JP7730633B2 - Method, device, and computer program for creating training data for predicting chemical composition - Google Patents

Description

The present invention relates to a method, device, and computer program for creating training data for predicting chemical compositions.

Patent Document 1 describes a technique for machine learning related to predicting vulcanized rubber compositions, in which if the number of learning input data before machine learning is less than a predetermined number, the learning input data is increased by interpolating samples created by regression analysis. Patent Document 2 describes a technique for estimating factor values and characteristic values using machine learning, in which if the learning data contains blank fields, the blank data is ignored and learning is performed using only valid data.

For chemical industry customers, it is important to predict the properties of targeted chemical compositions. Typical target chemical composition properties include thermal conductivity, particle size, particle distribution, shape, filler treatment, Morse's hardness, cure rate, hardness, bond line thickness, and electrical conductivity. For example, a thermally conductive silicone composition is manufactured by blending raw materials such as thermally conductive filler and silicone matrix. Traditionally, the blending amounts of a composition that meet the desired properties were determined through trial and error based on existing compositions. Furthermore, while the relationship between these blending amounts and the properties of a target chemical composition can be calculated to some extent through modeling, the modeling process is extremely complex and time-consuming.

Currently, there is technology that simplifies property prediction based on blend amounts by introducing machine learning into this modeling. However, while there are many candidate raw materials for preparing a chemical composition, there are often only a few raw materials required to create a chemical composition that meets the target properties. When a neural network is trained to learn training data that indicates the relationship between the properties of a chemical composition and its raw materials, the training data will show zero values for raw materials that are not used, resulting in the training data containing many zero values. In this case, the large number of zero values acts as a factor that inhibits the learning of the causal relationship between property values and raw materials. This leads to the problem that an optimal trained model cannot be created, resulting in reduced prediction efficiency.

Japanese Patent Application Laid-Open No. 2020-38495 Japanese Patent Application Laid-Open No. 2003-58582

The present invention aims to provide a method, device, and computer program for creating training data for predicting chemical compositions that solves the above-mentioned problems by reducing the computational complexity of machine learning and improving prediction accuracy.

According to a first aspect of the present invention, a method for creating training data for use in machine learning includes the steps of: aggregating the number of times raw materials are used in a dataset of multiple chemical compositions, the dataset including identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and raw material information for the raw materials that make up each of the multiple chemical compositions; ranking the raw materials based on the aggregated number of times they are used; identifying a predetermined number of raw materials from the ranked raw materials; and creating training data for machine learning by correlating the identified raw materials with the characteristic value information.

According to a second aspect of the present invention, the trained model creation method further includes a step of performing machine learning on the training data created by the first aspect to create a trained model.

According to a third aspect of the present invention, a method for predicting the blending materials of a chemical composition includes inputting desired property values of a desired chemical composition into a trained model created by the second aspect, and outputting blending materials and proportions that satisfy the desired property values.

According to a fourth aspect of the present invention, the method for producing a chemical composition further includes the step of producing a desired chemical composition based on the ingredients and proportions output by the third aspect.

According to a fifth aspect of the present invention, training data is created using the training data creation method for use in machine learning of the first aspect.

According to a fifth aspect of the present invention, a computer program causes a processor to perform the methods described in the first to fourth aspects.

According to a sixth aspect of the present invention, a training data creation device for use in machine learning includes: an aggregation unit that aggregates the number of times raw materials are used in a dataset of multiple chemical compositions, the dataset including identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and raw material information for the raw materials that make up each of the multiple chemical compositions; an aggregation unit that ranks the raw materials based on the aggregated number of times they are used; an identification unit that identifies a predetermined number of raw materials from the ranked raw materials; and a creation unit that correlates the identified raw materials with the characteristic value information to create training data for machine learning.

According to a seventh aspect of the present invention, the trained model creation device further includes a learning unit that performs machine learning on the training data created by the training data creation device of the sixth aspect to create a trained model.

According to an eighth aspect of the present invention, the device for predicting a blend of ingredients for a chemical composition further includes an output unit that inputs desired characteristic values of a desired chemical composition into the trained model created by the trained model creation device of the seventh aspect, and outputs blend ingredients and proportions to satisfy the desired characteristic values.

According to a ninth aspect of the present invention, the chemical composition manufacturing system further includes a manufacturing unit that manufactures a desired chemical composition based on the ingredients and proportions output by the ingredient mixture prediction device of the eighth aspect.

FIG. 1 is a diagram showing the minimum configuration of a teacher data creation device according to an embodiment of the present invention. 1 is a diagram showing the configuration of a chemical composition manufacturing system according to an embodiment of the present invention. FIG. 2 is a diagram showing a data set according to the present embodiment of the present invention. FIG. 2 is a diagram showing a processing flow of a teacher data creation method according to the present embodiment of the present invention. FIG. 1 illustrates ranked raw materials according to an embodiment of the present invention. FIG. 2 is a diagram showing training data according to the embodiment of the present invention. FIG. 1 is a diagram showing a process flow for producing a chemical composition using training data according to an embodiment of the present invention. FIG. 10 is a diagram showing a processing flow of a teacher data creation method according to a first modified example of this embodiment of the present invention. A diagram showing the processing flow of a teacher data creation method related to a second modified example of this embodiment of the present invention. FIG. 10 is a diagram illustrating the configuration of a teacher data creation device according to a third modified example of this embodiment of the present invention. FIG. 10 is a diagram showing a processing flow for determining optimal training data according to a third modified example of the present embodiment of the present invention. FIG. 10 is a diagram illustrating the configuration of a teacher data creation device according to a fourth modified example of this embodiment of the present invention. FIG. 10 is a diagram showing a processing flow for determining optimal training data according to a fourth modified example of the present embodiment of the present invention.

