Creation method of initial experimental data and experiment planning method

Abstract

To provide a creation method of initial experimental data and an experiment planning method for creating proper initial experimental data, while reducing calculation costs.SOLUTION: A creation method of initial experimental data includes an initial experimental data creation step for creating initial experimental data 34 by extracting data with which combinations have high adaptability from experiment candidate data 32. In the initial experimental data creation step, the initial experimental data 34 is created by repeating: a selection step in which N pieces of data are extracted from data 33 in process, after the experiment candidate data 32 has become the data 33 in process, and M pieces of data (where M<N) with which combinations have high adaptability are selected from the extracted N pieces of data, by using a genetic algorithm; and a data update step for updating the data 33 in process by combining the data selected in the selection step, and data remaining of the data 33 in process that was not extracted in the selection step.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for generating initial experimental data and a method for planning an experiment. [Background technology]

[0002] When manufacturing a composite material formed by combining multiple materials, a method for searching for a material composition that will achieve desired properties is, for example, an adaptive experimental design method using Bayesian optimization. In this method, initial experimental data to be used in an initial experiment is first created using an appropriate method from a huge number of combinations of material compositions, and the initial experiment is then performed using the material compositions from the created initial experimental data. Then, machine learning is performed using the results of the initial experiment to search for a material composition that will achieve the desired properties.

[0003] In the above-mentioned experimental design method, it is important to select the initial experimental data appropriately in order to quickly identify the composition of the material that can obtain the desired properties. Therefore, a method of creating the initial experimental data by selecting a combination of data that has a high adaptability, such as a D-optimal criterion value, has been known (see, for example, Patent Document 1).

[0004] However, when the number of data combinations for all material compositions is enormous, for example, millions to tens of millions, it is not realistic to calculate the fitness for all data combinations. Therefore, a conventional method for generating initial experimental data is a random sampling method in which data combinations that become initial experimental data are randomly selected, fitness is calculated repeatedly, and the data combination with the highest fitness is used as the initial experimental data. Another known method for generating initial experimental data is to use a genetic algorithm to obtain a data combination with the highest fitness from randomly selected data. [Prior art documents] [Patent documents]

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, in the above-mentioned random extraction method or the method using a genetic algorithm, for example, from a huge amount of data such as several million to tens of millions of data, a very small number of data such as several tens of data are randomly selected to obtain the fitness, and the search range becomes very limited. Therefore, in order to select a more appropriate combination of data, it is necessary to perform random selection of data and calculation of fitness many times, resulting in a large computational cost.

[0007] Therefore, an object of the present invention is to provide a method for creating initial experimental data and an experimental design method that can create appropriate initial experimental data while reducing the computational cost.

Means for Solving the Problems

[0008] The present invention aims to solve the above problems, and includes an experimental candidate data creation step for creating experimental candidate data, and an initial experimental data creation step for extracting data with a high fitness combination from the experimental candidate data to create initial experimental data. The initial experimental data creation step includes, after using the experimental candidate data as in-process data, extracting N data from the in-process data, and selecting M data (where M < N) with a high fitness combination from the extracted N data using a genetic algorithm. The selection step, and a data update step of combining the data selected in the selection step and the remaining data of the in-process data not extracted in the selection step to update the in-process data, and repeating the above steps to create the initial experimental data, and provides a method for creating initial experimental data.

[0009] Furthermore, for the purpose of solving the above-mentioned problems, the present invention provides an experimental planning method comprising: a training data creation step of creating training data including the initial experimental data created by the initial experimental data creation method and characteristic values ​​obtained as a result of conducting an experiment using each data item of the initial experimental data; a regression model creation step of creating a regression model using the training data that indicates the correlation between each parameter of the initial experimental data and the characteristic; a characteristic prediction step of predicting the value of each of the characteristic values ​​by applying each data item of the experimental candidate data to the regression model; and a high characteristic value data extraction step of extracting data from the experimental candidate data that will produce a high value of the characteristic based on the prediction result of the characteristic prediction step.

