WO2021079985A1

WO2021079985A1 - Property prediction device

Info

Publication number: WO2021079985A1
Application number: PCT/JP2020/039900
Authority: WO
Inventors: 祐子池田; 倫弘奥山; 幸仁中澤; 押山　智寛; 船津　公人
Original assignee: コニカミノルタ株式会社
Priority date: 2019-10-25
Filing date: 2020-10-23
Publication date: 2021-04-29
Also published as: JP7468542B2; JPWO2021079985A1

Abstract

The objective of the present invention is to provide a property prediction device capable of predicting the properties of a polymer composite material without using the molecular structure or a physical property value of a material as a descriptor. The present invention for achieving this objective relates to a property prediction device comprising: a prediction model storage unit for storing a prediction model, which, with respect to an input of material information regarding the material of a polymer composite material, outputs mechanical properties of the polymer composite material to be produced; and a learning processing unit for performing learning processing on the prediction model on the basis of the material information and the mechanical properties of the polymer composite material produced from the material. Regarding the material information, it is difficult to gather information indicating the physicochemical properties for all the candidate materials. As the material information, the property prediction device uses information including the material mixture ratio and information not directly indicating a physicochemical property.

Description

Characteristic predictor

The present invention relates to a characteristic prediction device.

According to Materials Informatics (MI), it is possible to mechanically design and search for materials that achieve desired properties based on past experimental results and simulation data, making the development of new materials more efficient. It is expected that it can be done in a targeted manner.

For example, in Non-Patent Document 1, a library of prediction models for predicting physical property values of organic molecules is created by machine learning using a descriptor obtained by fingerprinting, which is data obtained by quantifying the molecular structure, as an explanatory variable. Can be used for transfer learning.

Further, in Non-Patent Document 2, a database including the physical property values of the polymer resin, the mixing ratio of the filler, and the physical properties of the obtained polymer composite material is classified into classes by machine learning, and the physical properties of the polymer composite material are classified. It is described that the characteristics that contribute to the above are searched.

As described in Non-Patent Document 1 and Non-Patent Document 2, attempts have been made to predict and analyze the characteristics of composite materials containing polymer resins using MI. However, for the reasons described below, it is very difficult to apply MI to polymer composite materials, and even with the methods described in these documents, it cannot be said that the desired efficiency improvement has been achieved yet.

First, the experimental data reported so far regarding the molecular structure used for the descriptor in Non-Patent Document 1 and the physical property values used for classification in Non-Patent Document 2 were obtained by the same experimental method. However, there is a problem that these cannot be compared under the same conditions. For example, it is known that the experimental value of the average molecular weight varies depending on the temperature, flow rate, and type of the eluent used for the measurement. It is also known that the experimental value of the decomposition temperature also fluctuates depending on the temperature rise temperature at the time of measurement and the flow velocity of the gas. Further, regarding the physical characteristic values used for the classification in Non-Patent Document 2, the types of the physical characteristic values disclosed differ depending on the manufacturer that provides the material, and the physical characteristic values are known for all the materials. is not it. Furthermore, as a condition for purchasing materials, it is required not to measure the physical property values, and it is very difficult to prepare the data of the physical property values.

In the case of low-molecular-weight compounds, the descriptor converted into a fingerprint, which is a method for quantifying molecular structure information, is used as an explanatory variable, and machine learning is used to predict the desired physical property value of the new material. Proposals are being made. However, in the case of polymer compounds and fillers, even if they are commercially available products, as described above, information on the molecular structure, which is one of the physicochemical properties, is often not sufficiently disclosed. In some cases, the monomer structure is converted to a fingerprint as an alternative index, but even with the same polypropylene, basic physical properties such as viscosity, glass transition temperature, decomposition temperature, etc. are different due to differences in the degree of polymerization, average molecular weight, molecular weight distribution, copolymerization ratio, etc. Has a great effect on the mechanical characteristics because it changes. Therefore, it is very difficult to prepare appropriate data for analysis using MI for polymer materials.

