Material manufacturing method

Abstract

To more efficiently search for a composition for obtaining a target physical property value.SOLUTION: In a composition search method of a material, a computer executes steps of: causing to learn learning data in which information about a composition of a material is considered as an explanation variable, and a physical property value of the material is considered as an objective variable, to construct a prediction model; calculating a predicted value of a physical property by inputting prediction data for newly searching for the composition to the prediction model; calculating an influence degree of each explanation variable on prediction by using the learning data and the prediction model; calculating a weighted distance for the learning data of the prediction data, by using the influence degree; and displaying the relation between the predicted value and the weighted distance, and outputting the corresponding prediction data as a search candidate.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a compositional exploration method. [Background technology]

[0002] In material design, it is necessary to determine the parameters (composition or composition ratio, as well as constraints such as cost and manufacturing conditions) to obtain the physical properties of the target material. Traditionally, experimenters have often determined these parameters empirically or through trial and error. However, in the case of complex material design with a large number of parameters, it takes a long time and is extremely difficult to obtain the target physical properties.

[0003] In order to improve on this conventional material design, a technology has been proposed in recent years that uses machine learning to find optimal parameters using accumulated data that links the above-mentioned parameters with known physical property values.

[0004] As an example, Patent Document 1 below proposes an optimization method that generates a Bayesian model that searches for a combination of values ​​of multiple parameters that gives the optimal value as a physical property value for a target substance, and performs a search for combinations using the Bayesian model in a search space.

[0005] Furthermore, Non-Patent Document 1 below proposes a sequential search method using a prediction model, in which the next candidate point is determined using the distance between the predicted value and the training data, and the hyperparameters of the model are optimized. According to this method, the prediction method is not limited in the parameter search.

[0006] As another example, Patent Document 2 below proposes that when searching for parameters that will give desired physical properties using a prediction model that predicts physical property values ​​from the design parameters of a metal material, design conditions should be searched for so as to reduce the variability in multiple predicted values ​​based on multiple different learning data sets, and parameters should be searched for including a new region different from past performance data so as to increase the difference between the parameters and parameters in past performance data. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2020-187642 [Patent Document 2] International Publication No. 2010-152993 [Non-patent literature]

[0008] [Non-Patent Document 1] DOI: arxiv-2101.02289 Summary of the Invention [Problem to be solved by the invention]

[0009] However, the invention disclosed in Patent Document 1 uses a Bayesian model and is limited to optimization methods or Gaussian process regression, which means that it is not possible to flexibly use other prediction methods (e.g., gradient boosting, neural networks, etc.) that are expected to have high predictive performance, resulting in a problem of limited prediction methods.

[0010] On the other hand, the technology described in Non-Patent Document 1 does not limit the prediction method used in parameter search. Furthermore, in the case of the technology described in Non-Patent Document 1, the prediction accuracy verified with past parameters is weighted in the term of distance from the training data so that parameters far from past parameters can be searched for while taking into account the accuracy of the prediction model. However, in the case of the technology described in Non-Patent Document 1, weighting is applied uniformly to all parameters, and a uniform search is performed, including parameters that have a small relationship with the target variable. This results in the problem that it takes a long time to arrive at optimal parameters.

[0011] Furthermore, the invention disclosed in Patent Document 2 is configured to apply a weight to each parameter so as to increase the difference with the parameters in past performance data, but the weights are determined by the user and are arbitrary, which creates the problem that the search is not necessarily performed appropriately.

[0012] An object of the present invention is to provide a composition search method for more efficiently searching for a composition for obtaining a target physical property value. [Means for solving the problem]

[0013] The present invention has the following configuration.

[0014] [1] A composition search method for materials, the composition search method being executed by a computer, comprising the steps of: constructing a prediction model by learning training data in which information about the composition of the material is used as an explanatory variable and the physical property values ​​of the material are used as target variables; calculating predicted values ​​of physical properties by inputting prediction data for newly searching for a composition into the prediction model; calculating the influence of each explanatory variable on the prediction using the training data and the prediction model; calculating a weighted distance between the prediction data and the training data using the influence; and displaying the relationship between the predicted value and the weighted distance and outputting the corresponding prediction data as search candidates.

[0015] [2] The composition search method according to [1], characterized in that in the step of calculating the weighted distance, the weighted distance is scaled to have a value between 0 and 1.

