JP7806629B2 - Apparatus, program, and method for predicting physical properties of composite materials - Google Patents

Description

The present invention relates to a device, a program, and a method for predicting the physical properties of composite materials.

In recent years, a materials design system has been proposed for designing composite materials using the results of machine learning (see, for example, Patent Document 1).

The materials design system described in Patent Document 1 includes a data file containing a dataset of design conditions and material property values for the material to be designed, and a model learning unit that learns a model for estimating material property values from the design conditions based on the dataset to create a trained model, and verifies the predictive accuracy of the trained model through cross-validation.

Japanese Patent Application Laid-Open No. 2020-77257

In recent years, composite materials have been developed that combine multiple types of resins and compounding agents to impart new properties to the properties of the resin itself. Developing a new composite material requires adjusting the composition ratio of each component until the composite material has the desired properties. Because developing composite materials is therefore extremely costly and time-consuming, it is desirable to be able to predict, to some extent, the physical properties of the composite material to be tested at the experimental planning stage in order to improve the efficiency of composite material development.

Furthermore, in the case of composite materials such as wire coating materials, there are many types of compounding agents that can be added to the resin, and the physical properties can vary significantly depending on the composition ratio of the materials. This makes it difficult to predict the physical properties of composite materials, and there is a demand for technology that can predict these physical properties with high accuracy.

When creating and evaluating a regression model using a technology that predicts physical properties using a regression model generated by machine learning on a dataset consisting of multiple data sets including materials and physical properties, a hold-out method is generally used if the dataset contains a large number of data points (e.g., 100,000 or more), in which the training data and test data generation process is carried out only once to create and evaluate the regression model. On the other hand, if the dataset contains a small number of data points (e.g., less than 100), a leave-one-out method is used, in which one data point is used as test data and the remaining data as training data, and this process is carried out for each data point to create and evaluate a regression model for each. However, these methods are not appropriate when the number of data points in the dataset is somewhere between the two, or when the dataset is sparse.

The object of the present invention is to provide a property value prediction device, property value prediction program, and property value prediction method for composite materials that can make highly accurate predictions of property values corresponding to required properties.

In order to solve the above-mentioned problems, the present invention provides a data generation unit that divides a data set, which includes material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of physical properties related to the composite material, for each of a plurality of types of composite materials a plurality of times so that the combinations of the data sets are different from each other, thereby generating a plurality of training data sets and a plurality of evaluation data sets;
a first regression model creation unit that creates a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a regression method determination unit that evaluates the prediction accuracy of the plurality of first regression models for each of the regression methods based on the plurality of evaluation data groups, and determines the regression method with the highest evaluation among the plurality of regression methods;
a second regression model creation unit that creates a second regression model based on the plurality of data sets included in the data set using the regression method with the highest evaluation;
and a physical property prediction unit that predicts physical property values based on input material data using the second regression model.

The present invention makes it possible to predict with high accuracy the physical property values corresponding to the required physical properties.

1 is a block diagram showing a schematic configuration example of a physical property prediction device according to an embodiment of the present invention; FIG. 10 is a diagram illustrating an example of a data set. 1A is a conceptual diagram for explaining an example of data generation processing by a data generation unit, and FIG. 1B is a conceptual diagram for explaining data generation processing by a leave-one-out method as a comparative example. 10A and 10B are diagrams for explaining an example of a first regression model creation process by a first regression model creation unit. 10 is a diagram illustrating an example of a prediction accuracy calculation process performed by an error calculation unit. FIG. 10A and 10B are diagrams for explaining an example of a second regression model creation process by a second regression model creation unit. FIG. 2 is a block diagram for explaining an example of the overall flow of a physical property value prediction device. 10 is a flowchart illustrating an example of the overall operation of the physical property value prediction device.

[Embodiment]
1 is a block diagram showing a schematic configuration example of a physical property prediction device according to an embodiment of the present invention. This physical property prediction device 1 is composed of a computer or the like, and includes a prediction processing unit 2, a storage unit 3, an input unit 4, and a display unit 5, and has a physical property prediction function of predicting physical property values of composite materials using a regression model.

The composite materials targeted by the physical property prediction device 1 are manufactured by combining the materials that make up the composite material. Examples of composite materials include polymer compositions manufactured by combining multiple resins and multiple compounding agents. Examples of polymer compositions include wire coating materials, sheets, tubes, bonded magnets, and magnet rolls.