The present invention has been made based on the above findings, and an object of the present invention is to provide a property prediction device capable of predicting the properties of a polymer composite material without using the molecular structure or physical property value of the material as a descriptor. To do.

The above-mentioned problems include a prediction model storage unit that stores a prediction model that outputs mechanical properties of a manufactured polymer composite material in response to input of material information regarding the material of the polymer composite material, and the material information. The data including the mechanical properties of the polymer composite material produced from the material is used as training data, and the learning processing unit for performing the learning process on the prediction model is provided. The material information is information that makes it difficult to prepare information indicating physicochemical properties for all candidate materials, and the property predictor directly indicates physicochemical properties as the material information. Information including no information and the compounding ratio of the material is used.

Further, the above-mentioned problem is the material information acquisition unit for acquiring the material information regarding the material of the polymer composite material, and the polymer composite material manufactured by using the predicted model learned for the input material information. This is solved by a characteristic predictor having a characteristic information output unit that outputs mechanical characteristics. The material information is information that is difficult to prepare information indicating physicochemical properties for all candidate materials, and in the property prediction device, the input material information directly obtains physicochemical properties. The prediction model uses learning data including information not shown in the above and the compounding ratio of the material, and the prediction model includes the material information and the mechanical properties of the polymer composite material produced from the material. The learning process that was used has been applied.

INDUSTRIAL APPLICABILITY The present invention provides a property prediction device capable of predicting the properties of a polymer composite material without using the molecular structure or physical property value of the material as a descriptor.

FIG. 1 is a block diagram showing a hardware configuration of a characteristic prediction device according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example of the functional configuration of the characteristic prediction device according to the embodiment of the present invention. FIG. 3 is a flowchart showing a learning operation example of the characteristic prediction device according to the embodiment of the present invention. FIG. 4 is a flowchart showing a prediction operation example of the characteristic prediction device according to the embodiment of the present invention. FIG. 5 is a graph showing the relationship between _{the number of latent variables and ε 2} at the time of constructing the prediction model 2 constructed in the specific example. FIG. 6 shows the predicted elastic modulus output from the material information of each of the 180 experimental data using the prediction model 2 constructed in the specific example, and the measured elastic modulus in each of the 180 experimental data. It is a graph showing the relationship of. FIG. 7 shows the predicted value of the elastic modulus of the polymer composite material calculated from the data of the predicted value and the measured value shown in FIG. 6 and the absolute value of the residual between the predicted value and the measured value. It is a graph which shows the relationship. FIG. 8 shows the relationship between the predicted elastic modulus output from each material information using the predicted model 2 and the measured elastic modulus for the nine levels used for the verification of the predicted model 2. It is a graph showing. FIG. 9 is a graph showing the relationship between the filler content and the measured and predicted elastic modulus of the five data used for examining the factors contributing to the elastic modulus in the high elastic modulus region.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The present invention is not limited to the following forms.

[Characteristic prediction device]
One embodiment of the present invention relates to a property predictor that outputs the mechanical properties of a polymer composite material to be manufactured in response to input of material information which is information about the material of the polymer composite material. The characteristic prediction device has a prediction model for making the prediction, and learns (newly constructs or updates) the prediction model by machine learning.

FIG. 1 is a block diagram showing a hardware configuration of a characteristic prediction device according to an embodiment of the present invention.

The characteristic prediction device 100 is, for example, a computer including a CPU 110, a ROM 120, a RAM 130, an external storage device 140, a communication interface 150, an input device 160, an output device 170, and the like. Examples of the external storage device 140 include HDDs, SSDs, flash memories, and the like. Examples of the communication interface 150 include a communication controller for a LAN line and the like. Examples of the input device 160 include a keyboard, mouse, touch panel, scanner, bar code reader, and the like. Examples of the output device 170 include display devices such as CRTs and liquid crystals, as well as printers and the like.