[0010] Furthermore, in order to solve the above-mentioned problems, the present invention provides an experimental design method including a data extraction step of performing Bayesian optimization based on the initial experimental data created by the initial experimental data creation method and characteristic values ​​obtained as a result of conducting an experiment using each data of the initial experimental data, and extracting data that maximizes the value of the acquisition function. [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a method for generating initial experiment data and a method for designing an experiment that can generate appropriate initial experiment data while reducing calculation costs. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a schematic configuration diagram of an experimental planning system according to an embodiment of the present invention. [Figure 2] FIG. 10 is a diagram illustrating an example of experiment candidate data. [Figure 3] FIG. 10 is a diagram illustrating an initial experiment data creation process. [Figure 4] FIG. 10 is a diagram illustrating a genetic algorithm used in the selection process. [Figure 5] 1 is a flowchart of an experiment planning method including a method for creating initial experiment data according to an embodiment of the present invention. [Figure 6]FIG. 10 is a flow diagram of a process for creating initial experimental data. [Figure 7] FIG. 10 is a diagram illustrating a modified example of the initial experiment data creation process. DETAILED DESCRIPTION OF THE INVENTION

[0013] [Embodiment Mode] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings.

[0014] 1 is a schematic diagram of an experimental planning system 1 used in the method for creating initial experimental data and the method for planning experiments according to the present embodiment. The experimental planning system 1 according to the present embodiment is a system for determining the composition of materials to be used in an experiment in order to efficiently search for a material composition that will achieve desired properties when manufacturing a composite material formed by combining multiple materials. However, the present invention is not limited to composite materials and can be applied to other materials and other fields.

[0015] The experimental planning device 1 has a control unit 2 and a storage unit 3. The experimental planning device 1 is, for example, a computer such as a personal computer or a server device, and is equipped with a processing element such as a CPU, memories such as RAM and ROM, a storage device such as a hard disk, and a communication interface which is a communication device such as a LAN card.

[0016] The control unit 2 has a setting processing unit 21, an experiment candidate data creation processing unit 22, an initial experiment data creation processing unit 23, a learning data creation processing unit 24, a regression model creation processing unit 25, a characteristic prediction processing unit 26, and a high characteristic value data extraction processing unit 27. Details of each unit will be described later. The storage unit 3 is realized by a predetermined storage area of ​​a memory or storage device.

[0017] The experimental planning apparatus 1 also has a display 4 and an input device 5. The display 4 is, for example, a liquid crystal display, and the input device 5 is, for example, a keyboard or a mouse. The display 4 may be configured as a touch panel, and may also serve as the input device 5. The display 4 and the input device 5 may also be configured separately from the experimental planning apparatus 1 and configured to be able to communicate with the experimental planning apparatus 1 via wireless communication or the like. In this case, the display 4 or the input device 5 may be configured as a mobile terminal such as a tablet or a smartphone.

[0018] (Setting processing unit 21) The setting processing unit 21 performs setting processing (setting step) for making various settings of the experiment planning apparatus 1. The setting processing unit 21 can set information related to various processes, such as setting the conditions for creating the experiment candidate data 32 in the experiment candidate data creation processing described later, and setting the number (M) of initial experiment data 34 and the number (N) of data to be randomly sampled in the initial experiment data creation processing described later. The setting processing unit 21 can also register, update, delete, etc., various pieces of information stored in the storage unit 3. The input device 5 or the like can be used to input various pieces of information. The information on the various settings set in the setting processing is stored in the storage unit 3 as setting data 31.

[0019] (Experiment candidate data creation processing unit 22) The experiment candidate data creation processing unit 22 performs an experiment candidate data creation process (experiment candidate data creation step) for creating the experiment candidate data 32. In this embodiment, in the experiment candidate data creation step, the experiment candidate data 32 is created by automatically combining composition information of a plurality of materials in accordance with the creation conditions set in the setting process. FIG. 2 is a diagram showing an example of the experiment candidate data 32. As shown in FIG. 2, for example, when seven types of materials A to G are used and the composition of each material is set to five levels, the experiment candidate data creation processing unit 22 combines the compositions of all the materials to create five levels of the experiment candidate data 32. 7The experiment candidate data 32 consisting of 78,125 data items is created. The created experiment candidate data 32 is stored in the storage unit 3. Note that the experiment candidate data 32 may include not only composition information of materials but also manufacturing conditions such as temperature and time.

[0020] (Initial experiment data creation processing unit 23) The initial experiment data creation processing unit 23 performs an initial experiment data creation process (initial experiment data creation step) that extracts data that form combinations with high adaptability from the experiment candidate data 32 created in the experiment candidate data creation process, and creates initial experiment data 34. The initial experiment data creation processing unit 23 has a selection processing unit 23a and a data update processing unit 23b, and is configured to create the initial experiment data 34 by repeating the selection process (selection step) by the selection processing unit 23a and the data update process (data update step) by the data update processing unit 23b.