[0016] [3] The composition search method according to [1] or [2], characterized in that the prediction data is a combination of information about compositions that is comprehensively created in accordance with constraints such as preset step sizes and composition ratios, and that by repeating steps from calculating predicted values ​​of physical properties to calculating weighted distances, in the step of displaying the relationship between the predicted values ​​and the weighted distances, a plurality of relationships between the calculated predicted values ​​and the weighted distances are displayed.

[0017] [4] The composition search method according to [3], further comprising: a step of grouping the predicted values ​​by the weighted distance; and a step of displaying the relationship between the predicted values ​​and the weighted distance, wherein the prediction data is grouped and output.

[0018] [5] The composition search method described in [4], wherein in the step of displaying the relationship between the predicted value and the weighted distance, the corresponding prediction data is output as search candidates for each group in descending order of the predicted value.

[0019] [6] The composition search method according to [4] or [5], wherein in the grouping step, the weighted distance is divided equally into predetermined values ​​between 0 and 1 to form groups.

[0020] [7] The composition search method according to [4] or [5], wherein in the grouping step, the weighted distance is divided between 0 and 1 to form groups so that the number of predicted values ​​in each group after division is the same.

[0021] [8] In the step of displaying the relationship between the predicted value and the weighted distance, The composition search method according to any one of [3] to [6], wherein the number of prediction data to be output as search candidates is set by the user.

[0022] [9] Furthermore, the predicted value calculated from the prediction data and the weighted distance are used to calculate the acquisition function Acq(X i ), The composition search method according to [4], further comprising a step of outputting the corresponding prediction data as search candidates in descending order of the calculated acquisition function.

[0023]

number

[0024]

[10] A step of conducting an experiment based on information about the composition of the prediction data output as search candidates in the output step, and obtaining physical property values; The composition search method described in [3], further comprising the step of adding information on the obtained physical property values ​​and the corresponding compositions to the training data, wherein in the step of constructing the predictive model, the process from constructing the predictive model using the training data to which data has been added to the step of obtaining the physical property values ​​is repeated until the obtained physical property values ​​reach predetermined target values. [Effects of the Invention]

[0025] According to the present disclosure, compositions for obtaining target physical property values ​​can be more efficiently searched for. [Brief explanation of the drawings]

[0026] [Figure 1] FIG. 1 is a first diagram showing an example of the system configuration of a composition search system. [Figure 2] FIG. 2 is a diagram illustrating an example of the hardware configuration of the learning device and the prediction device. [Figure 3] FIG. 3 is a diagram illustrating an example of learning data and prediction data. [Figure 4] FIG. 4 is a first diagram showing an example of a graph displaying the relationship between the predicted value and the weighted distance. [Figure 5] FIG. 5 is a first flowchart showing the flow of the composition search process. [Figure 6] FIG. 6 is a second diagram showing an example of the system configuration of the composition search system. [Figure 7] FIG. 7 is a second diagram showing an example of a graph displaying the relationship between the predicted value and the weighted distance. [Figure 8] FIG. 8 is a third diagram showing an example of a graph displaying the relationship between the predicted value and the weighted distance. [Figure 9] FIG. 9 is a second flowchart showing the flow of the composition search process. [Figure 10] FIG. 10 is a third diagram showing an example of the system configuration of the composition search system. [Figure 11] FIG. 11 is a third flowchart showing the flow of the composition search process. [Figure 12] FIG. 12 is a diagram showing the number of times search is terminated in the example and the comparative example. DETAILED DESCRIPTION OF THE INVENTION

[0027] Hereinafter, each embodiment will be described with reference to the accompanying drawings. To facilitate understanding of the description, the same components in the drawings will be denoted by the same reference numerals as much as possible, and duplicate descriptions will be omitted.

[0028] [First embodiment] The composition search method according to the first embodiment includes the steps of constructing a prediction model by learning training data in which information about the composition of a material is used as an explanatory variable and the physical property values ​​of the material are used as target variables; calculating predicted values ​​of physical properties by inputting prediction data for searching for a new composition into the prediction model; calculating the influence of each explanatory variable on the prediction using the training data and the prediction model; calculating a weighted distance between the prediction data and the training data using the influence; and displaying the relationship between the predicted value and the weighted distance and outputting the corresponding prediction data as search candidates.

[0029] In this specification, the composition may be the elements that make up an alloy material, or various raw materials that make up an organic material or composite material. In this specification, information about the composition, such as the type, preparation ratio, and feature amount of raw materials, is also referred to as raw material parameters. Hereinafter, the composition search method according to the first embodiment will be described in detail with reference to FIGS. 1 to 5.