In this specification, physical properties refer to the properties required of a composite material, and may include, for example, mechanical properties such as tensile elongation at break and tensile strength, and thermal properties such as heat resistance and thermal expansion coefficient.

(Explanation of the physical property prediction function)
The physical property prediction function of the physical property prediction device 1 selects, from a plurality of predetermined regression methods, the regression method with the highest prediction accuracy suitable for the predetermined physical properties of the composite material, since the prediction accuracy of the regression method of the regression model used for prediction varies depending on the physical properties, creates a regression model corresponding to the selected regression method by learning from a data set, and predicts the physical property values of a new composite material using the learned regression model.

In addition, in cases where the number of data in a dataset is insufficient, in order to enable highly accurate prediction of physical property values, the dataset is divided into a training data group and an evaluation data group multiple times to create multiple regression models for each regression method, the average prediction accuracy of each regression model is used as the prediction accuracy of the regression method, and the regression method with the highest prediction accuracy is selected. The division of the dataset will be described later.

The prediction processing unit 2 is composed of a CPU (Central Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), an interface, etc. The prediction processing unit 2 functions as a data generation unit 21, a first regression model creation unit 22, a regression method determination unit 23, a second regression model creation unit 24, and a physical property prediction unit 25 by reading a physical property prediction program 31 stored in the storage unit 3 into the RAM and executing it with the CPU. The regression method determination unit 23 has an error calculation unit 231 and a regression method output unit 232. Details of each unit 21 to 25 of the prediction processing unit 2 will be described later. Note that each unit 21 to 25 of the prediction processing unit 2 may be composed of hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit).

The storage unit 3 is composed of ROM, RAM, a hard disk, etc., and stores various information such as a physical property prediction program 31 for executing the physical property prediction method according to this embodiment, a dataset 32 (see Figure 2), a group of regression methods 33, a training dataset 34, and an evaluation dataset 35.

The input unit 4 is composed of, for example, a keyboard, a mouse, etc. The display unit 5 is composed of, for example, a liquid crystal display, etc. Note that a touch display that integrates the input unit 4 and display unit 5 may also be used. Furthermore, a terminal device having the input unit 4 and display unit 5 may be connected to the prediction processing unit 2 via a network.

Dataset 32 contains data for multiple types of polymer compositions that are used to create and evaluate regression models. Furthermore, dataset 32 may, for example, not contain enough data and may be composed of sparse data. Details of dataset 32 will be described later.

The regression method group 33 includes a predetermined plurality (e.g., m) of regression methods RM ₁ to RM _m , each associated with identification information of the regression method (e.g., regression method name, etc.) and stored in the storage unit 3. Examples of regression methods included in the regression method group 33 include a Gaussian process regression method, a linear regression method, a hierarchical Bayes method, and a regression method using a neural network model. A regression method is an algorithm that uses machine learning to learn parameters of a regression model so that a target variable can be predicted from an explanatory variable.

The training data set 34 includes a plurality of (for example, n) training data groups LDG ₁ to LDG _n (see FIG. 3) each consisting of a plurality of training data LD generated by the data generating unit 21 .

The evaluation data set 35 includes a plurality of (for example, n) evaluation data groups TDG ₁ to TDG _n (see FIG. 3) made up of a plurality of evaluation data TD generated by the data generating unit 21 .

(Dataset composition)
FIG. 2 is a diagram showing an example of the dataset 32. The dataset 32 has a data set 321 for each of a plurality of types of polymer compositions. Specifically, the dataset 32 includes a composition ID 320 that identifies the polymer composition, and a data set 321 that is provided corresponding to the composition ID 320 and includes material data 321a that indicates blending information of materials that constitute the polymer composition and physical property data 321b that indicates physical property values of the polymer composition. While FIG. 2 shows an example of nine data sets 321, the number of data sets 321 (= the number of data) is not limited to nine.

Figure 2 shows an example of a polymer composition, such as a wire coating material. The wire coating material includes multiple resins, such as "Polymer 1," "Polymer 2," "Polymer 3," "Polymer 4," and "Polymer 5," and multiple compounding ingredients, such as "Filler 1," "Filler 2," "Filler 3," and "Crosslinking Agent." Physical property data 321b indicates, for example, the physical property value of "elongation" (e.g., tensile elongation at break) as an example of a predetermined physical property among the multiple physical properties possessed by the wire coating material. When predicting multiple physical property values, physical property data 321b may contain multiple physical property values, and a data set 32 may be prepared for each physical property value. For example, when predicting tensile elongation at break and tensile strength, a regression method suitable for tensile elongation at break and a regression method suitable for tensile strength are selected from the regression method group 33, respectively.