[0021] In this embodiment, the most commonly used D-optimal reference value is used as the fitness. However, the fitness is not limited to the D-optimal reference value, and for example, the A-optimal reference value or the I-optimal reference value can also be used.

[0022] Fig. 3 is a diagram illustrating the initial experiment data creation process. For simplicity of explanation, Fig. 3 shows a case where 20 pieces of data are selected as initial experiment data 34 from candidate experiment data 32 including 430 pieces of data.

[0023] 3, in the initial experiment data creation process, first, the experiment candidate data 32 (430 data items) created in the experiment candidate data creation process is stored in the storage unit 3 as processing data 33 (430 data items), which is temporary data during processing. In the initial experiment data creation process, a selection process using a genetic algorithm is combined with a tournament method to search for a data combination suitable for the initial experiment data 34 using all data included in the experiment candidate data 32, thereby maintaining a wide search range.

[0024] The selection processing unit 23a extracts N pieces of data from the in - process data 33, and from the N pieces of extracted data, uses a genetic algorithm to perform a selection process of selecting M pieces of data (where X < M < N, and X is the number of explanatory variables) that result in a combination with a high fitness value. In the example of FIG. 3, the case where N = 50 and M = 20 is shown, which shows the case of extracting a combination of 20 pieces of data with a high D - optimal criterion value from 50 randomly extracted pieces of data. Hereinafter, N may be referred to as the sampling number. Also, M may be referred to as the desired number of data. The specific processing content of the genetic algorithm used in the selection process will be described later.

[0025] In the present embodiment, the selection processing unit 23a randomly and non - repetitively extracts N pieces of data (50 pieces in the example of FIG. 3) from the in - process data 33, and from the N pieces of extracted data, uses a genetic algorithm to select M pieces of data (20 pieces in the example of FIG. 3) that result in a combination with a high fitness value, and repeats this until the number of unextracted in - process data 33 becomes less than N (50 pieces in the example of FIG. 3). The M pieces of data selected in the selection process are stored in the storage unit 3 together with the fitness value (here, the D - optimal criterion value). In the example at the top of FIG. 3, since the number of data in the in - process data 33 is 430, the process of randomly extracting 50 pieces of data and selecting 20 pieces of data with a high fitness value is repeated 8 times, and 30 pieces of unextracted data remain.

[0026] [[ID=⑧]]The data update processing unit 23b combines the data selected in the selection process (20 pieces × 8 times = 160 pieces in the example at the top of FIG. 3) and the remaining data of the in - process data 33 that was not extracted in the selection process (30 pieces in the example at the top of FIG. 3) to perform a data update process of updating to new in - process data 33 (160 pieces + 30 pieces = 190 pieces in the example at the top of FIG. 3).

[0027] By repeating the selection process and the data update process, the number of data items in the data being processed 33 gradually decreases, narrowing down to data items that have high fitness when combined with other data items. When the number of data items in the data being processed 33 after the data creation process is N or less (50 in the example of FIG. 3), the initial experimental data creation processing unit 23 stores M data items (20 in the example of FIG. 3) selected in the next selection process together with fitness (here, the D-optimal reference value) in the storage unit 3, and then sets the M data items with the highest fitness (here, the D-optimal reference value) among the combinations of M data items stored in the storage unit 3 as the initial experimental data 34. In the example of FIG. 3, by performing the selection process and the data update process in the top row, 430 data items are narrowed down to 190, and by performing the selection process and the data update process in the next row, the 190 data items are narrowed down to 100. By performing the selection process and the data update process in the next row, the 100 data items are narrowed down to 40, and as a result, the number of data items in the data being processed 33 becomes N (50) or less. Therefore, the initial experimental data creation processing unit 23 stores 20 pieces of data selected from the 40 pieces of data in the next selection step in the memory unit 3, and then stores 20 pieces of data among the 20 combinations of data stored in the memory unit 3 that have the highest fitness (here, the D optimal reference value) in the memory unit 3 as initial experimental data 34. Note that the inventors have found through their investigations that the fitness of the M pieces of data selected in the final selection step may not be the highest. Therefore, in this embodiment, the combination of M pieces of data selected in the selection process is stored each time together with the fitness, and the combination of M pieces of data that has the highest fitness is used as the initial experimental data 34.