The following describes the teacher data creation device according to this embodiment of the present invention, with reference to Figures 1 to 7.

(Minimum configuration of the teacher data creation device)
FIG. 1 is a diagram showing the minimum configuration of a teacher data creation device 1 according to this embodiment of the present invention. The teacher data creation device 1 includes a control unit 11 and a recording unit 12. The teacher data creation device 1 may be configured using a computer such as a server or a personal computer. The recording unit 12 records a dataset 121 for multiple chemical compositions related to a desired chemical composition, as shown in FIG. 3. The dataset 121 includes identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and raw material information for the raw materials constituting each of the multiple chemical compositions. The control unit 11 includes a counting unit 111, a ranking unit 112, an identifying unit 113, and a creating unit 114. The counting unit 111 counts the number of times each raw material is used in the dataset 121 based on the raw material information. The ranking unit 112 ranks the raw materials based on the counted number of times each raw material is used. The identifying unit 113 identifies a predetermined number of raw materials from the ranked raw materials. The creation unit 114 creates training data for machine learning by correlating the identified raw materials with the characteristic value information.

For example, the chemical composition may be a curable composition or a silicone polymer composition. Alternatively, the curable composition may be a curable silicone composition. Alternatively, the chemical composition may be a non-curable silicone composition.

Note that a program for implementing the functions of the control unit 11 in FIG. 1 may be recorded on a computer-readable recording medium within the teacher data creation device 1, and the program recorded on this recording medium may be loaded into a computer system and executed by a processor, thereby executing the operations of the teacher data creation device 1. Note that the term "computer system" here includes hardware such as an OS and peripheral devices. Furthermore, "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into a computer system. Furthermore, "computer-readable recording medium" also includes devices that retain a program for a certain period of time, such as volatile memory (RAM) within a computer system that acts as a server or client when a program is transmitted over a network such as the Internet or a communication line such as a telephone line. The program may also be a program that implements part of the functions described above. Furthermore, the program may be a so-called differential file (differential program) that can implement the functions described above in combination with a program already stored in the computer system.

The recording unit 12 in FIG. 1 is a portable medium such as a flexible disk, optical magnetic disk, ROM, or CD-ROM, a storage device such as a hard disk built into a computer system, or a recording medium such as RAM. Some or all of the data in the recording unit 12 may be recorded on an external device different from the teacher data creation device 1, and the control unit 11 may read the data from the external device and perform the operation in accordance with the operation of the control unit 11.

(Manufacturing system configuration)
FIG. 2 is a diagram showing the configuration of a chemical composition manufacturing system 5 according to this embodiment of the present invention. The manufacturing system 5 includes a teacher data creation device 1, a learning unit 2, an output unit 3, and a manufacturing unit 4. The learning unit 2 performs machine learning on the teacher data 122 created by the teacher data creation device 1 to create a trained model. A trained model creation device may be configured that includes the teacher data creation device 1 and the learning unit 2. The output unit 3 inputs desired property values of a desired chemical composition into the trained model created by the trained model creation device, and outputs the material formulation and proportions required to satisfy the desired property values. A chemical composition composition prediction device may be configured that includes the teacher data creation device 1, the learning unit 2, and the output unit 3. The manufacturing unit 4 manufactures the desired chemical composition based on the material formulation and proportions output by the material formulation prediction device. The chemical composition manufacturing system 5 may be configured that includes the teacher data creation device 1, the learning unit 2, the output unit 3, and the manufacturing unit 4. The output unit 3 may include a display that allows a user to view the output content. The manufacturing section 4 may be a device that performs the selection of raw materials, heating, kneading (mixing) machines, rolling machines, pressing machines, curing machines, etc. fully automatically, or some of the manufacturing steps may be performed manually.

Figure 3 is a diagram showing a dataset 121 according to this embodiment of the present invention. The dataset 121 includes identification names of multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and raw material information about the raw materials that make up each of the multiple chemical compositions. The raw material information includes raw materials consisting of only one type of substance (single component) and raw materials consisting of two or more types of substances. The raw material information may also include only raw materials consisting of two or more types of substances.

Although the data set 121 shown in FIG. 3 is omitted, the identification names include products 1 to 400, and the raw material information includes raw materials 1 to 52. The identification names, characteristic value information, and raw material information in data set 121 are information collected by a user. Data set 121 may be the raw data set collected by a user, i.e., a raw data set. Alternatively, if there is a raw material that is not used for all of the collected products in the raw data set, data set 121 may be a data set from which the column for that raw material has been removed, i.e., a cleaned data set.

For example, a chemical composition identified as "Product 1" is manufactured from "Raw Material 1" and "Raw Material 3" in a blending ratio (mass ratio) of 26:74 (Raw Material 1:Raw Material 3). "Product 1" is a chemical composition having the following properties: density: 1.1 g/cm ³ , hardness (Type A): 30, and tensile strength (tensile strength at break): 6.4 MPa. A chemical composition identified as "Product 2" is manufactured from "Raw Material 1,""Raw Material 2," and "Raw Material 3" in a blending ratio (mass ratio) of 58:33:9 (Raw Material 1:Raw Material 2:Raw Material 3). "Product 2" is a chemical composition having the following properties: density: 1.1 g/cm ³ , hardness (Type A): 31, and tensile strength: 5.2 MPa. As shown in FIG. 3 , Products 3 to 400 also have corresponding raw material information and corresponding characteristic value information.

A blank cell in dataset 121 (e.g., a cell with row: product 1 and column: raw material 2) indicates that raw material 2 is not used to produce product 1. The characteristic value information may be the product's "elongation (elongation at break)," modulus value (MPa), or oil resistance, and is not particularly limited as long as it is information that indicates the characteristics of the chemical composition.