The blending ratio of "Polymer 1" to "Polymer 5" is, for example, 15 to 45 parts by mass per 100 parts by mass of polymer. The blending ratio of "Filler 1" to "Filler 3" is, for example, 120 to 160 parts by mass per 100 parts by mass of polymer. The blending ratio of "Crosslinking Agent" is, for example, 0 to 90 parts by mass per 100 parts by mass of polymer. Note that blending ratios and amounts of ingredients such as fillers and crosslinking agents may also be expressed in terms of blending amounts. Blending ratios and amounts are examples of blending information.

Examples of resins include polyolefins such as high-density polyethylene, low-density polyethylene, and ethylene-acrylic acid copolymer, as well as elastomers such as chlorinated polyethylene. In addition to the crosslinking agents shown in Figure 2, examples of compounding agents include fillers such as talc, calcium carbonate, and silica, plasticizers, and stabilizers. It is anticipated that the polymer composition will contain 20 or more types of resins, compounding agents, and other materials, but this is not limited to this number.

The blank rate of the data in dataset 32, excluding the composition ID, is assumed to be sparse, at 30% to 80%, in order to facilitate the preparation of dataset 32. Here, "blank rate" refers to the ratio of the number of cells in dataset 32 in which no data has been entered to the total number of cells in dataset 32 in which data can be entered. For example, in the case shown in Figure 2, when the compounding agent is only a filler, the total number of cells is 90, and the number of cells in which no data has been entered is 36, resulting in a blank rate of 40%. Furthermore, the number of data sets 321 in dataset 32 (= number of data) is assumed to be between 100 and 100,000 (for example, 350 pieces of data), but is not limited to this.

Next, we will explain in detail each of the units 21 to 25 that make up the prediction processing unit 2.

(Configuration of data generation unit)
FIG. 3A is a conceptual diagram for explaining an example of data generation processing by the data generator 21, and FIG. 3B is a conceptual diagram for explaining data generation processing by the leave-one-out method as a comparative example.

The data generation unit 21 divides the dataset 32 multiple times (n times) so that the combinations of data sets 321 are different for each group, thereby generating n training data groups LDG ₁ to LDG _n , each consisting of multiple (seven in FIG. 3 ) training data LD, and n evaluation data groups TDG ₁ to TDG _n , each consisting of multiple (two in FIG. 3 ) evaluation data TD. The division method may be random or may be based on a certain rule, as long as the combinations of data sets 321 are different for each group. "The combinations of data sets 321 are different" means that there are no training data groups in which all of the multiple evaluation data after division are identical.

The data generation unit 21 stores the generated n training data groups LDG ₁ to LDG _n in the storage unit 3 as a training data set 34, and stores the generated n evaluation data groups TDG ₁ to TDG _n in the storage unit 3 as an evaluation data set 35.

Furthermore, the ratio of the number of training data included in the training data group generated by the data generation unit 21 to the number of evaluation data included in the evaluation data group is preferably in the range of 7:3 to 8:2, for example, taking into account the balance between the prediction accuracy and evaluation accuracy of the regression model. The number of divisions is preferably set to be greater the greater the blank rate of the data, in order to suppress bias in the data after division. For example, if the blank rate of the data is 30% or greater, the number of divisions is preferably 30 or more.

If the number of data points in dataset 32 is small (e.g., less than 100,000), or if dataset 32 is sparse (e.g., with a blank space rate of 30% or more), creating a single regression model using all of the data without dividing dataset 32 may result in overfitting of the regression model to dataset 32, potentially reducing the predictive accuracy for data other than dataset 32. Furthermore, the hold-out method, in which a regression model is created and evaluated by dividing a dataset only once, is prone to data bias, potentially reducing either the predictive accuracy or the evaluation accuracy of the regression model. Furthermore, the leave-one-out method, as shown in Figure 3(b), divides dataset 32 into multiple training data groups and one evaluation data group. Therefore, increasing the number of training data groups increases the accuracy of the regression model. However, since there is only one evaluation data group, there is a risk that the evaluation accuracy of the regression model may be reduced. Furthermore, the increased number of divisions increases the processing time.