[0028] Here, the specific processing contents of the genetic algorithm used in the selection process will be explained using Figure 4. As shown in Figure 4, in the selection process, first, for combinations of N randomly selected pieces of data (50 in this case), initial individuals are created, with 0 or 1 representing whether or not the data will be used. 1 for each individual means that the data will be used, and 0 means that the data will not be used. When creating the initial individuals, 0 or 1 is randomly determined for each piece of data. The number of initial individuals to be created can be set appropriately depending on the numbers M and N (this point will be discussed later). Here, the case is shown where the number of initial individuals is 300. The 300 initial individuals are collectively referred to as the initial population. After that, the D-optimal criterion value is found for each individual in the initial population.

[0029] Then, from the initial population, individuals with a high D-optimality criterion value are selected. In this embodiment, since M (20 in this case) data combinations are selected, constraints are set so that individuals with 20 "1"s are selected preferentially. Various specific methods for selecting individuals are known, such as elite methods and tournament methods, and individuals may be selected using an appropriate known method. The selected individuals are then subjected to crossover and mutation processes to create a next-generation population with an increased number of individuals. Crossover is a process of selecting two parent individuals and swapping parts of their data to form new individuals. Mutation is a process of swapping 0s and 1s in arbitrary data with an extremely low probability. The D-optimality criterion value of each individual in the created population is then calculated again, and individuals with a high D-optimality criterion value (and individuals with 20 "1"s) are selected. The selected individuals are used to create a next-generation population, and this process is repeated for a predetermined number of generations. The combination of individuals with the highest D-optimality criterion value in the final generation is the data combination selected in the selection process.

[0030] Here, we consider the number of data (sampling number) N to be randomly sampled in a genetic algorithm. When the number of data to be selected (desired number of data) M is set to 20, Table 1 shows the probability of the desired number of data appearing when the number of data to be randomly sampled (sampling number) N is changed. The desired number of data appearing probability represents the probability that an individual with M (20) 1s will appear when individuals are randomly created. Table 1 also shows the initial number of individuals and the number of individuals that appear with the desired number of data (M (20) 1s) (number of individuals with desired number of data appearing).

[0031] [Table 1]

[0032] As shown in Table 1, when the sampling number N is 50, the probability of an individual with 20 1s appearing is approximately 4%. In contrast, when the sampling number N is 60, the probability of an individual with 20 1s appearing is less than 1%, and when the sampling number N is 70 or more, the probability is less than 0.1%. As a result, there is a high possibility that the desired number of individuals will not appear in the initial population, resulting in processing failure. This problem can be improved by increasing the number of initial individuals (the number of initial populations). However, the larger the number of individuals in the initial population, the higher the computational cost of calculating the D-optimal criterion value for each individual and the longer the processing time. Therefore, from the perspective of suppressing processing failures and reducing computational costs, it is desirable for the sampling number N to be greater than the desired number of data M and less than 60, and more desirably less than 50.

[0033] (Learning data creation processing unit 24) The learning data creation processing unit 24 performs a learning data creation process (learning data creation step) to create learning data 35 including initial experimental data 34 created in the initial experimental data creation process and characteristic values ​​obtained as a result of conducting experiments using each data item in the initial experimental data 34. In the learning data creation process, the learning data 35 is created by linking the initial experimental data 34 with data on characteristic values ​​corresponding to each data item in the initial experimental data 34. The characteristic values ​​can be input using, for example, the input device 5. The created learning data 35 is stored in the storage unit 3.

[0034] (Regression model creation processing unit 25) The regression model creation processing unit 25 performs a regression model creation process (regression model creation step) to create a regression model 36 that shows the correlation between each parameter of the initial experimental data 34 and the characteristics, using the learning data 35 created in the learning data creation process. More specifically, in the regression model creation process, machine learning is performed using the learning data 35 as training data, the composition of each material as an explanatory variable, and the characteristics as a response variable, to create a regression model 36 that shows the correlation between the explanatory variables and the response variable. The created regression model 36 is stored in the storage unit 3. Note that the explanatory variables may include not only the composition of each material, but also manufacturing conditions such as temperature and time.

