[0134] For example, since the training dataset is obtained from experimental data, if the experimental data is insufficient, the training dataset will also be insufficient, and as a result, the accuracy of models 301 to 303 may also be insufficient. Furthermore, it takes time for the experimental data to become sufficient, and it is difficult to guarantee when it will become sufficient. Therefore, the AI ​​unit 108 may perform inference using the first model 301, and / or inference using the second and third models 302 and 303, in parallel with the training of models 301 to 303. For example, in conjunction with the switching of training (depending on which model's training has been completed), the AI ​​unit 108 may switch (select) the model to be used for inference.

[0135] In training the second model 302, the AI ​​unit 108 may identify or generate one or more material feature datasets 312 from a plurality of material feature datasets 312 as factors that cause the accuracy of the second model 302 to be equal to or greater than the second accuracy threshold. In training the third model 303, the AI ​​unit 108 may identify or generate one or more manufacturing recipe datasets 311 from a plurality of manufacturing recipe datasets 311 that correspond to one or more material feature datasets 312 identified or generated in training or inference of the second model 302. The AI ​​unit 108 may train the first model using one or more manufacturing recipe datasets 311 identified or generated in training or inference of the third model 303 and one or more material property datasets 313 corresponding to said one or more manufacturing recipe datasets 311. As a result, the material feature datasets 312 are narrowed down in training or inference of the third model 303, so that the accuracy of the third model 303 can be improved with a small computational load. Furthermore, since the manufacturing recipe dataset 311 corresponding to the refined material feature dataset 312 is used to train the first model 301, the accuracy of the first model 301 can be improved with a small computational load.

[0136] The material feature dataset 312 may be a dataset composed of numerical feature quantities for each feature item belonging to the material features. This reduces the computational load required for training and inference, and simplifies the second and third models 302 and 303.

[0137] When retraining is performed, the AI ​​unit 108 may discard list 600 and generate or update list 600 anew in inference using the first model 301 after retraining, and / or in inference using the second model 302 and the third model 303 after retraining. This prevents the IF unit 109 from providing recipes with lists that were generated or updated in inference using older models before their accuracy was improved.

[0138] Furthermore, as another variation, the following variation may be made. That is, for example, as shown in Figure 7, if the accuracy of the first model 301 is less than the first accuracy threshold (for example, if the accuracy of the first model 301 does not reach or exceed the first accuracy threshold even after training the first model 301 a predetermined number of times), the AI ​​unit 108 may, in addition to the manufacturing recipe dataset 311 identified or generated in the training of the second and third models 302 and 303, use the material feature dataset 312, which corresponds to the manufacturing recipe dataset 311 and was identified or generated in the training or inference of the second model 302, as input to the first model 301 to train the first model 301 (learning the correspondence between the set of manufacturing recipes and material features and the material properties). This is expected to increase the possibility of raising the accuracy of the first model 301 to or exceed the first accuracy threshold. In this case, the AI ​​unit 108 may generate or update a list 600 (i.e., a list 600 recording the relationship between manufacturing recipes, material features, and material properties) in inference using the first model 301, in addition to manufacturing recipes, by associating material features with material properties. The IF unit 109 may receive a query that specifies material features in addition to material properties, identify the material properties and manufacturing recipes corresponding to the material features specified in the query based on the list 600, and provide the identified manufacturing recipes.