In this embodiment, as shown in FIG. 3(a), in order to obtain high prediction accuracy and evaluation accuracy of the regression model without lengthening the processing time, cross-validation is employed, in which dataset 32 is divided multiple times to create multiple regression models (hereinafter also referred to as "first regression models"), and the errors of each first regression model are averaged to evaluate the prediction accuracy of the regression method. Furthermore, since the prediction accuracy of regression methods varies depending on the physical properties, this embodiment evaluates the prediction accuracy of each regression method using cross-validation for multiple regression methods, selects the regression method with the highest evaluation, and uses the entire dataset to create a regression model (hereinafter also referred to as "second regression model") corresponding to the regression method with the highest evaluation. Note that random forests, for example, may be used as a method for dividing dataset 32 into evaluation data and training data.

(Configuration of the first regression model creation unit)
The first regression model creation unit 22 acquires one regression method from a regression method group 33 including multiple regression methods RM ₁ to RM _m , and uses the acquired regression method to create n first regression models 1M ₁ to 1M _n based on n training data groups LDG ₁ to LDG _n acquired from the training dataset 34. That is, the first regression model creation unit 22 creates one first regression model corresponding to the regression method by training a regression model using material data of multiple training data included in one training data group as explanatory variables and physical property values as response variables. This is then performed for other training data groups to create n first regression models 1M ₁ to 1M _n for the one regression method. This is then performed for other regression methods to finally create m×n first regression models 1M ₁ to 1M _n .

(Configuration of regression method determination unit)
The regression method determination unit 23 has an error calculation unit 231 and a regression method output unit 232, and evaluates the prediction accuracy of the N first regression models 1M ₁ to 1M _n created by the first regression model creation unit 22 for each regression method based on the N evaluation data groups TDG ₁ to TDG _n , and determines the regression method RMx with the highest evaluation from the multiple regression methods RM ₁ to RM _m . This regression method determination unit 23 will be described in detail below.

The error calculation unit 231 calculates, for each regression method and each first regression model, first average error rates 1MPE ₁ to 1MPE _n by averaging the errors caused by the first regression models based on the plurality of evaluation data included in the evaluation data group. By averaging the errors caused by the first regression models, the influence of data bias among the evaluation data can be suppressed.

Here, the first average error rate 1MPEi (i = 1 to n) is the average value of the error rates PEi (i = 1 to k) calculated for each evaluation data TD for multiple (e.g., k) evaluation data TD included in the evaluation data group TDG, as shown in the following equation (1).
1MPEi=( _PE1 + _PE2 +...+ _PEk )/k...(1)

The error rate PEi (i = 1 to k) is the difference (= error) expressed in percentage between the predicted value PV by the first regression model 1M and the measured value MV, which is a physical property value included in the evaluation data TD, as shown in the following formula (2). The predicted value PV is a value calculated using the first regression model 1M with the material data of the evaluation data TD as an explanatory variable.
PEi=(|PV-MV|/MV)×100...(2)

Furthermore, the error calculation unit 231 calculates a second average error rate 2MPEi (i = 1 to m) by averaging the first average error rates 1MPE ₁ to 1MPE _n of the multiple first regression models 1M ₁ to 1M _n , as the prediction accuracy of the regression method RMi (i = 1 to m), as shown in the following formula (3). Here, the first average error rate is an example of a first prediction error. The second average error rate is an example of a second prediction error. Note that the prediction error is not limited to the average error rate, and may be a mean square error, a mean square deviation, etc.
2MPEi=(1MPE ₁ +1MPE ₂ +...+1MPE _n )/n...(3)
By calculating the second average error rate by averaging the first average error rates, the influence of data bias among the training data groups can be suppressed.

(Configuration of the regression method output unit)
The regression method output unit 232 compares the second average error rates _2MPE1 to _2MPEm calculated by the error calculation unit 231 as prediction accuracy for each of the regression methods _RM1 to _RMm , and determines the regression method RMx with the highest prediction accuracy that corresponds to the smallest second average error rate. The regression method output unit 232 outputs identification information 232a of the determined regression method RMx to the second regression model creation unit 24. This makes it possible to determine a regression method suitable for predetermined physical properties.