Each function described later in the characteristic prediction device 100 is realized, for example, by the CPU 110 referring to a processing program and various data stored in the ROM 120, the RAM 130, the external storage device 140, and the like. However, some or all of the above-mentioned functions may be realized by processing by a dedicated hardware circuit instead of or by processing by software.

(Functions of each component)
FIG. 2 is a block diagram showing an example of the functional configuration of the characteristic prediction device 100 according to the present embodiment.

As shown in FIG. 2, the characteristic prediction device 100 includes a material information acquisition unit 210, a characteristic information output unit 220, an experimental data acquisition unit 230, a learning processing unit 240, and a prediction model storage unit 250.

The material information acquisition unit 210 acquires the material information input from the input device 160.

The above material information is information about a material used for manufacturing a polymer composite material to be manufactured. In the present embodiment, the polymer composite material is a composite material containing a polymer resin and an additive such as a filler. Therefore, in the present embodiment, the material information includes at least information on the polymer resin and information on the filler.

It is expected that the mechanical properties (for example, elastic modulus) of the produced polymer composite material will be greatly influenced by the information indicating the physicochemical properties of the polymer resin and additives that are the materials. Therefore, when it is desired to predict the mechanical properties of a polymer composite material, it is considered desirable to construct and use a prediction model using the information indicating the physicochemical properties as an identifier. The information indicating the physicochemical properties includes, for example, the weight average molecular weight and the number average molecular weight of the polymer resin, the molecular weight distribution, the degree of copolymerization, the degree of cross-linking, various other physical property values, the size of the additive, and the like. Various physical property values and the like. Further, the above-mentioned additives are polymers that affect the mechanical properties of the polymer composite material to be produced, such as fillers, plasticizers, colorants, flame retardants, ultraviolet absorbers, antioxidants and elastomers. It is a compound other than resin.

However, in general, it is not possible to gather information showing these physicochemical properties for all of the material candidates that may be used in the production of polymer composite materials (hereinafter, also simply referred to as "candidate materials"). Have difficulty. For example, which of the information indicating these physicochemical properties is disclosed and which is not disclosed depends on the manufacturer that provides the material. Therefore, even if an attempt is made to construct a prediction model using information indicating these physicochemical properties as an identifier, the identifier data and the mechanical properties (elastic modulus) of the polymer composite material manufactured from the material are included. It is difficult to prepare the data of the experimental results and the teacher data that are complete.

On the other hand, in the present embodiment, information that does not directly indicate the physicochemical properties and the compounding ratio of the material are used as identifiers. Information that does not directly indicate the physicochemical properties is to specify the material, although it does not directly indicate the physicochemical properties of the material, such as information indicating the brand of the material. It can be ancillary information that can be used.

The characteristic information output unit 220 outputs the mechanical characteristics of the polymer composite material to be manufactured by using the learned prediction model. Specifically, the characteristic information output unit 220 is manufactured by inputting the material information into the prediction model when the material information acquisition unit 210 acquires the material information, specifically, the brand of the material and the blending ratio thereof. Obtain the output of the mechanical properties of polymer composites.

The output mechanical characteristics are transmitted to the output device and output so that they can be recognized from the outside.

The experiment data acquisition unit 230 acquires the experiment data input from the input device 160.

The above experimental data is data including the above material information and mechanical properties obtained by an experiment in which the above polymer material is manufactured and its mechanical properties are measured. The above experimental data is data including information that does not directly indicate the physicochemical properties, the compounding ratio of the material, and the mechanical properties measured from the obtained polymer composite material. It is used as teacher data for the learning processing unit 240 to train and process the prediction model.

The learning processing unit 240 uses the experimental data acquired by the experimental data acquisition unit 230 as learning data, and applies mechanical learning processing to the prediction model stored in the prediction model storage unit 250, and performs the above learning processing. The prediction model is stored in the prediction model storage unit 250.

In the present embodiment, the prediction model is a prediction model using the following linear regression equation (1) constructed by using partial least squares regression (PLS).