(Process flow for creating training data)
4 is a diagram showing the processing flow of the teacher data creation method according to this embodiment of the present invention. The processing from recording the data set 121 to creating the teacher data 122 will be described in order with reference to FIGS.

A user of the training data creation device 1 specifies the type of chemical composition desired by the customer. The type of chemical composition desired by the customer may be, for example, a curable composition, a curable silicone composition, a silicone polymer composition, a non-curable silicone composition, or, more specifically, silicone rubber, HCR (high-consistency rubber), LSR (liquid silicone rubber), etc. The user collects information on multiple chemical compositions related to the specified desired type of chemical composition, i.e., product information. Based on the collected product information, the recording unit 12 records a dataset 121 for multiple chemical compositions related to the desired chemical composition. The dataset 121 includes identification names of the multiple chemical compositions, characteristic value information indicating the characteristic values of each of the multiple chemical compositions, and raw material information for the raw materials that make up each of the multiple chemical compositions. If the collected product information contains a raw material that is not used in all of the collected products, the recording unit 12 may clean the collected product information, i.e., record a dataset excluding the column for that raw material.

The aggregation unit 111 aggregates the number of times each raw material is used in the dataset 121 based on the raw material information (step S401). The ranking unit 112 ranks the raw materials based on the aggregated number of times each raw material is used (step S402).

An example of counting the number of times of use and ranking will be described with reference to Figure 5. Figure 5 is a diagram showing ranked raw materials according to this embodiment of the present invention. The ranked raw materials in Figure 5 are generated based on the dataset 121 of Figure 3. Raw material 1 in the top row (rank 1) of Figure 5 is used in 302 products out of products 1 to 400 in dataset 121 of Figure 3, indicating that it has been used the most out of products 1 to 400. The next raw material 2 is used in 120 products out of products 1 to 400, indicating that it has been used the second most after raw material 1. The same applies to ranks 3 to 52.

Next, the identifying unit 113 identifies a predetermined number of raw materials from the ranked raw materials (step S403). For example, the identifying unit 113 calculates a predetermined number n that satisfies the following formula (1).
S is a preset value that satisfies 0<S<1, and _Sn is expressed by the following equation (2).
N _k indicates the number of times the ingredient is used, last indicates the total number of types of ingredients in the dataset 121, and n is a natural number. The identification unit 113 identifies a predetermined number n of top-ranked ingredients from the ranked ingredients.

For example, suppose the user has set S=0.7 in advance. In this case, calculate n that satisfies S _n-1 <0.7≦S _n . Specifically, calculate the value of S n when n=1, 2, 3, ... using the above formula (2), and determine the first value of n that makes S _n equal to or greater than 0.7. If the raw materials ranked as shown in FIG. 5 are used,
The value of is the total number of times used corresponding to ranks 1 to 52 (1977). When n=1,
The value of is 302. In this case, S ₁ = 0.152 (= 302/1977). Next, when n = 2,
The value of Sn is 422 (= 302 + 120). In this case, _S2 = 0.213 (= 422/1977). Continuing the calculation in the same way, _S13 = 0.674 (= 1334/1977) and _S14 = 0.706 (= 1397/1977). In other words, the value of Sn becomes 0.7 or greater for the first time when n = 14. In this case, the top 14 raw materials (1, 2, 3, 8, 4, 12, 13, 20, 24, 35, 29, 19, 42, 27) are identified from the ranked raw materials.

Next, the creation unit 114 creates training data 122 for machine learning by correlating the identified raw materials with the characteristic value information (step S404).

Next, Figure 6 is a diagram showing training data according to this embodiment of the present invention. An example of training data to be created will be described with reference to Figure 6. The training data in Figure 6 is information from dataset 121 in Figure 3, excluding information on ingredients other than the top 14 identified ingredients and their identification names. If there is an ingredient other than the top 14 identified ingredients that the learning unit 2 wants to learn, the ingredient may be selected by the user to remain in the training data.

(Action, effect)
As described above, the teacher data creation method according to this embodiment is a teacher data creation method used in machine learning. The method also includes a step of aggregating the number of times raw materials are used in a dataset of multiple chemical compositions. The dataset includes identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and raw material information for the raw materials constituting each of the multiple chemical compositions. The method also includes a step of ranking the raw materials based on the aggregated number of times they are used. The method also includes a step of identifying a predetermined number of raw materials from the ranked raw materials. The method also includes a step of correlating the identified raw materials with the characteristic value information to create teacher data for machine learning.

This significantly reduces the number of zeros contained in the training data used to learn the causal relationship between characteristic values and raw materials. This makes it possible to create training data for creating optimal trained models, improving the efficiency of predicting chemical composition formulations.

(Processing flow after creating training data)
The process up to the creation of the training data has been described in detail above. Next, with reference to FIG. 7 , the process from training the training unit 2 with the created training data to manufacturing the chemical composition will be described in order.

The teacher data creation device 1 transmits the created teacher data 122 to the learning unit 2. The learning unit 2 performs machine learning on the received teacher data 122 to create a trained model (step S701). Here, the learning unit 2 performs machine learning using the raw material information of the teacher data 122 as the correct label and the characteristic value information as the predictive material.

Next, the user inputs the desired characteristic values of the desired chemical composition into the trained model of the training unit 2 (step S702). The output unit 3 outputs the ingredients and proportions required to satisfy the desired characteristic values (step S703). The output unit 3 may present the output ingredients and proportions to the user via a display or the like provided on the output unit 3.