[0030] <System configuration of composition search system> First, the system configuration of a composition searching system that realizes a composition searching method according to a first embodiment will be described using Fig. 1 with reference to Figs. 3 and 4. Fig. 1 is a first diagram showing an example of the system configuration of a composition searching system. Fig. 3 is a diagram showing an example of learning data and prediction data. Fig. 4 is a first diagram showing an example of a graph displaying the relationship between a predicted value and a weighted distance.

[0031] As shown in FIG. 1, the composition search system 100 includes a learning device 110 and a prediction device 120.

[0032] A learning program is installed in the learning device 110, and the learning device 110 functions as a learning unit 112 by executing the program.

[0033] The learning unit 112 constructs a prediction model (trained model) using the learning data stored in the learning data storage unit 111. In this embodiment, the learning data used by the learning unit 112 when constructing the prediction model includes sets of raw material parameters (type, preparation ratio, feature amount) and measured physical property values ​​for multiple experimental samples (see FIG. 3(A)).

[0034] In addition, in this embodiment, the model learned by the learning unit 112 includes any method such as a random forest, Gaussian process regression, a neural network, or an ensemble learning model that combines multiple methods.

[0035] The prediction model (trained model) constructed by the learning unit 112 is set in the prediction unit 122 of the prediction device 120.

[0036] A prediction program is installed in the prediction device 120, and by executing this program, the prediction device 120 functions as a prediction data generation unit 121, a prediction unit 122, a display unit 123, an influence calculation unit 124, and a weighted distance calculation unit 125.

[0037] The calculation data generating unit 121 generates calculation data. The calculation data is composed of a comprehensive combination of compositions created in accordance with constraints that specify upper and lower limits and intervals of composition ratios, raw materials that cannot be used simultaneously, etc., or data on feature quantities related to the compositions (see FIG. 3(B)). The calculation data generating unit 121 inputs the generated calculation data to the prediction unit 122 and notifies the weighted distance calculation unit 125 of the data.

[0038] The prediction unit 122 calculates a predicted value from the prediction data using a prediction model, and notifies the display unit 123 of the calculated predicted value.

[0039] The influence calculation unit 124 calculates the influence of each explanatory variable on the prediction using the learning data stored in the learning data storage unit 111 and the prediction model. Specifically, the influence calculation unit 124 calculates the influence using various algorithms stored in various Python libraries.

[0040] For example, if the prediction model is a linear model, the influence calculation unit 124 calculates the influence using the coefficient of each variable. Also, if the prediction model is a model based on a decision tree, the influence calculation unit 124 calculates influence such as permutation importance or Gini importance. Alternatively, the influence calculation unit 124 may calculate the influence using the SAGE or SHAP algorithm of a Python library that can calculate influence using any method.

[0041] The weighted distance calculation unit 125 calculates the weighted distance of the prediction data to the learning data using the influence calculated by the influence calculation unit 124. Specifically, the weighted distance calculation unit 125 calculates the weighted distance using the following formulas (2) and (3).

[0042]

number

[0043]

number

[0044] The display unit 123 displays a plurality of relationships between the predicted values ​​calculated by the prediction unit 122 and the weighted distances calculated by the weighted distance calculation unit 125. For example, the display unit 123 displays a plurality of relationships between the predicted values ​​and the weighted distances using a two-dimensional graph with the horizontal axis representing the weighted distances and the vertical axis representing the predicted values ​​(see FIG. 4). The display unit 123 also outputs the corresponding prediction data as search candidates.

[0045] <Hardware configuration of the learning device and prediction device> Next, we will explain the hardware configurations of the learning device 110 and the prediction device 120 included in the composition search system 100. In this embodiment, the hardware configurations of the learning device 110 and the prediction device 120 are generally the same, so they will be explained together here using Figure 2. Figure 2 is a diagram showing an example of the hardware configurations of the learning device and the prediction device.

[0046] 2, the learning device 110 and the prediction device 120 include a processor 201, a memory 202, an auxiliary storage device 203, an I / F (Interface) device 204, a communication device 205, and a drive device 206. The hardware components of the learning device 110 and the prediction device 120 are connected to each other via a bus 207.

[0047] The processor 201 has various arithmetic devices such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc. The processor 201 reads various programs (for example, a learning program, a prediction program, etc.) into the memory 202 and executes them.

[0048] The memory 202 has a main storage device such as a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The processor 201 and the memory 202 form a so-called computer, and the processor 201 executes various programs read onto the memory 202, causing the computer to realize various functions.