[0139] One possible use case is as follows: The material characteristics may include (1) density and porosity quantified by the Archimedes method, and (2) SEM image features (e.g., composition area ratio and composition area deviation). The feature items (feature factors) that may influence the target properties (material properties desired by the researcher) may be porosity, crystal circumference, composition area ratio, and composition area deviation ("composition area deviation" is a feature that represents the variation in the composition ratio of multiple divided regions constituting a single image (in this case, an SEM image)). The first model 301 may be Ridge regression (the input may be manufacturing conditions such as composition ratio, material transfer, and firing temperature, and the output may be bending strength and thermal expansion coefficient). The second model 302 may be Ridge regression or LightGBM (the input may be the above-mentioned (1) density and porosity, and (2) SEM image features, and the output may be bending strength and thermal expansion coefficient). The third model 303 may be Ridge regression (the input may be manufacturing conditions such as composition ratio, material particle size, and firing temperature, and the output may be (1) density and porosity, and (2) SEM image features). A manufacturing recipe may be provided in response to an inquiry from researcher 5A (an inquiry specifying material properties) (it may be displayed on researcher terminal 11A). Furthermore, each of models 301 to 303 may include a regression equation for each element as an output (e.g., the dependent variable). In other words, each of models 301 to 303 may include one or more regression equations. Also, the input dataset may be common to all of models 301 to 303, but typically the explanatory variables and weightings differ for each regression equation in the model.

[0140] Furthermore, the significance of the process illustrated in Figure 5 is as follows, for example.

[0141] In other words, the objectives of the inference can be listed as (1) finding a manufacturing recipe for producing a material with desired material properties, and (2) finding candidate manufacturing recipes for adding experimental data in order to improve model accuracy. The inference in S504 in Figure 5 corresponds to objective (1) (because model training is complete). On the other hand, the inferences in S506, S508 and S509 in Figure 5 correspond to objective (2) (because model training is not complete).

[0142] In the inference corresponding to objective (2), a useful or effective dataset for model learning (e.g., manufacturing recipe) is selected from the combinations identified (e.g., combination of manufacturing recipe and material properties). Based on this dataset, a dataset obtained from actual experiments is added as a training dataset, and model learning is performed using the added dataset. Because the dataset is augmented in this way, the model can be learned with greater accuracy.

[0143] S505: If the answer is Yes, the accuracy of the inference using the second model 302 and the third model 303 can be considered sufficient. Therefore, an additional training dataset can be prepared based on the dataset input and output in that inference. In addition to the dataset input and output in the inference, a dataset conceived and generated by the researcher (e.g., the experimenter) may also be used to prepare the additional training dataset.

[0144] On the other hand, in the case of S505: No, the accuracy of the inference using the second model 302 and the third model 303 can be considered insufficient. Therefore, it is difficult to prepare additional training datasets based solely on the input and output datasets used in that inference, and it is acceptable to use datasets devised and generated by the researchers in order to prepare additional training datasets.

[0145] Since the training of the second model 302 takes place before the training of the third model 303, it is more efficient for the judgment in S505 to be performed before S507. Note that the order of judgments is not limited to the order exemplified in Figure 5 (i.e., S503 → S505 → S507).

[0146] Based on the above explanation, the following expressions are possible.

[0147] The AI ​​unit 108 may perform inference using the second model 302 and the third model 302 if the accuracy of the first model 301 is less than the first threshold, or if the accuracy of the first model 301 is lower than the accuracy of the second model 302 and the third model 303.

[0148] The AI ​​unit 108 may train the second model 302 before the third model 303. The AI ​​unit 108 may prioritize using the dataset input or output in inference using the second model 302 and the third model 303 to train at least one of the first to third models in the second case over the first case. First case: The case in which inference using the second model 302 and the third model 303 is performed when the accuracy of the third model 303 is less than the third threshold. Second case: The case in which inference using the second model 302 and the third model 303 is performed when the accuracy of the third model 303 is equal to or greater than the third threshold.

[0149] If the accuracy of the second model 302 is less than the second threshold, and the accuracy of the second model 302 and the third model 303 is higher than the accuracy of the first model 301, the AI ​​unit 108 may perform inference using the second model 302 and the third model 303. If the accuracy of the second model 302 and the third model 303 is the same as or lower than the accuracy of the first model 301, the AI ​​unit 108 may perform inference using the first model 301.

[0150] Incidentally, the processor 203 may implement the AI ​​unit 108X instead of the AI ​​unit 108 by, for example, executing a computer program. The AI ​​unit 108X includes a pre-processing unit 801 and a post-processing unit 811. The AI ​​unit 108X is expected to search for manufacturing recipes more efficiently than the AI ​​unit 108. Either the pre-processing unit 801 or the post-processing unit 811 may be located outside the system 10.