(Configuration of the second regression model creation unit)
The second regression model creation unit 24 creates a second regression model 2M based on all data sets 321 included in the dataset 32 using the regression method RMx determined by the regression method determination unit 23. That is, the second regression model creation unit 24 creates a second regression model 2M corresponding to the regression method RMx by using all data sets 321 included in the dataset 32 as a training data group, using material data of the training data included in the training data group as explanatory variables, and physical property values as target variables. The second regression model creation unit 24 outputs the calculated second regression model 2M to the physical property value prediction unit 25. By using all data sets 321 included in the dataset 32 for machine learning, it is possible to create a second regression model that can predict physical property values with high accuracy even if the dataset does not have a sufficient number of data and is sparse. Note that not all data in the dataset 32 does not need to be used for machine learning; a number of pieces of training data (e.g., 90% or more of the total) greater than the number of pieces of training data when the dataset 32 is divided (e.g., 70% or 80% of the total) may be used for machine learning.

(Configuration of the physical property value prediction unit)
The physical property value prediction unit 25 uses the second regression model 2M created by the second regression model creation unit 24 to predict the physical property values of the new polymer composition based on the material data indicating the blending information of the materials that make up the new polymer composition input from the input unit 4, and displays the physical property values on the display unit 5. The physical property values may be displayed on the display unit 5 together with the input material data.

<Operation of the physical property prediction device>
Next, an example of the operation of the physical property prediction device 1 will be described with reference to Fig. 3 to Fig. 8. Fig. 7 is a block diagram for explaining an example of the overall flow of the physical property prediction device 1. Fig. 8 is a flowchart for explaining an example of the overall operation of the physical property prediction device 1.

(1) Data generation process (S1)
This data generation process (S1) will be described with reference to Fig. 3(a) above, which is a conceptual diagram for explaining an example of the data generation process by the data generation unit 21.

In this data generation process (S1), as shown in Figure 3(a), the data generation unit 21 divides the data set 32 n times so that the combinations of data sets 321 are different for each, and generates a plurality of training data groups LDG ₁ to LDG _n , each consisting of a plurality of training data (seven in Figure 3), and a plurality of evaluation data groups TDG ₁ to TDG _n , each consisting of a plurality of evaluation data (two in Figure 3).

Next, the data generating unit 21 stores the n training data groups LDG ₁ to LDG _n as a training data set 34 in the storage unit 3 , and stores the n evaluation data groups TDG ₁ to TDG _n as an evaluation data set 35 in the storage unit 3 .

(2) First regression model creation process (S2)
This first regression model creation process (S2) will be described with reference to Fig. 4 and Fig. 7. Fig. 4 shows the process in which the first regression model creation unit 22 creates n first regression models 1M ₁ to 1M _n for one regression method.

4 and 7 , in this first regression model creation process (S2), the first regression model creation unit 22 acquires one regression method _RM1 from the regression method group 33, acquires a training data group _LDG1 from the training dataset 34, and creates a first regression model _1M1 corresponding to the regression method _RM1 based on the training data group _LDG1 . Next, the first regression model creation unit 22 acquires a training data group _LDG2 from the training dataset 34 and creates a first regression model _1M2 corresponding to the regression method _RM1 based on the training data group _LDG2 . For the other training data groups _LDG3 to _LDGn , the first regression model creation unit 22 creates first regression models _1M3 to _1Mn corresponding to the regression method RM1 based on the other training data groups _LDG3 to _LDGn in the same manner as described _above .

The first regression model creation unit 22 executes the above process for the other regression methods RM ₂ to RM _m to create a total of m×n first regression models 1M ₁ to 1M _n . The first regression model creation unit 22 links the created m×n first regression models 1M ₁ to 1M _n to the identification information of the corresponding regression model, and outputs them to the regression method determination unit 23.

(3) Prediction accuracy calculation process (S3)
This prediction accuracy calculation process (S3) will be described with reference to Figures 5(a), 5(b), and 7. Figure 5(a) shows the process by which the error calculation unit 231 calculates the prediction accuracy of the first regression models _1M1 to _1Mn for one regression method. Figure 5(b) shows the process by which the error calculation unit 231 calculates the second average error rates _2MPE1 to _2MPEm .