Wherein (1), beta is a constant term, c _i is the i-th partial regression coefficients, x _i is the i-th explanatory variable, y is the dependent variable, N is the in the number of explanatory variables is there.

The prediction model may be composed of any known statistical model, and a neural network, a decision tree, and a random PP and a filler are constructed by a method such as Niforest, Kernel-Based PLS (KPLS) support vector regression, or the like. It may be a prediction model. Further, it may be a prediction model constructed by a regression analysis method other than PLS, such as principal component regression (PCR) or ridge regression.

The prediction model storage unit 250 stores the prediction model learned by the learning processing unit 240.

The prediction model storage unit 250 may store only one prediction model, or may store a plurality of prediction models having different types of input material information or output mechanical properties.

(Example of learning operation)
Next, an example of the learning operation of the characteristic prediction device 100 according to the present embodiment will be described. FIG. 3 is a flowchart showing an example of learning operation of the characteristic prediction device 100.

First, the experimental data acquisition unit 230 was measured from the information input to the input device 160 that does not directly indicate the physicochemical properties, the blending ratio of the material, and the obtained polymer composite material. Obtain experimental data including mechanical properties (step S110). The experiment data acquisition unit 230 outputs the acquired experiment data to the learning processing unit 240. The experimental data acquisition unit 230 may output to the learning processing unit 240 each time one experimental data is acquired, or may output to the learning processing unit 240 after acquiring a plurality of experimental data.

Next, the learning processing unit 240 uses the material information (information that does not directly indicate the physicochemical properties and the blending ratio of the material) included in the experimental data as explanatory variables in the prediction model storage unit 250. , It is determined whether or not there is a prediction model whose output value is the mechanical characteristics included in the above experimental data (step S120).

As a result of the determination, when the prediction model exists (step S120, YES), the learning processing unit 240 uses the material information included in the experimental data and the mechanical characteristics as training data to perform the learning process of the prediction model. (Step S130). After that, the process proceeds to step S150. The learning process in step S130 uses a large amount of experimental data, and the error between the mechanical characteristics of each experiment included in the experimental data and the output value from the prediction model becomes small (converges). ) May be repeated.

On the other hand, when the prediction model does not exist (step S120, NO), the learning processing unit 240 newly creates a prediction model in which the material information included in the experimental data is used as an explanatory variable and the mechanical characteristics are used as the output value. Build (step S140).

After that, the learning processing unit 240 stores the learned (or newly constructed) prediction model in the prediction model storage unit 250 (step S150).

(Example of predicted operation)
Next, an example of the prediction operation of the characteristic prediction device 100 according to the present embodiment will be described. FIG. 4 is a flowchart showing an example of prediction operation of the characteristic prediction device 100.

The experimental data was created by the following procedure.

1-1. Preparation of Polymer Composite Material PP and filler were kneaded with a twin-screw kneader (manufactured by Xplore Instruments, product name MC15) to obtain a kneaded product. The temperature at the time of kneading was 200 ° C., the rotation speed at the time of adding the material was 80 rpm, and after the material was added, the temperature was raised to 130 rpm and kneaded for 5 minutes. An injection molding machine (manufactured by Xplore Instruments, product name IM12) was used to produce a molded product having a shape conforming to ISO527-2-1BA (2012). The molding conditions were a cylinder temperature of 200 ° C., a mold temperature of 60 ° C., an injection pressure, and a time of 10 barr to 15 barr, 18 seconds.

1-2. Measurement of elastic modulus The elastic modulus was measured by a tensile test using a Tencilon universal material tester (manufactured by A & D Co., Ltd.) according to an initial load of 0.3 N and a moving speed of 1 mm / min.

2. Construction of prediction model 2-1. Setting of explanatory variables The explanatory variables of PP are described by the following equation (2).