[0151] The pre-processing unit 801 performs pre-learning and pre-inference. The post-processing unit 811 performs post-learning and post-inference.

[0152] The preliminary learning includes constructing a preliminary model 800, which is a model representing the relationship between material composition and material properties, using preliminary learning data 900, which is data containing data representing material composition and data representing material properties for each of a plurality of material compositions. The subsequent learning includes constructing a subsequent model 850, which is a model representing the relationship between manufacturing recipe and material properties, using subsequent learning data 950, which is data containing data representing material properties and data representing a manufacturing recipe that includes each of at least some of the plurality of material compositions represented by the preliminary learning data 900. Model construction as used herein may include model learning. In the second embodiment, the relationship between the manufacturing recipe and material properties is represented as the first model 301, or as a pair of the third model 303 and the second model 302. In the subsequent model 850, either the first model 301 or the pair of the second and third models 302 and 303 may be omitted. "Data representing material composition" may include data containing values ​​for at least one of the one or more composition items related to material composition. "Data representing material properties" may include data containing values ​​for at least one of the one or more property items related to material properties. "Data representing manufacturing recipe" may include data containing values ​​for at least one of the one or more condition items related to the manufacturing process, which are items related to the manufacturing process and materials.

[0153] The preliminary inference includes inputting preliminary inference input data, which is data containing values ​​for one or more first composition items for each of one or more material compositions, into a constructed (trained) preliminary model 800 to output preliminary inference output data 1430. The preliminary inference output data 1430 includes, for each of one or more material compositions, predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items, and / or predicted values ​​for one or more characteristic items of the material properties. That is, as shown in the example in Figure 18, the preliminary inference output data 1430 includes a predicted value for thermal conductivity, which is one characteristic item of the material properties, but instead of or in addition to that predicted value, the preliminary inference output data 1430 may include, for each of one or more material compositions, predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items in the preliminary inference. Here, "one or more first compositional elements" may be, for example, a chemical formula and elemental characteristics, and "one or more second compositional elements" may include elemental characteristics as the same second compositional element as one of the first compositional elements, but may not include a chemical formula as one of the first compositional elements as a second compositional element. "One or more first compositional elements" and "one or more second compositional elements" may have some matching compositional elements. In other words, "one or more first compositional elements" and "one or more second compositional elements" do not have to be exactly the same. The subsequent inference process involves inputting subsequent inference input data 1450, which is data that includes values ​​for at least one composition item and values ​​for one or more property items of the target material properties, for each of the one or more material compositions whose predicted values ​​in the preceding inference output data 1430 satisfy the requirements, into a constructed (trained) subsequent model 850, thereby outputting manufacturing recipe data 1460, which is data that represents a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data 1450 for materials having the target material properties.

[0154] Among manufacturing recipes, material composition is one of the factors that greatly influences material properties. In the search for a manufacturing recipe, there are many explanatory variables when material properties are the dependent variable, such as manufacturing conditions and material conditions. However, by narrowing down the material composition in the earlier stages, it is expected that manufacturing recipes can be searched accurately in a shorter amount of time in the later stages.

[0155] Multiple items related to the manufacturing recipe may include one or more composition items and one or more condition items. The pre-processing unit 801 may select one or more composition items from among the multiple composition items represented by the pre-training data 900 based on the degree of contribution of the pre-training data 900 to the material properties. The post-training data may include values ​​for condition items determined based on the selected one or more composition items. The selection of one or more composition items based on the degree of contribution corresponds to composition refinement. Determining the condition items based on the selected one or more composition items may be done manually, or it may be done automatically by the post-processing unit 811 using a rule-based method or other method. In the example shown in Figure 13, the selected one or more composition items are "atomic weight," "crystal structure," "effective mass," and "carrier density," and the determined condition items are "production method," "synthesis temperature," and "raw material particle size." In this way, the subsequent training data 950 becomes training data (supervised data) that is more suitable for subsequent training, so an improvement in the accuracy of the subsequent model 850 is expected. Furthermore, a reduction in the number of training iterations required to achieve a desired accuracy or higher for the subsequent model 850 is expected.