[0035] (Characteristics prediction processing unit 26) The characteristic prediction processing unit 26 performs a characteristic prediction process (characteristic prediction step) in which each data item in the experimental candidate data 32 is applied to the regression model 36 created in the regression model creation process to predict the value of each characteristic. In the characteristic prediction process, predicted values ​​of the characteristics are obtained for all data items (combinations of all material compositions) included in the experimental candidate data 32. The obtained predicted values ​​of the characteristics are stored in the storage unit 3 as predicted characteristic data 37.

[0036] (High characteristic value data extraction processing unit 27) The high characteristic value data extraction processing unit 27 performs a high characteristic value data extraction process (high characteristic value data extraction process) to extract data with high characteristic values ​​from the candidate experiment data 32 based on predicted characteristic data 37, which is the prediction result of the characteristic prediction process. In the high characteristic value data extraction process, data with the highest characteristic value (or multiple data with high characteristic values) is extracted. The extracted data is stored in the storage unit 3 as high characteristic value data 38. Furthermore, the high characteristic value data extraction processing unit 27 displays the high characteristic value data 38, which is the extraction result, on the display device 4 or the like to present the extraction result to the user.

[0037] Alternatively, Bayesian optimization may be performed based on the characteristic values ​​obtained as a result of an experiment using each data of the initial experimental data 34, and the data that maximizes the value of the acquisition function may be extracted and the extracted results may be presented to the user. As this acquisition function, for example, PI (Probability of Improvement), PTR (Probability in Target Range), EI (Expected Improvement), and UCB (Upper Confidence Bound) may be used, or a combination of these may be used.

[0038] The user conducts an experiment using a combination of material compositions from the high characteristic value data 38 and measures the property values ​​of the resulting composite material. If the measured property values ​​satisfy the desired properties, the composition of the material from the high characteristic value data 38 can be identified as a composition that achieves the desired properties. If the desired properties are not satisfied, the data obtained from the experiment is added to the training data 35, a new regression model 36 is created, property prediction processing and high characteristic value data extraction processing are performed, and an experiment is conducted using the resulting combination of material compositions, and this process is repeated. In this way, a material composition that achieves the desired properties can be identified.

[0039] (Methods for creating initial experimental data and planning experiments) 5 is a flow diagram of an experiment planning method including a method for creating initial experiment data according to this embodiment. As shown in FIG. 5, first, in step S1, a setting process is performed to set various settings for each process. The various settings performed in the setting process are stored in the storage unit 3 as setting data 31. Then, in step S2, an experiment candidate data creation process is performed. In the experiment candidate data creation process, the experiment candidate data creation processing unit 22 creates data for all combinations of material compositions based on the preset material types and composition levels of each material, and stores the data in the storage unit 3 as experiment candidate data 32.

[0040] Then, in step S3, an initial experiment data creation process is performed. Note that steps S2 and S3 in the flow diagram of FIG. 5 correspond to the flow diagram of the initial experiment data creation method according to this embodiment. In the initial experiment data creation process, as shown in FIG. 6, first, in step S31, in-process data 33 with the same content as the experiment candidate data 32 is created and stored in the storage unit 3. Then, a selection process is performed in steps S32 to S34. First, in step S32, N pieces of data are extracted randomly from the in-process data 33 without overlapping. Then, in step S33, M pieces of data that form a combination with high fitness (here, the D-optimal reference value) are selected from the extracted N pieces of data using a genetic algorithm (see FIG. 4). The selected M pieces of data are stored in the storage unit 3 together with their fitness (here, the D-optimal reference value). Then, in step S34, it is determined whether the number of unextracted pieces of data in-process 33 is less than N. If the determination in step S34 is NO (N), the process returns to step S32 and continues the selection process.

[0041] If the determination in step S34 is YES (Y), a data update process is performed in step S35. In the data update process, the data selected in step S33 is combined with unextracted data from the in-process data 33 to update it into new in-process data 33. Thereafter, in step S36, it is determined whether the number of updated in-process data 33 is N or less. If the determination in step S36 is NO (N), the process returns to step S32, and the selection process and data update process are repeated. If the determination in step S36 is YES (Y), in step S37, a genetic algorithm is used as in step S33 to select M pieces of data that form a combination with high fitness. The selected M pieces of data are stored in the storage unit 3 together with their fitness (here, the D-optimal reference value). Thereafter, in step S38, the M pieces of data with the highest fitness (here, the D-optimal reference value) from the data stored in steps S33 and S37 are stored in the storage unit 3 as initial experimental data 34. Then, the process returns.