[0049] The auxiliary storage device 203 stores various programs and various data used when the various programs are executed by the processor 201. For example, the learning data storage unit 111 is realized in the auxiliary storage device 203.

[0050] The I / F device 204 is a connection device that connects to an operation device 211, which is an example of a user interface device, and a display device 212. The communication device 205 is a communication device for communicating with an external device (not shown) via a network.

[0051] The drive device 206 is a device for loading a recording medium 213. The recording medium 213 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, a magneto-optical disk, etc. The recording medium 213 may also include semiconductor memories that record information electrically, such as ROMs, flash memories, etc.

[0052] The various programs to be installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 213 in the drive device 206 and reading the various programs recorded on the recording medium 213 by the drive device 206. Alternatively, the various programs to be installed in the auxiliary storage device 203 may be installed by being downloaded from a network via the communication device 205.

[0053] <Flow of composition search process in the composition search system> Next, we will explain the flow of composition search processing in composition search system 100. Figure 5 is a first flowchart showing the flow of composition search processing.

[0054] In step S501, the learning device 110 constructs a prediction model. As described above, in this embodiment, the learning data used by the learning device 110 when constructing the prediction model includes pairs of raw material parameters (type, preparation ratio, feature amount) and measured physical property values ​​for multiple experimental samples (see FIG. 3(A)).

[0055] Furthermore, as described above, the predictive model constructed by the learning device 110 is a trained model obtained by performing machine learning using the training data in which the parameters of the raw materials of the training data are used as explanatory variables and the measured physical property values ​​are used as objective variables.

[0056] In step S502, the prediction device 120 generates prediction data. As described above, the prediction data generated by the prediction device 120 in this embodiment is composed of data on feature quantities related to the compositions or on the composition combinations that are comprehensively generated in accordance with constraints that specify the upper and lower limits and intervals of the composition ratios, raw materials that cannot be used simultaneously, etc. (see FIG. 3(B)).

[0057] In step S503, the prediction device 120 calculates a predicted value from the prediction data using the prediction model constructed in step S501.

[0058] In step S504, the prediction device 120 uses the learning data and the prediction model to calculate the degree of influence of each explanatory variable on the prediction.

[0059] In step S505, the prediction device 120 calculates a weighted distance between the prediction data and the learning data using the influence calculated in step S504.

[0060] In step S506, the prediction device 120 checks whether or not calculation of predicted values ​​and weighted distances has been performed for all of the prediction data. If calculation of predicted values ​​and weighted distances has been performed for all of the prediction data (YES in step S506), the process proceeds to step S507. On the other hand, if there is any prediction data for which calculation of predicted values ​​and weighted distances has not been performed (NO in step S506), the process returns to step S503.

[0061] In step S507, the prediction device 120 displays multiple relationships between the predicted values ​​and the weighted distances, and outputs the corresponding prediction data as search candidates. As described above, when displaying multiple relationships between the predicted values ​​and the weighted distances, the prediction device 120 plots and displays the predicted values ​​on a two-dimensional graph with the horizontal axis representing the weighted distances and the vertical axis representing the predicted values ​​(see FIG. 4).

[0062] <Effects of the composition search method according to the first embodiment> Next, we will explain the effects of the composition search method according to the first embodiment. In the composition search method according to the first embodiment, the user can select search candidates taking into consideration the predicted value and the weighted distance of the prediction data to the learning data.

[0063] In the first place, when predicting physical properties, if there is a large difference in important parameters among information about the composition, there is a high possibility that the actual physical properties will differ significantly. On the other hand, if there is a large difference in important parameters, the reliability of the predicted values ​​predicted by the prediction device 120 will be low.

[0064] Here, the unweighted distance is not suitable for use as an index of the reliability of a predicted value because important parameters are buried in information about the composition and treated uniformly. In other words, the weighted distance used in the first embodiment is more suitable as an index indicating whether the reliability of a predicted value is high or challenging than the unweighted distance. As a result, according to the first embodiment, for example, by selecting a composition with a long weighted distance, the user can obtain challenging search candidates in which important parameters are searched for with a focus.

[0065] As described above, according to the first embodiment, it is possible to select search candidates while striking a balance between the reliability of the predicted values ​​and the level of challenge, thereby making it possible to more efficiently search for compositions to obtain target physical property values.

[0066] [Second embodiment] Next, a composition searching method according to the second embodiment will be described, focusing on the differences from the first embodiment.