[0156] The subsequent model 850 may include a first model 301 representing the relationship between a manufacturing recipe and material properties, a second model 302 representing the relationship between material features and material properties, and a third model 303 representing the relationship between a manufacturing recipe and material features. In subsequent inference, if the accuracy of the first model 301 is equal to or greater than a first threshold, or if the accuracy of the first model 301 is higher than the accuracy of the second model 302 and the third model 303, the subsequent processing unit 811 may input subsequent inference input data 1450 to the first model 301 of the subsequent model 850. This is expected to improve the accuracy of the first model 301 representing the relationship between a manufacturing recipe and material properties, and therefore, it is possible to increase the likelihood that the first model 301, which does not require prediction of material features in the search for a manufacturing recipe, will be used. In other words, it is expected that the transition from inference using the second and third models 302 and 303 to inference using the first model 301 will be accelerated. Furthermore, it is expected that the number of training iterations required to achieve the desired accuracy or higher for the first model 301 will be reduced.

[0157] The subsequent model 850 may include second and third models 302 and 303 in place of or in addition to the first model 301. The subsequent training data 950 represents a set of material characteristics, in addition to material properties and manufacturing recipes, for each of at least some of the material compositions represented by the preceding inference output data 1430. In the subsequent training data 950, at least some of the condition items related to the manufacturing recipe may be condition items determined based on one or more selected composition items, and at least some of the condition items related to material characteristics may be feature items determined based on one or more selected composition items. The subsequent processing unit 811 may input the subsequent inference input data 1450 to the second model 302 and the third model 303 in subsequent inference. Since the subsequent training data 950 becomes training data more suitable for subsequent training, an improvement in the accuracy of the second and third models 302 and 303 is expected, and consequently, an improvement in the accuracy of inference using the second and third models 302 and 303 is expected. Furthermore, a reduction in the number of training iterations required to achieve a desired accuracy or higher for the second and third models 302 and 303 is expected. Note that the second and third models 302 and 303 may be used in subsequent inference when the accuracy of the first model 301 is below the first threshold, or when the accuracy of the first model 301 is lower than the accuracy of the second and third models 302 and 303.

[0158] The subsequent training data 950 may include data representing material properties and data representing manufacturing recipes for each of the material compositions among the multiple material compositions represented by the preceding training data 900. It is difficult to associate manufacturing recipes with all of the multiple material compositions represented by the preceding training data 900 through experiments, etc. Therefore, the subsequent training data 950 will contain data for fewer material compositions than those represented by the preceding training data 900. However, as described above, the condition items and feature items in the subsequent training data 950 are items determined based on one or more selected composition items. For example, if the material property is thermal conductivity, since the composition items reflected in the subsequent training data 950 are limited, it is possible to determine which feature items affect thermal conductivity as an intermediate evaluation. This is expected to improve the accuracy of recipe search using the second and third models 302 and 303.

[0159] Furthermore, a material manufacturing system including the material creation support system 10, or a material manufacturing system that works in conjunction with the material creation support system 10, may be constructed. Such a material manufacturing system may include, for example, a manufacturing apparatus such as the manufacturing apparatus 111 shown in Figure 1, and may also include, in addition to the manufacturing apparatus, an evaluation apparatus such as the evaluation apparatus 112 shown in Figure 1. The following material manufacturing method may be realized by such a material manufacturing system. That is, the material manufacturing method may include a preliminary inference step in which a computer (for example, the material creation support system 10) performs preliminary inference using a preliminary model 800 constructed in preliminary learning, a later inference step in which a computer performs later inference using a later model 850 constructed in later learning, and a material manufacturing step in which a material is manufactured according to the manufacturing conditions represented by the manufacturing recipe data output in later inference. The material manufacturing method may further include a preliminary learning step in which a computer performs preliminary learning, and a later learning step in which a computer performs later inference.