[0042] After the initial experiment data creation process in step S3, an initial experiment is performed using the material composition of the initial experiment data 34, and the properties of the obtained composite material are measured. Then, the property values ​​obtained as a result of the measurement are input into the experimental planning system 1 using the input device 5 or the like (step S4).

[0043] Thereafter, in step S5, a learning data creation process is performed. In the learning data creation process, the learning data creation processing unit 24 creates learning data 35 including the initial experimental data 34 and the characteristic values ​​input in step S4, and stores the data in the storage unit 3. Thereafter, in step S6, a regression model creation process is performed. In the regression model creation process, the learning data 35 created in step S5 is used to create a regression model 36 that shows the correlation between the explanatory variables and the dependent variables, with each parameter (composition of each material) of the initial experimental data 34 as an explanatory variable and the characteristic as a dependent variable, and the regression model 36 is stored in the storage unit 3.

[0044] Thereafter, in step S7, a characteristic prediction process is performed. In the characteristic prediction process, each data item of the experiment candidate data 32 created in step S2 is applied to the regression model 36 created in step S6 to predict the respective characteristic values, and these are stored in the storage unit 3 as predicted characteristic data 37. Thereafter, in step S8, a high characteristic value data extraction process is performed. In the high characteristic value data extraction process, data with the highest characteristic value is extracted from the experiment candidate data 32 based on the predicted characteristic data 37, and this is stored in the storage unit 3 as high characteristic value data 38.

[0045] Thereafter, an experiment is conducted using the composition of the material of the high characteristic value data 38, and the characteristics of the obtained composite material are measured. The characteristic values ​​obtained as a result of the measurement are then input into the experimental planning system 1 using the input device 5 or the like (step S9). Thereafter, in step S10, it is determined whether the characteristic values ​​input in step S9 achieve the target. If the determination in step S10 is NO (N), in step S11, the high characteristic value data 38 and the characteristic values ​​input in step S9 are added to the training data 35 to update the training data 35, and then the process returns to step S6. If the determination in step S10 is YES (Y), the process ends.

[0046] (Example) Initial experimental data 34 was obtained using seven types of materials, with five levels of composition for each material. The experimental candidate data 32 was a combination of all the material compositions, with five levels. 7 = 78,125 data were created. The number of data in the initial experiment 34 was 20. The combination of selecting 20 data from 78,125 data is as follows: 78125 C 20 =2.9×10 79 Therefore, it is not realistic to calculate the fitness of all combinations. In an example in which the initial experimental data 34 was generated by the method for generating initial experimental data according to this embodiment, the calculation time was 923 seconds, and the D-optimal reference value of the obtained initial experimental data 34 was 1.1 × 10 11 It was.

[0047] In Comparative Example 1, the same experimental candidate data 32 as in the example was used, and 20 pieces of data were randomly selected to determine the D-optimal reference value, and this process was repeated 1 million times. The data with the highest D-optimal reference value was used as the initial experimental data 34. The calculation time was an extremely long 4634 seconds, and the D-optimal reference value was 4.0 × 10 9 This was a value about two orders of magnitude lower than that of the example.

[0048] In Comparative Example 2, 50 data were randomly extracted from the experimental candidate data 32, and 20 data with a high D-optimal criterion value were selected using a genetic algorithm. The calculation time was short at 1 second, but the search range was very narrow, so the reliability was low, and the D-optimal criterion value was also 3.8 × 10 9 , which was a value about two orders of magnitude lower than that of the embodiment. In Comparative Example 3, 50 data were randomly extracted from the experimental candidate data 32, and 20 data that would result in a combination with a high D-optimal criterion value were selected using a genetic algorithm 930 times so that the calculation time (940 seconds) would be the same as that of the embodiment. In this Comparative Example 3, the D-optimal criterion value was 9.0 × 10 9 The results are shown in Table 2.

[0049] [Table 2]

[0050] From the results in Table 2, it was confirmed that, according to this embodiment, appropriate initial experimental data 34 that provides a high D optimum reference value can be created in a shorter calculation time than in the conventional method.

[0051] (Variation) In this embodiment, in the selection process, random sampling and selection are repeated until the number of unextracted in-process data 33 becomes N or less, but this is not limiting, and the selection process may be terminated when the number of unextracted in-process data 33 is greater than N. For example, as shown in Fig. 7, random sampling and selection may be performed only once, and the selected data may be combined with the unextracted in-process data 33 to repeatedly update the in-process data 33.