[0067] <System configuration of composition search system> First, the system configuration of a composition searching system that realizes a composition searching method according to a second embodiment will be described using Fig. 6 and with reference to Figs. 7 and 8. Fig. 6 is a second diagram showing an example of the system configuration of a composition searching system. Figs. 7 and 8 are second and third diagrams showing an example of a graph displaying the relationship between predicted values ​​and weighted distances.

[0068] The system configuration shown in FIG. 6 differs from the system configuration described with reference to FIG. 1 in the first embodiment in that the prediction device 120 includes a classification unit 601 and the function of the display unit 602 is different from the function of the display unit 123.

[0069] The classification unit 601 groups the predicted values ​​calculated by the prediction unit 122 based on the weighted distance between the prediction data and the learning data. The classification unit 601 also notifies the display unit 602 of the grouping results. Note that the method of grouping by the classification unit 601 is arbitrary, and for example, it may select either a method of equally dividing the weighted distance by a predetermined value between 0 and 1, or a method of dividing so that the number of data in each group after division is the same. The number of groups may be a number set in advance, or may be a number set by the user.

[0070] Furthermore, the classification unit 601 calculates an acquisition function that serves as a criterion for determining whether or not the prediction data is a search candidate, and notifies the display unit 602. Specifically, the classification unit 601 calculates the acquisition function using, for example, the following formula (4).

[0071]

number

[0072] The display unit 602 displays multiple relationships between predicted values ​​and weighted distances, and outputs the corresponding prediction data as search candidates for each group in descending order of the acquisition function. Specifically, the display unit 602 selects prediction data (information on composition) from each group in descending order of the acquisition function and outputs them as search candidates.

[0073] The number of search candidates output from each group can be set appropriately for each group, and can be set by the user taking into consideration the experimental environment. For example, the user may set it so that search candidates are output evenly in each group. Alternatively, the user may set it so that a larger number of search candidates are output from groups with a greater weighted distance. In this case, a search can be performed that places emphasis on compositions with a greater weighted distance to the training data.

[0074] The example in Figure 7 shows how, when displaying multiple relationships between predicted values ​​and weighted distances, predicted values ​​are plotted on a two-dimensional graph with the horizontal axis representing weighted distance and the vertical axis representing predicted values. Then, predicted values ​​with high acquisition functions are numbered and displayed, and the corresponding prediction data are output as search candidates.

[0075] In the above explanation, it has been explained that the classification unit 601 groups the predicted values ​​and then calculates the acquisition function, and the display unit 602 numbers and displays the predicted values ​​with the highest acquisition function for each group, and outputs the corresponding prediction data as search candidates.

[0076] However, the functions of the classification unit 601 and the display unit 602 are not limited to this. For example, the classification unit 601 may calculate the acquisition function without grouping the predicted values, and the display unit 602 may number and display the predicted values ​​with high acquisition functions, and output the corresponding prediction data as search candidates.

[0077] In this case, the classification unit 601 may select prediction data based on an acquisition function calculated using, for example, the following formula (5) or (6), and output the data as search candidates.

[0078]

number

[0079]

number

[0080] According to the classification unit 601, by appropriately setting the weighting coefficient α included in the acquisition function, the user can obtain the predicted value f(X i ) and the weighted distance D i or 1-D i For example, in the case of formula (5), if α is increased, it is possible to emphasize the composition with a long weighting distance to the learning data while giving priority to the composition with a high predicted value f(X i Conversely, in the case of formula (6), if α is made small, it is possible to search for a high predicted value f(X i ) can be explored.

[0081] In the case of the classification unit 601, the display unit 602 selects prediction data in descending order of the acquisition function and outputs them as search candidates (see FIG. 8). Note that the display unit 602 can use either the above formula (5) or formula (6) as the acquisition function, or can use both together. When using both formulas, the number of search candidates to be output for each formula may be set appropriately, taking into account the total number of search candidates to be output.

[0082] <Flow of composition search process in the composition search system> Next, we will explain the flow of the composition search process in the composition search system 100. Figure 9 is a second flowchart showing the flow of the composition search process.

[0083] In FIG. 9, the processes from steps S501 to S506 are the same as those described in the first embodiment with reference to FIG. 5, and therefore will not be described here.

[0084] In the following step S901, the prediction device 120 groups the predicted values ​​according to the weighted distance.