[0160] For example, in Figure 1, manufacturing recipe data may be input from the IF unit 109 to the experimental system 110. Specifically, for example, manufacturing recipe data may be transmitted from system 100 to the experimental system 110 via a communication network, and in the experimental system 110, the manufacturing recipe data may be input to a control device (e.g., a computer) in the experimental system 110. The control device may manufacture the material by controlling one or more condition items for controlling each of one or more manufacturing devices 111 according to the manufacturing recipe represented by the manufacturing recipe data (for example, values ​​for each condition item related to the manufacturing conditions). The control device may also perform an evaluation (e.g., measurement) of the manufactured material by controlling one or more condition items for controlling each of one or more evaluation devices 112 according to the manufacturing recipe represented by the manufacturing recipe data. In other words, the conditions related to the manufacturing method or material in the manufacturing recipe may also include conditions related to evaluation (measurement).

[0161] 10...Material creation support system 100...MI platform system

Claims

1. The system comprises a pre-processing unit that performs pre-learning and pre-inference, and a post-processing unit that performs post-learning and post-inference, wherein the pre-learning includes constructing a pre-model, which is a model representing the relationship between material composition and material properties, using pre-learning data, which is data including data representing material composition and data representing material properties for each of a plurality of material compositions, and the post-learning includes constructing a post-model, which is a model representing the relationship between a manufacturing recipe and material properties, using post-learning data, which is data including data representing material properties and data representing a manufacturing recipe that includes each of at least some of the plurality of material compositions, wherein the data representing material composition includes a value for at least one of one or more composition items relating to material composition, the data representing material properties includes a value for at least one of one or more property items relating to material properties, and the data representing a manufacturing recipe includes a value for at least one of one or more condition items relating to the manufacturing process among the manufacturing process and materials. The preceding preliminary inference includes outputting preliminary inference output data by inputting preliminary inference input data, which is data containing values ​​for one or more first composition items for each of one or more material compositions, into the preceding model, wherein the preceding inference output data includes at least one of the following (A) and (B) for each of the one or more material compositions: (A) predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items; (B) predicted values ​​for one or more characteristic items of material properties. The subsequent inference includes outputting manufacturing recipe data, which is data containing a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the desired material properties, by inputting subsequent inference input data, which is data containing values ​​for at least one composition item for each of the one or more material compositions whose predicted values ​​included in the preliminary inference output data satisfy the requirements, into the subsequent model.

2. The material creation support system according to claim 1, wherein the multiple items relating to the manufacturing method recipe include one or more composition items and one or more condition items, the preceding processing unit selects one or more composition items from among the multiple composition items represented by the preceding learning data based on the degree of contribution of the preceding learning data to the material properties, and the subsequent learning data includes values ​​for condition items determined based on the selected one or more composition items.

3. The material creation support system according to claim 2, wherein the subsequent model includes a first model representing the relationship between a manufacturing recipe and material properties, a second model representing the relationship between material characteristics and material properties, and a third model representing the relationship between a manufacturing recipe and material characteristics, and the subsequent processing unit inputs the subsequent inference input data to the first model of the subsequent model if, in the subsequent inference, the accuracy of the first model is equal to or greater than a first threshold, or the accuracy of the first model is higher than the accuracy of the second and third models.

4. The material creation support system according to claim 3, wherein the multiple items relating to the manufacturing recipe include one or more composition items and one or more condition items, the pre-processing unit selects one or more composition items from among the multiple composition items represented by the pre-processing data based on the degree of contribution to the material properties represented by the pre-processing data, the post-processing data is data that includes data representing material properties and data representing the manufacturing recipe for each of the at least some material compositions, as well as data representing material features, in the post-processing data, at least some of the condition items relating to the manufacturing recipe are condition items determined based on the selected one or more composition items, at least some of the feature items relating to material features are feature items determined based on the selected one or more composition items, and the post-processing unit inputs the post-processing input data to the second model and the third model in the post-processing inference if the accuracy of the first model is less than a first threshold, or if the accuracy of the first model is lower than the accuracy of the second model and the third model.