5A, the error calculation unit 231 calculates first average error rates 1MPE1 to 1MPEn as the prediction accuracy of n first regression models _1M1 to _1Mn created corresponding to one regression method. That is, the error calculation unit 231 calculates the first average error rate _1MPE1 of the first regression model _1M1 corresponding to the regression method _RM1 based on the evaluation data group _TDG1 . Next, the error calculation unit ₂₃₁ calculates the first average error rate _1MPE2 of the first regression model _1M2 corresponding to the regression method _RM1 based on the evaluation data group _TDG2 . For the first regression models _1M3 to _1Mn , the error calculation unit 231 calculates the first average error rates _1MPE3 to _1MPEn of the first regression models _1M3 to _1Mn corresponding to the regression method _RM1 based on the other evaluation data groups _TDG3 to _TDGn in the _same manner as described above.

The error calculation unit 231 executes the above process for the other regression methods RM ₂ to RM _m , and calculates a total of m×n first average error rates 1MPE ₁ to 1MPE _n .

Next, as shown in Figure 5(b), the error calculation unit 231 calculates a second average error rate 2MPE1 by averaging the first average error rates _1MPE1 to _1MPEn calculated for one regression method _RM1 . The error calculation unit 231 similarly calculates second average error rates _2MPE2 to 2MPEm for the other regression methods _RM2 to _RMm . As shown in Figures ₅ (b) and 7, the error calculation unit 231 links the calculated _m second average error rates _2MPE1 to _2MPEm to the identification information of the corresponding regression methods and outputs them to the regression method output unit 232.

(4) Determining the regression method (S4)
7, the regression method output unit 232 compares the m second average error rates 2MPE ₁ to 2MPE _m output from the error calculation unit 231, and determines the regression method RMx with the highest prediction accuracy that corresponds to the smallest second average error rate. The regression method output unit 232 outputs identification information 232a of the determined regression method RMx to the second regression model creation unit 24.

(5) Creation of a second regression model (S5)
The creation of the second regression model (S5) will be described with reference to Fig. 6 and Fig. 7. Fig. 6 is a diagram for explaining an example of the second regression model creation process by the second regression model creation unit 24.

In creating this second regression model (S5), the second regression model creation unit 24 acquires, from the regression method group 33, the regression method RMx corresponding to the identification information 232a of the regression method RMx output from the regression method output unit 232 of the regression method determination unit 23, as shown in FIG. 7, and creates a second regression model 2M based on all data sets 321 included in the dataset 32 using the regression method RMx, as shown in FIG. 6, and outputs the second regression model 2M to the physical property value prediction unit 25.

(6) Prediction of physical properties (S6)
In this prediction of physical property values (S6), the physical property value prediction unit 25 uses the second regression model 2M output from the second regression model creation unit 24 to predict physical property values based on the material data of the new polymer composition input from the input unit 4, and the display unit 5 displays the physical property values.

<Actions and Effects of the Present Embodiment>
According to this embodiment, the following actions and effects are achieved.
(a) Even if the data set 32 does not have a sufficient number of data and is sparse, the number of times the data set 32 is divided is set according to the blank rate of the data, a first regression model is created for each of the divisions for each regression method, and the errors of each first regression model are averaged to evaluate the regression method. This reduces the influence of bias in the data that occurs when the data is divided, and makes it possible to select a regression method that corresponds to the required physical properties.
(b) By selecting the regression method with the highest prediction accuracy from among a plurality of predetermined regression methods and creating a regression model corresponding to the selected regression method using all the data in the data set 32, it is possible to perform highly accurate predictions of physical property values corresponding to required physical properties.

The number of data points in Dataset 32 was set to 350, the blank rate in Dataset 32 was set to 40%, the number of training data points and the number of evaluation data points were divided into 7:3, and an experiment was conducted to calculate the average error rate using the Gaussian process regression method.

As a result, in Comparative Example 1, in which dataset 32 was divided once using the holdout method to create and evaluate a regression model, the first average error rate was 6%. In Comparative Example 2, in which dataset 32 was divided once again using the holdout method to create and evaluate a regression model, the first average error rate was 11%. On the other hand, in this example, in which dataset 32 was divided 30 times, the second average error rate was 2% or less, demonstrating that more accurate prediction of physical property values can be achieved than with the holdout method.

(Summary of the embodiment)
Next, the technical ideas grasped from the above-described embodiments will be described by using the reference numerals and the like in the embodiments. However, the reference numerals in the following description do not limit the components in the claims to the members and the like specifically shown in the embodiments.