The output unit 3 transmits the outputted compounded ingredients and their proportions to the manufacturing unit 4. The manufacturing unit 4 manufactures the chemical composition based on the received compounded ingredients and their proportions (step S704). This completes the processing flow from the learning unit 2 learning the created training data 122 to the manufacturing of the chemical composition.

(Modification of this embodiment)
The above provides a detailed explanation of the teacher data creation device 1 according to this embodiment, but the specific aspects of the teacher data creation device 1 are not limited to those described above, and various design changes can be made within the scope of the gist.

(First Modification of the Present Embodiment)
For example, as a first modification of this embodiment, the raw material information of the dataset 121 recorded by the recording unit 12 may be composed of information on raw materials consisting of only one type of substance (single-component information). That is, the dataset 121 includes identification names of multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and single-component information on a single component constituting each of the multiple chemical compositions.

(Processing flow for creating training data in the first modified example)
8 is a diagram showing a processing flow of a teacher data creation method according to a first modified example of this embodiment of the present invention. The processing from recording the data set 121 to creating the teacher data 122 will be described in order.

A user of the training data creation device 1 specifies the type of chemical composition desired by a customer. The user collects information on multiple chemical compositions related to the specified desired type of chemical composition, i.e., product information. The recording unit 12 records a dataset 121 for multiple chemical compositions related to the desired chemical composition based on the collected product information. The dataset 121 includes identification names of the multiple chemical compositions, characteristic value information indicating the characteristic values of each of the multiple chemical compositions, and single-component information for the single components that make up each of the multiple chemical compositions. If the collected product information contains a single component that is not used in all of the collected products, the recording unit 12 may clean the collected product information, i.e., record a dataset excluding the column for that single component.

In the first variant, the single-component information of the dataset 121 may be created by converting the raw material into only one type of substance after the dataset 121 containing raw materials consisting of two or more types of substances is recorded in step 401 of FIG. 4.

The aggregation unit 111 aggregates the number of times each single component is used in the dataset based on the single component information (step S801). The ranking unit 112 ranks the single components based on the aggregated number of times each component is used (step S802). Once the single components have been ranked, a table in Figure 5 is created in which the "Raw Material" is "Single Component."

Next, the identification unit 113 identifies a predetermined number of single components from the ranked single components (step S803). For example, the identification unit 113 may identify a predetermined number n of single components using equations (1) and (2) in this embodiment.

Next, the creation unit 114 creates training data for machine learning by correlating the identified single component with the characteristic value information (step S804). If there is a single component other than the identified single component that the learning unit 2 wants to learn, the user may select that single component to remain in the training data.

(Action, effect)
As described above, the teacher data creation method according to the first modified example of this embodiment is a teacher data creation method used for machine learning. The method also includes a step of aggregating the number of times each single component is used in a dataset of multiple chemical compositions. The dataset includes identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and single component information for each single component constituting each of the multiple chemical compositions. The method also includes a step of ranking the single components based on the aggregated number of times each component is used. The method also includes a step of identifying a predetermined number of single components from the ranked single components. The method also includes a step of correlating the identified single components with the characteristic value information to create teacher data for machine learning.

This significantly reduces the number of zeros contained in the training data used to learn the causal relationships between characteristic values and single components. This makes it possible to create training data for building optimal trained models, improving the efficiency of predicting chemical composition formulations.

(Second Modification of the Present Embodiment)
As a second modification of this embodiment, the raw material information in the dataset 121 recorded by the recording unit 12 may be composed of characteristic information about the characteristic points of the raw materials. That is, the dataset 121 includes identification names of multiple chemical compositions, characteristic value information indicating the characteristic values of each of the multiple chemical compositions, and characteristic information about the characteristic points of the raw materials constituting each of the multiple chemical compositions. Here, the characteristic points of the raw materials include at least one of the vinyl content, molecular weight, viscosity, silicon-bonded hydrogen atom weight, type of organosiloxane unit constituting the raw material, functional groups on the organosiloxane unit constituting the raw material, number or ratio of organosiloxane units constituting the raw material, particle size of the filler, surface area of the filler, chemical composition of the filler, density of the filler, electrical conductivity of the filler, thermal conductivity of the filler, shape and aspect ratio of the filler, glass transition temperature, and type, amount, and bonding position of the functional group.

Here, the organosiloxane units constituting the raw materials will be further explained. When the raw material is an organosilicon compound, and particularly an organopolysiloxane compound, the raw material has a siloxane bond represented by Si—O and is composed of organosiloxane units selected from monofunctional siloxane units (also referred to as M units) represented by R ₃ SiO _1/2 (wherein R is independently a group selected from a monovalent organic group, a hydroxyl group, an alkoxy group, a hydrogen atom, and a halogen atom), difunctional siloxane units (also referred to as D units) represented by R ₂ SiO _2/2 (wherein R is the same group as above), trifunctional siloxane units (also referred to as T units) represented by RSiO _3/2 (wherein R is the same group as above), and tetrafunctional siloxane units (also referred to as Q units) represented by SiO _4/2 . Here, the reactivity (including reaction rate) and properties of the reactants (including physical properties such as hardness of the cured product) of the raw materials, chemical compositions containing the raw materials, and the raw materials or chemical compositions are greatly affected by the type of organosiloxane units constituting the raw materials, the functional groups (corresponding to the above-mentioned R) on the organosiloxane units constituting the raw materials, and the number or ratio of the organosiloxane units constituting the raw materials. Therefore, by creating training data using characteristic information including these characteristic points, it is possible to design chemical compositions with desired properties. The monovalent organic group represented by R may contain heteroatoms such as oxygen, nitrogen, and sulfur atoms. Monovalent organic groups include reactive functional groups such as alkenyl groups, epoxy groups, and (meth)acryloxy groups, as well as non-reactive functional groups such as alkyl groups and aryl groups, and the type is not particularly limited depending on the type of raw material. Of the above characteristic points, the vinyl content, molecular weight, viscosity, silicon-bonded hydrogen atom weight, glass transition temperature, and type and bonding position of the functional groups are, of course, affected by the characteristic points related to the organosiloxane units constituting the raw materials.