5. The material creation support system according to claim 2, wherein the subsequent model includes a second model representing the relationship between material characteristics and material properties, and a third model representing the relationship between a manufacturing recipe and material characteristics, wherein the multiple items relating to the manufacturing recipe include one or more composition items and one or more condition items, the preceding processing unit selects one or more composition items from among the multiple composition items represented by the preceding learning data based on the contribution of the preceding learning data to the material properties, the subsequent learning data is data that includes data representing material properties and data representing a manufacturing recipe for each of the at least some material compositions, as well as data representing material characteristics, in which case at least some of the condition items relating to the manufacturing recipe are condition items determined based on the selected one or more composition items, and at least some of the feature items relating to material characteristics are feature items determined based on the selected one or more composition items, and the subsequent processing unit inputs the subsequent inference input data to the second model and the third model in the subsequent inference.

6. The material creation support system according to claim 2, wherein the subsequent learning data includes data representing material properties and data representing manufacturing recipes for each of the partial material compositions among the plurality of material compositions represented by the preceding learning data.

7. A material creation support method that performs pre-learning, post-learning, pre-inference, and post-inference using a computer, wherein the pre-learning includes constructing a pre-model, which is a model representing the relationship between material composition and material properties, using pre-learning data, which is data including data representing material composition and data representing material properties for each of a plurality of material compositions, wherein the post-learning includes constructing a post-model, which is a model representing the relationship between manufacturing recipe and material properties, using post-learning data, which is data representing a set of data including data representing material properties and data representing a manufacturing recipe that includes each of at least some of the plurality of material compositions, wherein the data representing material composition includes a value for at least one of one or more composition items relating to material composition, the data representing material properties includes a value for at least one of one or more property items relating to material properties, and the data representing a manufacturing recipe includes a value for at least one of one or more condition items relating to the manufacturing conditions of at least one of the manufacturing conditions of the manufacturing method and materials. The preceding preliminary inference includes inputting preliminary inference input data, which is data containing values ​​for one or more first composition items for each of one or more material compositions, into the preceding model to output preliminary inference output data, wherein for each of the one or more material compositions, the preliminary inference output data includes at least one of the following (A) and (B): (A) predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items; (B) predicted values ​​for one or more characteristic items of material properties. The subsequent inference includes inputting subsequent inference input data, which is data containing values ​​for at least one composition item and values ​​for one or more characteristic items of the target material properties, into the subsequent model, thereby outputting manufacturing recipe data, which is data representing a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the target material properties.

8. A computer program that causes a computer to perform pre-learning, post-learning, pre-inference, and post-inference, wherein the pre-learning includes constructing a pre-model, which is a model representing the relationship between material composition and material properties, using pre-learning data, which is data including data representing material composition and data representing material properties for each of a plurality of material compositions, wherein the post-learning includes constructing a post-model, which is a model representing the relationship between a manufacturing recipe and material properties, using post-learning data, which is data representing a set of data including data representing material properties and data representing a manufacturing recipe that includes the material composition for each of at least some of the plurality of material compositions, wherein the data representing material composition includes a value for at least one of one or more composition items relating to material composition, the data representing material properties includes a value for at least one of one or more property items relating to material properties, and the data representing a manufacturing recipe includes a value for at least one of one or more condition items relating to the manufacturing conditions of the manufacturing method and materials. The preceding preliminary inference includes outputting preliminary inference output data by inputting preliminary inference input data, which is data containing values ​​for one or more first composition items for each of one or more material compositions, into the preceding model, wherein the preceding inference output data includes at least one of the following (A) and (B) for each of the one or more material compositions: (A) predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items; (B) predicted values ​​for one or more characteristic items of material properties. The subsequent inference includes outputting manufacturing recipe data, which is data containing a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the desired material properties, by inputting subsequent inference input data, which is data containing values ​​for at least one composition item for each of the one or more material compositions whose predicted values ​​included in the preliminary inference output data satisfy the requirements, into the subsequent model.