Equation _{(2), x pi} is PP explanatory variables, _{p i} is PP stock (i = 1 ~ 11), α indicates that the addition of material indicated by the explanatory variable.
First, the material information acquisition unit 210 includes information that is input to the input device 160 and that does not directly indicate the physicochemical properties of the material of the polymer composite material to be manufactured, and the mixing ratio of the material. Acquire material information including, (step S210). The material information acquisition unit 210 outputs the acquired material information to the characteristic information output unit 220.

Next, the characteristic information output unit 220 acquires the predicted mechanical characteristic value of the polymer composite material, which is output as a result of inputting the material information into the prediction model stored in the prediction model storage unit 250. (Step S220).

Next, the characteristic information output unit 220 outputs the acquired mechanical characteristic value to the output device 170 (step S230). The output mechanical characteristic value is displayed on the output device 170 or the like as a mechanical characteristic value predicted from the material information.

[Concrete example]
The specific methods and results of the construction of the prediction model by the characteristic prediction device 100 and the actual prediction are shown below.

1. 1. Preparation of experimental data Candidate materials were 11 types of polypropylene (PP), 18 types of fillers, and 20 types of other additives. Data including the selection of the material and the compounding ratio thereof of the polymer composite material produced by the combination arbitrarily selected from the above candidate materials and the elastic modulus measured for the obtained 180 kinds of polymer composite materials are provided. It was used as experimental data. The other additives are additives other than fillers, such as colorants, flame retardants, ultraviolet absorbers, antioxidants and elastomers.

The explanatory variables of the filler are described by the following equation (3).

Wherein _{(3), x fi} filler explanatory variables, _{f i} filler stock _{(i = 1 ~ 18),} c fi the ratio of the added amount of the fillers with respect to the addition amount of PP (wt%), alpha is It indicates that the material indicated by the explanatory variable was added.

The explanatory variables of other additives are described by the following formula (4).

In formula (4), x _ai is the explanatory variable _{of the other additive, a i} is the brand of the other additive (i = 1 to 20), and c _ai is the amount of the other additive added relative to the amount of PP added. (Mass%) and α indicate that the material indicated by the explanatory variable was added.

Using the equations (2) to (4), the explanatory variables by the vector representation illustrated in the equation (5) were set. Incidentally, the formula (5) is, PP stock _{p 1,} 10 wt% of a filler stock _{f 1} with respect to the addition amount of PP, and other additives stock _{a 1} added 5% by weight, based on the weight of the PP It is shown that a polymer composite material was prepared using the agent.

2-2. Construction of Prediction Model Using the above experimental data, a prediction model using linear regression equation (1) was constructed by partial least squares regression (PLS). At this time, a plurality of prediction models were constructed by changing the number of latent variables.

2-2-1. Construction of Prediction Model 1 At the time of construction of Prediction Model 1, 153 (85%) randomly selected out of 180 experimental data were used as teacher data to construct a prediction model, and the remaining 27 (15%) were tested. It was used as data.

At this time, the number of latent variables, expressed by the following equation (6), mean square square error epsilon ₁ in the test data _(Root Mean Square Error: RMSE) is the smallest value.

In formula (6), N is the number of experimental data (180), _{yobs, i} are the measured values of elastic modulus in the test data i, and _{ypred, i} are the material information of the test data i using the constructed learning data. The output predicted value of elastic modulus is shown.

In the prediction model 1, when the number of latent variables was 13, ε ₁ was the minimum of 156 MPa, and the coefficient of determination R ^{2 at} this time was 0.95. From the value of the coefficient of determination, it can be said that the accuracy of the prediction model 1 is generally good.

2-2-2. Construction of Prediction Model 2 At the time of construction of Prediction Model 2, we tried to construct a prediction model in which the number of latent variables was smaller than that of Prediction Model 1, the generalization was improved regardless of the selection of teacher data, and the error was smaller. ..

Specifically, the number of latent variables was set to the value that minimizes the _{error ε 2} expressed by the following equation (7), which is obtained by Leave-One-Out cross-validation (LOOCV).