[0052] Furthermore, in this embodiment, a case has been described in which the function of creating initial experiment data 34 is incorporated into the experimental planning device 1, but this is not limited to this. It is also possible to configure an initial experiment data creation device that is separate from the experimental planning device 1 and includes an experiment candidate data creation processing unit 22 and an initial experiment data creation processing unit 23.

[0053] Also, in the present embodiment, in step S31, processing data 33 having the same content as the experimental candidate data 32 is created and stored in the storage unit 3. That is, in order to use the experimental candidate data 32 for characteristic prediction processing or the like, the initial experimental data creation process is performed using the processing data 33 different from the experimental candidate data 32. However, as another modification example, for example, without using another processing data 33, after the experimental candidate data 32 itself is used as the processing data 33, the initial experimental data creation process may be performed so as to directly update the experimental candidate data 32 itself which is the processing data 33. That is, the initial experimental data creation process (initial experimental data creation step) extracts N pieces of data from the experimental candidate data 32 which is the processing data 33 after using the experimental candidate data 32 as the processing data 33, and from the extracted N pieces of data, using a genetic algorithm, a selection process (selection step) of selecting M pieces (where M < N) of data that form a combination with a high fitness, and a data update process (data update step) of updating the experimental candidate data 32 by combining the data selected in the selection process (selection step) and the remaining data of the experimental candidate data 32 that was not extracted in the selection process (selection step) may be repeated to create the initial experimental data 34. In the case of this another modification example, it is preferable to create experimental candidate data 32 having the same content as the experimental candidate data 32 created in step S2 and store it in the storage unit 3 as a backup for the experimental candidate data 32 to be directly updated.

[0054] (Actions and Effects of the Embodiment) As described above, in the method for creating initial experimental data according to the present embodiment, an experimental candidate data creation step for creating experimental candidate data 32 and an initial experimental data creation step for extracting data having a high degree of fitness from the experimental candidate data 32 to create initial experimental data 34 are provided. In the initial experimental data creation step, after setting the experimental candidate data 32 as in-processing data 33, N data are extracted from the in-processing data 33, and from the extracted N data, using a genetic algorithm, a selection step of selecting M (where M < N) data that form a combination with a high degree of fitness, and a data update step of combining the data selected in the selection step and the remaining data of the in-processing data 33 that were not extracted in the selection step to update to new in-processing data 33 are repeated to create the initial experimental data 34.

[0055] Thereby, while reducing the calculation cost, it becomes possible to create appropriate initial experimental data 34 using all the data included in the experimental candidate data 32. As a result, in the experimental design method using the initial experimental data 34 created in the present embodiment, it can be expected to specify a combination of the composition of a material that can quickly obtain a composite material having desired characteristics.

[0056] (Summary of the embodiment) Next, the technical idea grasped from the above-described embodiment will be described by referring to the reference numerals and the like in the embodiment. However, each reference numeral and the like in the following description are not limited to the members and the like that specifically show the components in the claims in the embodiment.

[0057] [1] An experimental candidate data creation step for creating experimental candidate data (32), and an initial experimental data creation step for extracting data with a high degree of fitness from the experimental candidate data (32) to create initial experimental data (34). The initial experimental data creation step includes: after setting the experimental candidate data (32) as in-process data (33), extracting N data from the in-process data (33), and using a genetic algorithm to select M data (where M < N) with a high degree of fitness from the extracted N data. A selection step; a data update step of combining the data selected in the selection step and the remaining data of the in-process data (33) not extracted in the selection step to update the in-process data (33), and repeating the above steps to create the initial experimental data (34). A method for creating initial experimental data.

[0058] [2] In the selection step, N data are randomly and non-repeatedly extracted from the in-process data (33), and M data with a high degree of fitness are selected from the extracted N data using a genetic algorithm. This process is repeated until the number of unextracted in-process data (33) is less than N. The method for creating initial experimental data according to [1].

[0059] [3] In the initial experimental data creation step, when the number of data in the in-process data (33) after the data creation step is N or less, after performing the next selection step, among the combinations of M data selected in each selection step, the combination of M data with the maximum degree of fitness is used as the initial experimental data (34). The method for creating initial experimental data according to [1].

[0060] [4] The degree of fitness is the D-optimal criterion value. The method for creating initial experimental data according to [1].