The molecular weight, which is a characteristic point, may be the weight average molecular weight or the number average molecular weight, or may be the polydispersity of the polymer. Furthermore, the viscosity, which is a characteristic point, may be the viscosity measured by at least one of a capillary viscometer, a falling ball viscometer, and a rotational viscometer, depending on the chemical composition, or may be the kinematic viscosity.

(Processing flow for creating training data in the second modified example)
9 is a diagram showing a processing flow of a teacher data creation method according to a second modified example of this embodiment of the present invention. The processing from recording the data set 121 to creating the teacher data 122 will be described in order.

A user of the training data creation device 1 specifies the type of chemical composition desired by a customer. The user collects information on multiple chemical compositions related to the specified desired type of chemical composition, i.e., product information. The recording unit 12 records a dataset 121 for multiple chemical compositions related to the desired chemical composition based on the collected product information. The dataset 121 includes identification names of the multiple chemical compositions, characteristic value information indicating the characteristic values of each of the multiple chemical compositions, and characteristic information on the characteristic points of the raw materials that make up each of the multiple chemical compositions. If unrelated characteristic points exist for all of the collected products in the collected product information, the recording unit 12 may clean the collected product information, i.e., record a dataset excluding the column of the characteristic point.

In the second variant, the feature information of the dataset 121 may be created by converting the ingredient information into feature points for each ingredient after the dataset 121 containing ingredient information is recorded in step 401 of FIG. 4.

The aggregation unit 111 aggregates the number of times that feature points appear in the dataset based on the feature information (step S901). The ranking unit 112 ranks the feature points based on the aggregated number of times they appear (step S902). Once the feature points have been ranked, a table is created in which "raw materials" are "feature points" as in Figure 5.

Next, the identification unit 113 identifies a predetermined number of feature points from the ranked feature points (step S903). For example, the identification unit 113 may identify the predetermined number n of feature points using equations (1) and (2) in this embodiment.

Next, the creation unit 114 creates training data for machine learning by correlating the identified feature points with the characteristic value information (step S904). If there are feature points other than the identified feature points that the learning unit 2 wants to learn, the user may select those feature points to remain in the training data.

(Action, effect)
As described above, the teacher data creation method according to the second modified example of this embodiment is a teacher data creation method used for machine learning. The method also includes a step of aggregating the frequency of occurrence of feature points in a dataset of multiple chemical compositions. The dataset includes identification names of the multiple chemical compositions, characteristic value information indicating characteristic values of each of the multiple chemical compositions, and characteristic information about feature points of raw materials constituting each of the multiple chemical compositions. The method also includes a step of ranking the feature points based on the aggregated frequency of occurrence. The method also includes a step of identifying a predetermined number of feature points from the ranked feature points. The method also includes a step of correlating the identified feature points with the characteristic value information to create teacher data for machine learning.

This significantly reduces the number of zeros contained in the training data used to learn the causal relationship between characteristic values and feature points. This makes it possible to create training data for creating an optimal trained model, improving the efficiency of predicting chemical composition formulations.

(Configuration of the third modified example of this embodiment)
10 is a diagram showing the configuration of a teacher data creation device according to a third modified example of this embodiment of the present invention. The teacher data creation device 1 may further include a determination unit 115. The determination unit 115 determines whether the teacher data created by the teacher data creation method of this embodiment, the teacher data creation method of the first modified example, or the teacher data creation method of the second modified example is optimal. Furthermore, if the determination unit 115 determines that the teacher data created by the used method is not optimal, the determination unit 115 creates teacher data by a creation method that has not yet been used.

(Processing flow of the third modified example of this embodiment)
11 is a diagram showing a process flow for determining optimal training data according to a third modification of this embodiment of the present invention. The process flow from creating training data to determining optimal training data will be described below.

The teacher data creation device 1 creates teacher data 122 using the teacher data creation method according to this embodiment. The teacher data creation device 1 transmits the created teacher data 122 to the learning unit 2. The learning unit 2 performs machine learning on the received teacher data 122 to create a trained model (step S1101). Next, the user inputs desired characteristic values of a desired chemical composition into the trained model of the learning unit 2 (step S1102). The output unit 3 outputs the blended ingredients and proportions required to satisfy the desired characteristic values (step S1103). The output unit 3 transmits the output blended ingredients and proportions to the manufacturing unit 4. The manufacturing unit 4 manufactures the chemical composition based on the received blended ingredients and proportions (step S1104). Steps S1101 to S1104 up to this point are similar to steps S701 to S704 in FIG. 7.

Next, the determination unit 115 determines whether the created teacher data is optimal. Specifically, the determination unit 115 measures the characteristic values of the manufactured chemical composition (step S1105). Next, the determination unit 115 determines whether the measured characteristic values match the desired characteristic values (step S1106). If the measured characteristic values match the desired characteristic values (step S1106: Yes), the determination unit 115 determines that the created teacher data is optimal (step S1107). If the measured characteristic values do not match the desired characteristic values (step S1106: No), the determination unit 115 determines that the created teacher data is not optimal, and creates teacher data using a teacher data creation method that has not yet been used, i.e., the first or second modified method (step S1108).