9. A recording medium storing a computer program that causes a computer to perform pre-learning, post-learning, pre-inference, and post-inference, wherein the pre-learning includes constructing a pre-model, which is a model representing the relationship between material composition and material properties, using pre-learning data, which is data including data representing material composition and data representing material properties for each of a plurality of material compositions, the post-learning includes constructing a post-model, which is a model representing the relationship between a manufacturing recipe and material properties, using post-learning data, which is data representing a set of data including data representing material properties and data representing a manufacturing recipe that includes the material composition for each of at least some of the plurality of material compositions, the data representing material composition includes a value for at least one of the one or more composition items that relate to material composition, the data representing material properties includes a value for at least one of the one or more property items that relate to material properties, and the data representing a manufacturing recipe includes a value for at least one of the one or more condition items that relate to the conditions of the manufacturing method among the manufacturing method and materials. The preceding preliminary inference includes outputting preliminary inference output data by inputting preliminary inference input data, which is data containing values ​​for one or more first composition items for each of one or more material compositions, into the preceding model, wherein the preliminary inference output data includes at least one of the following (A) and (B) for each of the one or more material compositions: (A) predicted values ​​for one or more second composition items, which include the same or different composition items as at least one of the one or more first composition items; (B) predicted values ​​for one or more characteristic items of material properties. The subsequent inference includes outputting manufacturing recipe data, which is data containing a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the desired material properties, by inputting subsequent inference input data, which is data containing values ​​for at least one composition item for each of the one or more material compositions whose predicted values ​​included in the preliminary inference output data satisfy the requirements, into the subsequent model, wherein the subsequent inference includes outputting a manufacturing recipe data that represents a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the desired material properties.

10. The process comprises: a pre-inference step in which a computer performs pre-inference using a pre-model constructed in pre-learning; a post-inference step in which a computer performs post-inference using a post-model constructed in post-learning; and a material manufacturing step in which a material is manufactured according to the manufacturing conditions represented by the manufacturing recipe data output in the post-inference, wherein the pre-learning includes constructing the pre-model, which is a model representing the relationship between material composition and material properties, using pre-learning data which is data including data representing material composition and data representing material properties for each of a plurality of material compositions; the post-learning includes constructing the post-model, which is a model representing the relationship between manufacturing recipe and material properties, using post-learning data which is data representing a set of data including data representing material properties and data representing a manufacturing recipe that includes each of at least some of the plurality of material compositions; the data representing material composition includes a value for at least one of one or more composition items relating to material composition; and the data representing material properties includes a value for at least one of one or more property items relating to material properties. The data representing the manufacturing recipe includes a value for at least one of one or more condition items, which are items relating to the conditions of the manufacturing process among the manufacturing process and materials; the preliminary inference includes outputting preliminary inference output data by inputting preliminary inference input data, which is data including values ​​for one or more first composition items for each of one or more material compositions, into the preliminary model; the preliminary inference output data includes at least one of the following (A) and (B) for each of the one or more material compositions: (A) predicted values ​​for one or more second composition items, which include composition items that are the same as or different from at least one of the one or more first composition items; (B) predicted values ​​for one or more characteristic items of the material properties.A material manufacturing method comprising: inputting subsequent inference data into a subsequent model, which is data that includes, for each of the one or more material compositions among the one or more material compositions whose predicted values ​​included in the preceding inference output data satisfy the requirements, a value for at least one composition item and a value for one or more property items of the target material properties, thereby outputting manufacturing recipe data that represents a manufacturing recipe searched from the range of material compositions represented by the subsequent inference input data for a material having the target material properties.

11. The material manufacturing method according to claim 10, further comprising a preliminary learning step in which the preliminary learning is performed by a computer, and a subsequent learning step in which the subsequent inference is performed by a computer.

Images