[0085] In step S902, the prediction device 120 displays the relationship between the predicted values ​​and the weighted distance, and outputs the corresponding prediction data as search candidates for each group in descending order of the acquisition function. When displaying the relationship between the predicted values ​​and the weighted distance, the prediction device 120 plots the predicted values ​​on a two-dimensional graph with the weighted distance on the horizontal axis and the predicted value on the vertical axis, as shown in Fig. 7, and then numbering and displaying the predicted values ​​with the highest acquisition function, and outputs the corresponding prediction data as search candidates.

[0086] <Summary> As is clear from the above explanation, the composition search method according to the second embodiment groups the predicted values ​​by weighted distance and displays the relationship between the predicted values ​​and the weighted distance. As a result, the composition search method according to the second embodiment can select prediction data with high predicted values ​​based on the level of challenge of each group and output them as search candidates.

[0087] In addition, the composition search method according to the second embodiment calculates an acquisition function for the prediction data, and outputs the prediction data corresponding to a high predicted value of the calculated acquisition function as a search candidate. This allows the composition search method according to the second embodiment to output search candidates while balancing the reliability of the predicted value with the challenge.

[0088] [Third embodiment] Next, a composition searching method according to the third embodiment will be described, focusing on the differences from the first and second embodiments.

[0089] <System configuration of composition search system> First, the system configuration of a composition searching system that realizes a composition searching method according to the third embodiment will be described with reference to Fig. 10. Fig. 8 is a third diagram showing an example of the system configuration of a composition searching system.

[0090] The system configuration shown in FIG. 10 differs from the system configuration described with reference to FIG. 6 in the second embodiment in that it includes an experimental device 1010.

[0091] The experimental apparatus 1010 is an apparatus used by the experimenter 1011 when evaluating the physical properties of the output search candidate compositions. The experimenter 1011 checks whether the physical property values ​​obtained by evaluating the physical properties using the experimental apparatus 1010 reach target values, and if the target values ​​are reached, ends the composition search. On the other hand, if the target values ​​are not reached, the experimenter 1011 adds a pair of information about the composition of the search candidate for which the experiment was performed and the obtained physical property values ​​to the training data, and stores the data in the training data storage unit 111.

[0092] <Flow of composition search process in the composition search system> Next, we will explain the flow of the composition search process in the composition search system 100. Figure 11 is a third flowchart showing the flow of the composition search process.

[0093] In FIG. 11, the processes from steps S501 to S902 are the same as those described in the second embodiment with reference to FIG. 9, and therefore will not be described here.

[0094] In the next step S1101, the experimenter 1011 uses the experimental equipment 1010 to evaluate the physical properties of the search candidate compositions output in step S902, and obtains the physical property values.

[0095] In step S1102, the experimenter 1011 checks whether the physical property values ​​obtained in step S1101 have reached their target values. If the target values ​​have been reached (YES in step S1102), the composition search ends. On the other hand, if the target values ​​have not been reached (NO in step S1102), the process proceeds to step S1103.

[0096] In step S1103, the experimenter 1011 adds a set of information about the composition of the search candidate experimented in step S1101 and the obtained physical property value to the training data, and then returns to step S501. Using the updated training data, the composition search system 100 repeats steps S501 to S1103 described above until the physical property value reaches the target value in step S1102.

[0097] <Summary> As is clear from the above explanation, in the composition search method of the third embodiment, the physical properties of the search candidate composition are evaluated, and if the physical property value does not reach the target value, a pair of information about the search candidate composition and the obtained physical property value is added to the training data.

[0098] In this way, by configuring search candidates that have a balance between high reliability of predicted values ​​and high challenge, and evaluating their physical properties through experiments, the number of experiments required to reach the target physical property value can be reduced.

[0099] Although the above description does not mention processing when a physical property value reaches a target value, it is assumed that when a physical property value reaches a target value, for example, a material is designed and manufactured based on the corresponding search candidate, thereby enabling the design and manufacturing of a material having the target physical property value. [Example]

[0100] Hereinafter, a specific example of the composition searching method according to the third embodiment of the above-described embodiments will be described.

[0101] In this example, the dataset used for training and prediction data was from a paper by Turab Lookman et al. (https: / / www.nature.com / articles / s41598-018-21936-3#Sec12), which describes the compositions of metal compounds and the feature quantities and physical properties associated with those compositions. This dataset is a data set of elastic moduli for 223 M2AX compound compositions (M: transition metal, A: p-block element, X: nitrogen (N) or carbon (C)), a portion of which is shown in Table 1. Columns 2 through 8 of Table 1 list the p, d, and s orbital radii of each element at the element site (M, A, X). These are used as explanatory variables for the training and prediction data. The Young's modulus in column 9 is used as the objective variable for the training data.