For example, if the next method for creating teacher data to be used is the method according to the first modified example, and it is determined that the teacher data created in the same manner is not optimal (step S1106: No), teacher data is created using a method for creating teacher data that has not yet been used, i.e., the method according to the second modified example (step S1108). This is repeated until a method for creating teacher data that is determined to be optimal is found.

If training data is created for all methods and the optimal training data creation method cannot be found, the training data that is deemed most effective may be selected. This completes the processing flow from creating training data to determining the optimal training data.

In the description of Figure 11, it was first determined whether the training data according to this embodiment was optimal, but it may also be determined first whether the training data according to the first or second variant is optimal. Which training data is determined first may be selected by the user, or may be automatically selected based on the type of desired chemical composition, for example, whether it is millable-type rubber.

(Action, effect)
As described above, the teacher data creation method according to the third modification of this embodiment includes a step of determining whether the teacher data created by the teacher data creation method according to this embodiment, the first modification, or the second modification is optimal. If the teacher data is determined to be not optimal, the method includes a step of creating teacher data by a creation method not previously used.

This allows the optimal training data to be selected from multiple training data created using multiple methods, further improving the efficiency of predicting chemical composition formulations.

(Configuration of the fourth modified example of this embodiment)
FIG. 12 is a diagram showing the configuration of a teacher data creation device according to a fourth modified example of this embodiment of the present invention. The teacher data creation device 1 may further include a determination unit 115. The determination unit 115 determines whether the teacher data created by the teacher data creation method of this embodiment, the teacher data creation method of the first modified example, or the teacher data creation method of the second modified example is optimal. Furthermore, if the determination unit 115 determines that the teacher data created by the previously used method is not optimal, the determination unit 115 creates teacher data by a creation method that has not yet been used. Furthermore, the recording unit 12 includes verification data 123 for verifying the usefulness of the teacher data 122. The verification data 123 is data independent of the dataset 121 and the teacher data 122. The verification data 123 includes characteristic values obtained by measuring a chemical composition previously manufactured, as well as the ingredients and proportions of the chemical composition.

(Processing flow of the fourth modified example of this embodiment)
13 is a diagram showing a processing flow for determining optimal training data according to a fourth modification of the present embodiment of the present invention. The processing flow from creating training data to determining optimal training data will be described below.

The teacher data creation device 1 creates teacher data 122 using the teacher data creation method according to this embodiment. The teacher data creation device 1 transmits the created teacher data 122 to the learning unit 2. The learning unit 2 performs machine learning on the received teacher data 122 to create a trained model (step S1301). Next, the manufacturing system 5 inputs the characteristic values in the verification data 123 into the trained model of the learning unit 2 as the desired characteristic values of the desired chemical composition (step S1302). The output unit 3 outputs the blended materials and proportions required to satisfy the desired characteristic values (step S1303).

Next, the determination unit 115 determines whether the outputted ingredients and proportions match the ingredients and proportions corresponding to the input characteristic values in the verification data 123 (step S1304). If the outputted ingredients and proportions match the ingredients and proportions in the verification data 123 (step S1304: Yes), the determination unit 115 determines that the created training data is optimal (step S1305). If the outputted ingredients and proportions do not match the ingredients and proportions in the verification data 123 (step S1304: No), the determination unit 115 determines that the created training data is not optimal, and creates training data using a training data creation method that has not yet been used, i.e., the first or second modified method (step S1306).

For example, if the next method for creating teacher data to be used is the method according to the first modified example, and it is determined that the teacher data created in the same manner is not optimal (step S1304: No), teacher data is created using a method for creating teacher data that has not yet been used, i.e., the method according to the second modified example (step S1306). This is repeated until a method for creating teacher data that is determined to be optimal is found.

If training data is created for all methods and the optimal training data creation method cannot be found, the training data that is deemed most effective may be selected. This completes the processing flow from creating training data to determining the optimal training data.

In step S1304 of FIG. 13, the determination unit 115 determined whether the outputted ingredients and proportions match the ingredients and proportions in the verification data 123. Here, the determination unit 115 may determine whether there is a match by evaluating the degree of match between the ingredients and proportions in the outputted ingredients and proportions and the ingredients and proportions in the verification data 123. For example, the determination unit 115 may numerically evaluate the degree of deviation between the ingredients and proportions in the outputted ingredients and proportions and the ingredients and proportions in the verification data 123, and determine that there is a match if the evaluation value is less than a predetermined threshold.

In the explanation of Figure 13, it was first determined whether the training data according to this embodiment was optimal, but it may also be determined first whether the training data according to the first or second variant is optimal. Which training data is determined first may be selected by the user, or may be automatically selected based on the type of desired chemical composition, for example, whether it is millable-type rubber.

(Action, effect)
As described above, the teacher data creation method according to the fourth modification of this embodiment includes a step of determining whether the teacher data created by the teacher data creation method according to this embodiment, the first modification, or the second modification is optimal. The determining step includes a step of machine learning the created teacher data in a learning unit of a computer, and a step of inputting characteristic values in the verification data to the learning unit and outputting the ingredients and their proportions. The determining step also includes a step of determining whether the output ingredients and their proportions match the ingredients and their proportions in the verification data corresponding to the input characteristic values. The determining step also includes a step of determining that the created teacher data is optimal if they match, and a step of determining that the created teacher data is not optimal if they do not match.

This allows the optimal training data to be selected from multiple training data created using multiple methods without manufacturing, further improving the efficiency of predicting chemical composition formulations.

As described above, this embodiment has been described, but this embodiment and its modifications are presented as examples and are not intended to limit the scope of the invention. This embodiment and its modifications can be implemented in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. This embodiment and its modifications are included in the scope of the invention and its equivalents as set forth in the claims, as well as within the scope and spirit of the invention.