[0061] [5] In the experimental candidate data creation step, the experimental candidate data (32) is created by automatically combining the composition information of multiple materials. The method for creating initial experimental data according to [X].

[0062] [6] An experimental design method comprising: a learning data creation step of creating learning data (35) including the initial experimental data (34) created by the initial experimental data creation method according to any one of [1] to [5] and characteristic values ​​obtained as a result of an experiment using each piece of the initial experimental data (34); a regression model creation step of creating a regression model (36) using the learning data (35) that indicates a correlation between each parameter of the initial experimental data (34) and the characteristic; a characteristic prediction step of predicting values ​​of the characteristic by applying each piece of the experimental candidate data (32) to the regression model (36); and a high characteristic value data extraction step of extracting data from the experimental candidate data (32) that will have a high value for the characteristic based on a prediction result of the characteristic prediction step.

[0063] [7] An experimental design method comprising a data extraction step of performing Bayesian optimization based on the initial experimental data (34) created by the initial experimental data creation method described in any one of [1] to [5] and characteristic values ​​obtained as a result of an experiment using each data of the initial experimental data (34), and extracting data that maximizes the value of the acquisition function.

[0064] (Addendum) Although the embodiments of the present invention have been described above, the invention according to the claims is not limited to the above-described embodiments. It should be noted that not all of the combinations of features described in the embodiments are necessarily essential to the means for solving the problems of the invention. Furthermore, the present invention can be appropriately modified and implemented within the scope of its spirit. [Explanation of symbols]

[0065] 1...Experimental planning equipment 2...Control unit 21...Settings processing unit 22...Experiment candidate data creation processing section 23...Initial experiment data creation processing section 23a...Selection processing unit 23b...Data update processing section 24...Learning data creation processing unit 25...Regression model creation processing section 26...Characteristics prediction processing unit 27...High characteristic value data extraction processing section 3...Storage section 31...Setting data 32...Experiment candidate data 33...Data in processing 34...Initial experimental data 35...Learning data 36...Regression model 37...Predicted characteristic data 38...High characteristic value data

Claims

1. an experiment candidate data creation step of creating experiment candidate data; an initial experiment data creation step of extracting data that will be a combination with high adaptability from the experiment candidate data to create initial experiment data; The initial experimental data creation step includes: After the candidate experiment data is set as data in process, a selection step of extracting N pieces of data from the data being processed, and selecting M pieces of data (M<N) that form a combination with high fitness from the extracted N pieces of data using a genetic algorithm; a data updating step of combining the data selected in the selection step with the remaining data of the data in process that was not extracted in the selection step to update the data in process, thereby creating the initial experimental data. How to generate initial experimental data.

2. In the selection step, N pieces of data are extracted from the data in process at random and without overlapping, and M pieces of data that form a combination with high fitness are selected from the extracted N pieces of data using a genetic algorithm, and this selection is repeated until the number of unextracted data in process becomes less than N. The method for generating initial experimental data according to claim 1 .

3. In the initial experimental data creation step, when the number of data items in the processing data after the data creation step is N or less, after the next selection step is performed, the combination of M data items selected in each selection step and having the maximum fitness is set as the initial experimental data. The method for generating initial experimental data according to claim 1 .

4. The fitness is a D-optimal criterion value. The method for generating initial experimental data according to claim 1 .

5. In the experiment candidate data creation step, composition information of a plurality of materials is automatically combined to create the experiment candidate data. The method for generating initial experimental data according to claim 1 .

6. a learning data creation step of creating learning data including the initial experimental data created by the initial experimental data creation method according to any one of claims 1 to 5 and characteristic values ​​obtained as a result of an experiment using each data of the initial experimental data; a regression model creation step of creating a regression model that indicates a correlation between each parameter of the initial experimental data and the characteristics, using the learning data; a characteristic prediction step of predicting values ​​of the characteristics by applying each data of the candidate experimental data to the regression model; and a high characteristic value data extraction step of extracting data having a high value of the characteristic from the candidate experimental data based on the prediction result of the characteristic prediction step. Experimental design methods.

7. a data extraction step of performing Bayesian optimization based on the initial experimental data created by the initial experimental data creation method according to any one of claims 1 to 5 and characteristic values ​​obtained as a result of an experiment using each data of the initial experimental data, and extracting data that maximizes the value of the acquisition function; Experimental design methods.

Images