[1] A data generation unit (21) that divides a data set (321) of material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of the physical properties of the composite material into a data set (321) for each of a plurality of types of composite materials so that the combinations of the data set (321) are different from each other, thereby generating a plurality of training data groups and a plurality of evaluation data groups;
a first regression model creation unit (22) that creates a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a regression method determination unit (23) that evaluates the prediction accuracy of the plurality of first regression models for each regression method based on the plurality of evaluation data groups, and determines the regression method with the highest evaluation among the plurality of regression methods;
a second regression model creation unit (24) that creates a second regression model based on the plurality of data sets (321) included in the data set (32) using the regression method with the highest evaluation;
a property value prediction unit (25) that predicts a property value based on input material data using the second regression model.

[2] The device (1) for predicting physical properties of a composite material described in [1], wherein the regression method determination unit (23) includes an error calculation unit (231) that calculates a first prediction error by averaging the errors of the first regression model for each regression method and each first regression model based on the multiple evaluation data included in the evaluation data group, and calculates a second prediction error by averaging the first prediction errors of the multiple first regression models as the prediction accuracy of the regression method, and a regression method output unit (232) that compares the prediction accuracy calculated for each regression method and outputs the regression method with the highest evaluation.

[3] The device (1) for predicting physical properties of a composite material described in [1] or [2], wherein the second regression model creation unit (24) creates the second regression model based on the data set that constitutes the data set (32).

[4] The composite material property value prediction device (1) described in any one of [1] to [3], wherein the number of data sets (321) constituting the data set (32) is 100 or more and less than 100,000.

[5] The composite material property value prediction device (1) described in any one of [1] to [4], wherein the blank rate of the dataset, where no data exists, is 30% or more and 80% or less.

[6] The number of times the data set is divided by the data generation unit is a number corresponding to a blank rate where no data exists in the data set.
The apparatus for predicting physical properties of a composite material according to any one of [1] to [5].

[7] The composite material property value prediction device (1) described in any one of [1] to [6], wherein the composite material is a polymer composition containing multiple resins and multiple compounding agents.

[8] A computer includes a data generation unit (21) that divides a data set (321) of material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of the physical properties of the composite material into a data set (32) for each of a plurality of types of composite materials so that the combinations of the data set (321) are different from each other, thereby generating a plurality of training data groups and a plurality of evaluation data groups;
a first regression model creation unit (22) that creates a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a regression method determination unit (23) that evaluates the prediction accuracy of the plurality of first regression models for each regression method based on the plurality of evaluation data groups, and determines the regression method with the highest evaluation among the plurality of regression methods;
a second regression model creation unit (24) that creates a second regression model based on the plurality of data sets (321) included in the data set (32) using the regression method with the highest evaluation;
a physical property value prediction unit (25) that predicts physical property values based on input material data using the second regression model;

[9] A generation step (S1) of dividing a data set (321) of material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of the physical properties of the composite material into a data set (32) for each of a plurality of types of composite materials so that the combinations of the data sets (321) are different from each other, thereby generating a plurality of training data groups and a plurality of evaluation data groups;
a first creation step (S2) of creating a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a determination step (S4) of evaluating the prediction accuracy of the plurality of first regression models for each regression method based on the plurality of evaluation data groups, and determining the regression method (RM _x ) having the highest evaluation among the plurality of regression methods;
a second creation step (S5) of creating a second regression model based on a plurality of the data sets (321) included in the data set (32) using the regression method with the highest evaluation;
a prediction step (S6) of predicting physical property values based on input material data using the second regression model.

The above describes an embodiment of the present invention, but the present invention is not limited to the above embodiment and various modifications and implementations are possible.

In addition, some of the components of the above embodiments may be omitted or modified. Furthermore, steps may be added, deleted, modified, or rearranged in the flow of the above embodiments.