In formula (7), N is the number of experimental data (180), y _i is the measured elastic modulus in the test data i,

Shows the predicted value of the elastic modulus output from the material information of the test data i using the prediction model constructed by constructing 179 data excluding the test data i as the teacher data.

FIG. 5 is a graph showing the relationship between the number of latent variables and ε _{2 at this time.} _{From FIG. 5, it can be seen that ε 2} is the minimum value (309 MPa) when the number of latent variables is four. Therefore, the prediction model when there are four latent variables is set as the prediction model 2. The coefficient of determination R ² of the prediction model 2 was 0.73.

FIG. 6 shows the relationship between the predicted elastic modulus output from the material information of each of the 180 experimental data using the prediction model 2 and the measured elastic modulus in each of the 180 experimental data. It is a graph. From the value of the coefficient of determination and the result of FIG. 5, it can be said that the accuracy of the prediction model 2 is generally good.

3. 3. Verification of prediction model 3-1. Verification of the residual between the predicted value and the measured value Fig. 7 shows the predicted value of the elastic modulus of the polymer composite material calculated from the data of the predicted value and the measured value shown in FIG. 6, and the predicted value and the measured value. It is a graph which shows the relationship with the absolute value of the residual between the values.

In FIG. 7, in the range where the predicted elastic modulus is 2500 MPa or less, the total number of data for which the absolute value of the residual can be set to 300 MPa or less (a value smaller than _{309 MPa, which is ε 2 of the prediction model) is the total.} It was 90% of. In the range where the predicted elastic modulus was larger than 2500 MPa, the number of data for which the absolute value of the residual could be 300 MPa or less was smaller, but the number of teacher data was smaller (see FIG. 6). ) It is thought that it depends.

3-2. Verification of accuracy by prediction of additional experimental results Prediction model 2 for material information of 570,000 polymer composite materials in which the combination of PP, filler and other additives, and the amount of filler and other additives added were changed. Was used to calculate the predicted elastic modulus.

Of the above 570,000 combinations, the explanatory variables (Equation (5)) include (i) a level close to the experimental data used to create the prediction model and (ii) a level far from the experimental data. , Nine levels were selected to make a polymer composite and the elastic modulus was measured. The preparation of the polymer composite material and the measurement of the elastic modulus were carried out in the same manner as when the experimental data were prepared.

FIG. 8 is a graph showing the relationship between the predicted value of the elastic modulus output from each material information using the prediction model 2 and the measured value of the elastic modulus for the above nine levels.

For levels B to H whose predicted values are in the range of 1500 to 3500 MPa ((i) the explanatory variables correspond to levels close to the experimental data), the absolute value of the residual between the predicted values and the measured values is within 300 MPa. Therefore, it can be said that the accuracy of the prediction model 2 is good. For levels A and H whose predicted values are outside the range of 1500 to 3500 MPa ((ii) explanatory variables correspond to levels far from the experimental data), the absolute value of the residual between the predicted values and the measured values is It was larger than 300 MPa and the prediction accuracy was low. It is considered that this is because the number of teacher data was small (see FIG. 6).

3-3. Examination of factors contributing to elastic modulus in the high elastic modulus region Regarding the 180 experimental data and the predicted elastic modulus output from these material information, the residual between the measured elastic modulus and the predicted elastic modulus is the largest. When the characteristics of the explanatory variables were examined for the data, the data had a high filler content (40% by mass).

Therefore, five data having the same filler (f ₄ ) content of 40%, 30%, 20%, 10%, and 1% in the material information were extracted from 180 experimental data and predicted values, and these were extracted. The measured value and the predicted value of the data of were examined.

FIG. 9 is a graph showing the relationship between the filler content and the measured and predicted elastic moduli of the above five data.

As shown in FIG. 9, the predicted value of the elastic modulus by the prediction model 2 tends to increase linearly as the filler content increases. However, it can be seen that the measured elastic modulus has a point (appropriate point) where the increase in elastic modulus due to the content of the filler reaches a plateau.