[0102] [Table 1] Using the above dataset, the search for the optimal composition by repeatedly outputting (selecting and proposing) search candidates and experimentally evaluating (measuring) their physical properties was reproduced in Example 1 and Comparative Examples 1 and 2. Specifically, the number of attempts required to find the composition with the highest Young's modulus in the dataset was compared. The fewer attempts, the more efficiently the method can search for the optimal composition.

[0103] Example 1 shows a case where a composition search was performed according to the flowchart of Fig. 11, which is a composition search method according to the third embodiment. Comparative Example 1 shows a case where a composition search was performed without performing the processes of steps S504 and S505 of the flowchart of Fig. 11, in order to compare the effect of weighting. Furthermore, Comparative Example 2 shows a case where a composition search was performed using a composition search method that does not take into account the distance from the training data, and simply outputs the corresponding prediction data as search candidates in descending order of predicted value.

[0104] The procedure of Example 1 will be specifically described below.

[0105] In steps S501 and S502, the learning device 110 extracts, as the initial learning data, combinations of the orbital radius and Young's modulus of each of 24 elements with low Young's modulus from among the 223 compound compositions included in the dataset. The learning device 110 also uses the remaining 199 compound compositions included in the dataset as explanatory variables of the prediction data (the orbital radius of each element). The learning device 110 then constructs a prediction model by performing learning using a random forest regression model from scikit-learn as a prediction modeling method.

[0106] In step S503, the prediction device 120 calculates a predicted value from the prediction data using the prediction model constructed in step S501.

[0107] In step S504, the prediction device 120 calculates the Gini importance built into scikit-learn as the influence.

[0108] In step S505, the prediction device 120 calculates a weighted distance using the influence calculated in step S504. The prediction device 120 repeats steps S503 to S506 to calculate predicted values ​​and weighted distances for all prediction data, and then proceeds to step S901.

[0109] In step S901, the prediction device 120 divides the prediction data into groups based on the weighted distance. Here, the prediction data is divided into three groups by dividing the weighted distance by a fixed value.

[0110] In step S902, the prediction device 120 outputs one composition from each group as a search candidate. Specifically, the prediction device 120 uses the above-mentioned formula (4) as the acquisition function, and calculates s g is set to 0 for all groups, and the corresponding prediction data is output as search candidates in descending order of the acquisition function in each group.

[0111] In step S1101, the experimenter 1011 acquires the Young's modulus corresponding to the outputted search candidate (=prediction data) from the data set, instead of conducting an experiment and measurement on the outputted search candidate.

[0112] In step S1102, the experimenter 1011 checks whether the Young's modulus obtained in step S1101 has reached the target value (the highest value in the data set). If it has, the search is terminated and the number of times the search has ended is obtained. If it has not, the process proceeds to the next step, S1103.

[0113] In step S1103, the experimenter 1011 adds the output information about the composition of the search candidate and the acquired physical property values ​​to the training data to update it, and returns to step S501 of constructing a prediction model. The experimenter 1011 repeated the above steps until the Young's modulus reached the target value in step S1102. That is, by adopting one search candidate from each group, the prediction data for each of the three groups as a whole was reduced by three, and the orbital radius of each element that was in the prediction data and the corresponding Young's modulus were added to the training data.

[0114] Note that the random forest regression model used in Example 1 involves randomness in search, and it is possible that the search candidate with the highest Young's modulus may be found by chance the first time. Therefore, in order to appropriately compare the number of times until the search ends, in Example 1, Comparative Example 1, and Comparative Example 2, the procedure until the target value is reached in step S1102 was repeated 100 times to obtain 100 search end times, and the average value and standard deviation were calculated and compared.

[0115] The differences between the procedure of Comparative Example 1 and that of Example 1 will be specifically described.

[0116] In Comparative Example 1, the process corresponding to step S503 in Example 1 is not performed, and in step S504, the influence w of the explanatory variable in the above formula (2) is calculated. tThe unweighted distance is calculated by setting all of the weighted distances to 1. In step S901, the unweighted distance is used instead of the weighted distance. The other steps are the same as in the first embodiment.

[0117] The differences between the procedure of Comparative Example 2 and that of Example 1 will be specifically described.

[0118] In Comparative Example 2, the processes corresponding to steps S503, S504, S901, and S1101 in Example 1 are not performed, and the three corresponding prediction data obtained in step S502 are output as search candidates in descending order of the prediction value, and then step S1101 is performed. The other procedures are performed in the same way as in Example 1.