REFERENCE SIGNS LIST 1 Teacher data creation device 11 Control unit 111 Counting unit 112 Ranking unit 113 Identification unit 114 Creation unit 115 Determination unit 12 Recording unit 121 Data set 122 Teacher data 123 Verification data 2 Learning unit 3 Output unit 4 Manufacturing unit 5 Manufacturing system

Claims (26)

A method for creating training data for use in machine learning by a computer, comprising:
a step of aggregating the number of times raw materials are used in a dataset of a plurality of chemical compositions, the dataset including identification names of the plurality of chemical compositions, characteristic value information indicating characteristic values of each of the plurality of chemical compositions, and raw material information on raw materials constituting each of the plurality of chemical compositions;
ranking the raw materials based on the counted number of uses;
identifying a predetermined number of ingredients from the ranked ingredients;
A teacher data creation method including a step of creating teacher data for machine learning by correlating the identified raw materials with the characteristic value information. The step of identifying a predetermined number of raw materials from the ranked raw materials includes:
A step of calculating a predetermined number n that satisfies the following formula (1):
S is a preset value that satisfies 0<S<1,
S _n is expressed by the following formula (2):
N _k indicates the number of times the raw material is used,
last indicates the total number of raw material types in the dataset, and n is a natural number;
The training data creation method according to claim 1 , further comprising the step of identifying a predetermined number n of top-ranked raw materials from the ranked raw materials. A method for creating training data for use in machine learning by a computer, comprising:
a step of aggregating the number of times a single component is used in a dataset of a plurality of chemical compositions, the dataset including identification names of the plurality of chemical compositions, characteristic value information indicating characteristic values of each of the plurality of chemical compositions, and single component information for each of the single components constituting each of the plurality of chemical compositions;
ranking the single components based on the aggregated number of uses;
identifying a predetermined number of single components from the ranked single components;
A teacher data creation method including a step of creating teacher data for machine learning by correlating the identified single component with the characteristic value information. A method for creating training data for use in machine learning by a computer, comprising:
a step of counting the number of occurrences of feature points in a dataset of a plurality of chemical compositions, the dataset including identification names of the plurality of chemical compositions, feature value information indicating feature values of each of the plurality of chemical compositions, and feature information on feature points of raw materials constituting each of the plurality of chemical compositions;
ranking the feature points based on the counted number of occurrences;
identifying a predetermined number of feature points from the ranked feature points;
A teacher data creation method including a step of creating teacher data for machine learning by correlating the identified feature points with the characteristic value information. The training data creation method of claim 4, wherein the characteristic points include at least one of the following: vinyl content, molecular weight, viscosity, weight of silicon-bonded hydrogen atoms, type of organosiloxane units constituting the raw material, functional groups on the organosiloxane units constituting the raw material, number or ratio of organosiloxane units constituting the raw material, particle size of the filler, surface area of the filler, chemical composition of the filler, density of the filler, electrical conductivity of the filler, thermal conductivity of the filler, shape and aspect ratio of the filler, glass transition temperature, and type, amount, and bonding position of the functional groups. The explanatory variables of the teacher data are characteristic value information,
The objective variable of the training data is a raw material, a single component, or a feature point,
a step of having a learning unit perform machine learning on the created teacher data;
inputting desired characteristic values of a desired chemical composition into the learning unit, and outputting ingredients and proportions to satisfy the desired characteristic values;
causing the device to produce a chemical composition based on the outputted formulation ingredients and proportions;
measuring a characteristic value of the produced chemical composition using an apparatus and determining whether the measured characteristic value matches the desired characteristic value;
If there is a match, determining that the created training data is optimal;
The teacher data creation method according to any one of claims 1 to 5, further comprising a step of determining that the created teacher data is not optimal if there is no match. a step of creating teacher data by the teacher data creation method according to claim 1, 3, or 4, in which explanatory variables are characteristic value information and objective variables are raw materials, single components, or feature points;
a step of causing a learning unit to perform machine learning on the created teacher data;
inputting desired characteristic values of a desired chemical composition into the learning unit, and outputting ingredients and proportions to satisfy the desired characteristic values;
causing the device to produce a chemical composition based on the outputted formulation ingredients and proportions;
measuring a characteristic value of the produced chemical composition using an apparatus and determining whether the measured characteristic value matches the desired characteristic value;
If there is a match, determining that the created training data is optimal;
If they do not match, determining that the created training data is not optimal;
A teacher data creating method comprising the step of creating teacher data by a teacher data creating method of claim 1, 3 or 4 that has not been used before, if it is determined that the data is not optimal. The explanatory variables of the teacher data are characteristic value information,
The objective variable of the training data is a raw material, a single component, or a feature point,
a step of causing a learning unit to perform machine learning on the created teacher data;
inputting the characteristic values in the verification data to the learning unit and outputting the ingredients and their proportions;
a step of determining whether the outputted ingredients and proportions match the ingredients and proportions in the verification data corresponding to the input characteristic values;
If there is a match, determining that the created training data is optimal;
If there is no match, determining that the created teacher data is not optimal;
The teacher data creation method according to any one of claims 1 to 5, further comprising: a step of creating teacher data by the teacher data creation method of claim 1, 3, or 4, in which explanatory variables are characteristic value information and objective variables are raw materials, single components, or feature points;
a step of causing a learning unit to perform machine learning on the created teacher data;
inputting the characteristic values in the verification data to the learning unit and outputting the ingredients and their proportions;
a step of determining whether the outputted ingredients and proportions match the ingredients and proportions in the verification data corresponding to the input characteristic values;
If there is a match, determining that the created training data is optimal;
If they do not match, determining that the created training data is not optimal;