In FIG. 8, the level H, which also has a high elastic modulus and a decrease in prediction accuracy, is a polymer composite material to which a _{filler (f 10) of a type different from the data in FIG. 9 is added.} This suggests that the prediction accuracy may decrease at a high level of filler content regardless of the type of filler.

From these results, it can be said that the accuracy of the prediction model 2 is good in the range where the filler content is more than 0% by mass and 35% by mass or less. In order to improve the prediction accuracy in the range where the filler content is high, it is conceivable to introduce a non-linear term into the explanatory variables or utilize a non-linear prediction model.

[Other Embodiments]
It should be noted that the above-described embodiment is merely an example of the embodiment of the present invention, and the technical scope of the present invention should not be construed in a limited manner by the above-described embodiment. The present invention can be practiced in various forms without departing from its gist or its main features.

For example, in the above embodiment, the brand of the material is used as an identifier, but the identifier is not limited to the brand. For example, a catalog number, a production number, a lot number, or the like may be used as an identifier. Further, the manufacturing time (manufacturing year, etc.) may be combined with these to form an identifier.

Further, in the above embodiment, the elastic modulus of the polymer composite material is used as the objective variable, but the objective variable is not limited to the elastic modulus. For example, strength such as tensile strength, compressive strength and shear strength, hardness, elongation at break, impact resistance, wear resistance, flame retardancy, heat resistance, light resistance, weather resistance, acid resistance, alkali resistance, resistance to alkali. The objective variables may be solvent-based and color tone.

This application is an application claiming priority based on Japanese Application No. 2019-194086 filed on October 25, 2019, and the contents described in the claims, specification and drawings of the application are the present. Incorporated in the application.

The present invention is suitable as a property predictor for predicting the mechanical properties of a polymer composite material.

100 Characteristic prediction device 110 CPU
120 ROM
130 RAM
140 External storage device 150 Communication interface 160 Input device 170 Output device 210 Material information acquisition unit 220 Characteristic information output unit 230 Experimental data acquisition unit 240 Learning processing unit 250 Prediction model storage unit

Claims

A predictive model storage unit that stores a predictive model that outputs the mechanical properties of the polymer composite material to be manufactured in response to input of material information regarding the material of the polymer composite material.
A learning processing unit that performs learning processing on the prediction model using data including the material information and mechanical properties of the polymer composite material manufactured from the material as learning data.
Have,
The material information is information that makes it difficult to prepare information indicating physicochemical properties for all candidate materials.
As the material information, information including information that does not directly indicate the physicochemical properties and the compounding ratio of the material is used.
Characteristic predictor.
The Material Information Acquisition Department, which acquires material information regarding materials for polymer composite materials,
A characteristic information output unit that outputs the mechanical properties of the polymer composite material to be manufactured using the learned prediction model for the input material information.
Have,
The material information is information that makes it difficult to prepare information indicating physicochemical properties for all candidate materials.
The input material information includes information that does not directly indicate physicochemical properties and a compounding ratio of the material.
The prediction model is subjected to learning processing using learning data including the material information and the mechanical properties of the polymer composite material produced from the material.
Characteristic predictor.
The characteristic prediction device according to claim 1 or 2, wherein the polymer composite material is a composite material containing a polymer resin and an additive.
The characteristic prediction device according to claim 3, wherein the additive is a filler.
The characteristic prediction device according to any one of claims 1 to 4, wherein the information that does not directly indicate the physicochemical characteristics is information that indicates the brand of the material.
The characteristic prediction device according to any one of claims 1 to 5, wherein the mechanical property is an elastic modulus of the polymer composite material.
The characteristic prediction device according to any one of claims 1 to 6, wherein the prediction model is a prediction model using the following linear regression equation (1).

(In the formula (1), beta is a constant term, c i is the i-th partial regression coefficients, x i is the i-th explanatory variable, y is the dependent variable, N is the number of explanatory variables Is.)
The characteristic prediction device according to claim 7, wherein the partial regression coefficient is a coefficient obtained by using partial least squares regression.