1... physical property value prediction device, 2... prediction processing unit, 3... storage unit, 4... input unit, 5... display unit,
21...data generation unit, 22...first regression model creation unit, 23...regression method determination unit,
24... second regression model creation unit, 25... physical property value prediction unit, 31... physical property value prediction program,
32...dataset, 33...regression method group, 34...training dataset,
35...Evaluation dataset, 231...Error calculation unit,
232...Regression method output unit, 232a...Identification information, 320...Composition ID,
321...data set, 321a...material data, 321b...physical property data,
LDG ₁ to LDG _n ... training data group, TDG ₁ to TDG _n ... evaluation data group,
LD... training data, TD... evaluation data, RM ₁ to RM _m ... regression method,
1M ₁ to 1M _n ...first regression model, 1MPE ₁ to 1MPE _n ...first average error rate,
2MPE ₁ to 2MPE _m ... second mean error rate, 2M... second regression model

Claims (8)

a data generation unit that divides a data set, which includes material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of physical properties related to the composite material, for each of a plurality of types of composite materials a plurality of times so that the combinations of the data sets are different from one another, thereby generating a plurality of training data groups and a plurality of evaluation data groups;
a first regression model creation unit that creates a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a regression method determination unit that evaluates the prediction accuracy of the plurality of first regression models for each of the regression methods based on the plurality of evaluation data groups, and determines the regression method with the highest evaluation among the plurality of regression methods;
a second regression model creation unit that creates a second regression model based on the plurality of data sets included in the data set using the regression method with the highest evaluation;
a physical property value prediction unit that predicts a physical property value based on input material data using the second regression model;
A physical property value prediction device for a composite material, comprising :
The number of times the data set is divided by the data generation unit is equal to or greater than the numerical value of the blank rate where no data exists in the data set.
A device for predicting the physical properties of composite materials. The regression method determination unit includes an error calculation unit that calculates, for each of the regression methods and each of the first regression models, a first prediction error by averaging errors due to the first regression model based on a plurality of evaluation data included in the evaluation data group, and calculates, as a prediction accuracy of the regression method, a second prediction error by averaging the first prediction errors due to the plurality of first regression models; and a regression method output unit that compares the prediction accuracy calculated for each of the regression methods and outputs the regression method with the highest evaluation.
The apparatus for predicting physical property values of a composite material according to claim 1 . the second regression model creation unit creates the second regression model based on the data sets that constitute the data set.
The apparatus for predicting physical property values of a composite material according to claim 1 or 2. The number of the data sets constituting the data set is 100 or more and less than 100,000.
The apparatus for predicting physical property values of a composite material according to claim 1 . The blank rate of the dataset, where no data exists, is 30% or more and 80% or less.
The apparatus for predicting physical property values of a composite material according to claim 4. The composite material is a polymer composition containing a plurality of resins and a plurality of compounding agents.
The apparatus for predicting physical property values of a composite material according to claim 1 . Computer,
a data generation unit that divides a data set, which includes material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of physical properties related to the composite material, for each of a plurality of types of composite materials a plurality of times so that the combinations of the data sets are different from one another, thereby generating a plurality of training data groups and a plurality of evaluation data groups;
a first regression model creation unit that creates a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a regression method determination unit that evaluates the prediction accuracy of the plurality of first regression models for each of the regression methods based on the plurality of evaluation data groups, and determines the regression method with the highest evaluation among the plurality of regression methods;
a second regression model creation unit that creates a second regression model based on the plurality of data sets included in the data set using the regression method with the highest evaluation;
a physical property value prediction unit that predicts a physical property value based on input material data using the second regression model;
A physical property value prediction program for a composite material that functions as
The number of times the data set is divided by the data generation unit is equal to or greater than the numerical value of the blank rate where no data exists in the data set.
A program for predicting the physical properties of composite materials. The computer
a generation step of dividing a data set, which includes material data indicating blending information of materials constituting a composite material and physical property data indicating physical property values of physical properties related to the composite material, for each of a plurality of types of composite materials a plurality of times so that the combinations of the data sets are different from one another to generate a plurality of training data sets and a plurality of evaluation data sets;
a first regression model creation step of creating a plurality of first regression models for each of a plurality of predetermined regression methods based on the plurality of learning data groups;
a determination step of evaluating the prediction accuracy of the plurality of first regression models for each of the regression methods based on the plurality of evaluation data groups, and determining the regression method with the highest evaluation among the plurality of regression methods;
a second generation step of generating a second regression model based on a plurality of the data sets included in the data set using the best-performing regression method;
a prediction step of predicting a physical property value based on input material data using the second regression model;
A method for predicting physical properties of a composite material, comprising :
The number of times the dataset is divided in the generating step is equal to or greater than the numerical value of the blank rate where no data exists in the dataset.
Methods for predicting the physical properties of composite materials.