A teacher data creating method comprising the step of creating teacher data by a teacher data creating method of claim 1, 3 or 4 that has not been used before, if it is determined that the data is not optimal. The training data creation method according to any one of claims 1 to 9, wherein the chemical composition is a curable composition or a silicone polymer composition. The training data creation method described in claim 10, wherein the curable composition is a curable silicone composition. The training data creation method described in any one of claims 1 to 9, wherein the chemical composition is a non-curable silicone composition. A step of creating teacher data by the teacher data creation method according to any one of claims 1 to 12, wherein an explanatory variable is characteristic value information and a target variable is a raw material, a single component, or a feature point;
A trained model creation method comprising a step of performing machine learning on the training data by a computer to create a trained model. A step of creating a trained model by the trained model creation method according to claim 13;
A method for predicting the formulation of chemical compositions, comprising the steps of inputting desired property values of a desired chemical composition into the trained model by a computer, and outputting formulation materials and proportions that will satisfy the desired property values. A teacher data creation device used in machine learning,
a counting unit that counts the number of times raw materials are used in a data set of a plurality of chemical compositions, the data set including identification names of the plurality of chemical compositions, characteristic value information indicating characteristic values of each of the plurality of chemical compositions, and raw material information about raw materials that constitute each of the plurality of chemical compositions;
a ranking unit that ranks the raw materials based on the counted number of times of use;
an identification unit that identifies a predetermined number of raw materials from the ranked raw materials;
A teacher data creation device including a creation unit that creates teacher data for machine learning by correlating the identified raw materials with the characteristic value information. The identification unit
A predetermined number n that satisfies the following formula (1) is calculated,
S is a preset value that satisfies 0<S<1, and _Sn is expressed by the following formula (2):
N _k indicates the number of times the raw material is used, last indicates the total number of types of raw materials in the dataset, and n is a natural number.
Identifying a predetermined number n of top-ranked raw materials from the ranked raw materials;
The teacher data creation device according to claim 15 . A teacher data creation device used in machine learning,
a counting unit that counts the number of times a single component is used in a data set of a plurality of chemical compositions, the data set including identification names of the plurality of chemical compositions, characteristic value information indicating characteristic values of each of the plurality of chemical compositions, and single component information about a single component that constitutes each of the plurality of chemical compositions;
a ranking unit that ranks the single components based on the counted number of times of use;
an identifying unit that identifies a predetermined number of single components from the ranked single components;
A teacher data creation device including a creation unit that creates teacher data for machine learning by correlating the identified single component with the characteristic value information. A teacher data creation device used in machine learning,
a counting unit that counts the number of occurrences of feature points in a dataset of a plurality of chemical compositions, the dataset including identification names of the plurality of chemical compositions, feature value information indicating feature values of each of the plurality of chemical compositions, and feature information on feature points of raw materials that constitute each of the plurality of chemical compositions;
a ranking unit that ranks the feature points based on the counted number of occurrences;
an identifying unit that identifies a predetermined number of feature points from the ranked feature points;
a creation unit that creates teacher data for machine learning by correlating the identified feature points with the characteristic value information. The training data creation device of claim 18, wherein the characteristic points include at least one of vinyl content, molecular weight, viscosity, weight of silicon-bonded hydrogen atoms, type of organosiloxane units constituting the raw material, functional groups on the organosiloxane units constituting the raw material, number or ratio of organosiloxane units constituting the raw material, particle size of the filler, surface area of the filler, chemical composition of the filler, density of the filler, electrical conductivity of the filler, thermal conductivity of the filler, shape and aspect ratio of the filler, glass transition temperature, and type, amount and bonding position of the functional groups. The explanatory variables of the teacher data are characteristic value information,
The objective variable of the training data is a raw material, a single component, or a feature point,
A determination unit,
The created teacher data is subjected to machine learning by a learning unit of a computer,
inputting desired characteristic values of a desired chemical composition into the learning unit, and outputting blending materials and proportions to satisfy the desired characteristic values;
causing the device to produce a chemical composition based on the outputted ingredients and proportions;
causing an apparatus to measure a characteristic value of the produced chemical composition and determining whether the measured characteristic value matches the desired characteristic value;
If they match, it is determined that the created training data is optimal;
The teacher data creation device according to claim 15 , further comprising a determination unit configured to determine that the created teacher data is not optimal if there is no match. The explanatory variables of the teacher data are characteristic value information,
The objective variable of the training data is a raw material, a single component, or a feature point,
A determination unit,
The created teacher data is subjected to machine learning by a learning unit of a computer,
Inputting the characteristic values in the verification data into the learning unit and outputting the blended materials and their proportions;
determining whether the outputted ingredients and proportions match the ingredients and proportions in the verification data corresponding to the input characteristic values;
If they match, it is determined that the created training data is optimal;
The teacher data creation device according to any one of claims 15 to 19 , further comprising a determination unit configured to determine that the created teacher data is not optimal if there is no match. The training data creation device according to any one of claims 15 to 21 , wherein the chemical composition is a curable composition or a silicone polymer composition. The training data creation device according to claim 22 , wherein the curable composition is a curable silicone composition. The training data creation device according to any one of claims 15 to 21 , wherein the chemical composition is a non-curable silicone composition. The teacher data creation device according to any one of claims 15 to 24 ,
a learning unit that performs machine learning on training data created by the training data creation device, the training data including explanatory variables that are characteristic value information and objective variables that are raw materials, single components, or feature points, to create a trained model;
A trained model creation device including: The trained model creation device according to claim 25 ;
A chemical composition composition material prediction device that includes an output unit that inputs desired characteristic values of a desired chemical composition into a trained model created by the trained model creation device and outputs the blended materials and proportions to satisfy the desired characteristic values.