[0119] The results are shown in Table 2 and Figure 12. The average number of times search was terminated was 5.2 times for Example 1, 7.7 times for Comparative Example 1, and 26.0 times for Comparative Example 2, with Example 1 showing the lowest number of times. Table 2 also shows the average number of times search was terminated and the standard deviation. Figure 12 plots the average number of times search was terminated for Example 1 and Comparative Examples 1 and 2, with the standard deviation shown as error bars.

[0120] Comparative Example 2 clearly has a larger average search termination count, and therefore can be said to be less efficient than Example 1 and Comparative Example 1. The difference between the results of Example 1 and Comparative Example 1 was tested using the null hypothesis, which assumes that the absence of a difference between the two groups results in nothing. The null hypothesis states that there is no difference in the average values ​​between the two groups. A Student's t-test was performed as a specific statistical method. As a result of the test, the p-value was found to be less than the significance level of 0.01, rejecting the null hypothesis and determining that there is a significant difference at the significance level of 1% between the search termination counts of Example 1 and Comparative Example 1. This confirms that the composition search method according to the third embodiment is a method capable of efficiently searching for compositions.

[0121] [Table 2] This application claims priority based on Japanese Patent Application No. 2021-163338, filed on October 4, 2021, the entire contents of which are incorporated herein by reference. [Industrial Applicability]

[0122] The composition search method of the present invention can be used in material design for alloy materials, organic materials, composite materials, and the like.

Claims

1. A method for exploring the composition of a material, comprising: A step of constructing a prediction model by learning learning data using information about the composition of the material as an explanatory variable and physical property values ​​of the material as a target variable; a step of calculating predicted values ​​of physical properties by inputting prediction data for searching for a new composition into the prediction model; calculating the degree of influence of each explanatory variable on the prediction using the learning data and the prediction model; calculating a weighted distance of the prediction data from the learning data using the influence degree; a step of displaying the relationship between the predicted value and the weighted distance, and outputting the corresponding prediction data as search candidates.

2. 2. The composition search method according to claim 1, wherein in the step of calculating the weighted distance, the weighted distance is scaled to a value between 0 and 1.

3. the prediction data is a combination of information about compositions that is comprehensively created in accordance with preset step widths and constraints on composition ratios, By repeating the steps from calculating the predicted value of the physical property to calculating the weighted distance, 2. The composition search method according to claim 1, wherein in the step of displaying the relationship between the predicted value and the weighted distance, a plurality of relationships between the calculated predicted value and the weighted distance are displayed.

4. The computer further performs a step of grouping the predicted values ​​according to the weighted distances; 4. The composition search method according to claim 3, wherein in the step of displaying the relationship between the predicted value and the weighted distance, the prediction data is output in groups.

5. In the step of displaying the relationship between the predicted value and the weighted distance, 5. The composition search method according to claim 4, wherein the prediction data corresponding to each group are output as search candidates in descending order of the predicted values.

6. 5. The composition search method according to claim 4, wherein in the grouping step, the weighted distance is divided equally into predetermined values ​​between 0 and 1 to perform grouping.

7. 5. The composition search method according to claim 4, wherein in the grouping step, the weighted distance is divided between 0 and 1 to group the predicted values ​​so that the number of predicted values ​​in each group after division is the same.

8. In the step of displaying the relationship between the predicted value and the weighted distance, 4. The composition search method according to claim 3, wherein the number of prediction data to be output as search candidates is set by a user.

9. Furthermore, the predicted value calculated from the prediction data and the weighted distance are used to obtain the acquisition function Acq(X ｉ ) 5. The composition search method according to claim 4, further comprising a step of outputting the corresponding prediction data as search candidates in descending order of the calculated acquisition function. [Equation 1] Here, X ｉ is the i-th prediction data, f(X ｉ ) is scaled to have values ​​between 0 and 1, ｉ The predicted value of s ｇ is the weighting coefficient for the gth group, D ｉ is X ｉ is the weighted distance.

10. a step of performing an experiment based on information about the composition of the prediction data output as search candidates in the output step, and obtaining physical property values; and adding information about the obtained physical property values ​​and the corresponding compositions to the training data, 4. The composition exploration method according to claim 3, wherein, in the step of constructing a prediction model, a process of constructing a prediction model using training data to which data has been added, through a process of obtaining the physical property value, in the step of obtaining the physical property value, are repeated until the obtained physical property value reaches a predetermined target value.

Images