Property prediction model generation method, property prediction model generation apparatus, property prediction method, property prediction apparatus, property prediction program, composition prediction method, composition prediction apparatus, and composition prediction program

Abstract

To provide a technique for extracting features based on combinations of materials, and efficiently predicting physical properties of rubber compound.SOLUTION: A property prediction model generation method includes: a co-occurrence matrix generation step of generating a co-occurrence matrix by using the number of rubber compounds as a co-occurrence relationship between materials, the rubber compounds having a combination of arbitrary two materials in a data mart; a material vector generation step of calculating the occurrence frequency and rarity of the materials in the data mart based on the co-occurrence matrix, to generate material vectors indicating features of the combinations of the materials; a composition vector generation step of generating composition vectors by dimensionally reducing the material vectors and linearly combining them using a composition amount; and a training step of training a learning-type arithmetic model configured to predict physical properties of rubber compound, based on the data mart in which at least one of components of the composition vectors is added to an explanatory variable.SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present invention relates to a method for generating a physical property prediction model in a rubber compound, a physical property prediction model generation device, a physical property prediction method, a physical property prediction device, a physical property prediction program, a formulation prediction method, a formulation prediction device, and a formulation prediction program.

Background Art

[0002] For example, a rubber compound used for a tire or the like mounted on a vehicle is a composite material in which a polymer, which is a rubber material as a main raw material, is added with a reinforcing agent and various chemicals. Rubber compounds are formulated based on various materials according to their uses, and those having different physical properties have been developed.

[0003] Patent Document 1 discloses a conventional method for designing a vulcanized rubber composition. In this design method, using a prediction module of a machine-learned computer, with the characteristic amount of the vulcanized rubber composition as an objective function and the combination of the formulations of raw materials as design variables, optimization is performed to extract a combination of formulations that realizes the set target value of the objective function. When performing the optimization, constraint conditions are imposed to create a combination of the formulations of raw materials as design variables and input them into the prediction module. The constraint condition is that it includes at least one or more combination candidates of formulations that are sets of the set raw materials. The combination candidate of formulations is a set of raw materials in which a combination of formulations is created by the formulation amounts of the raw materials in the learning input data.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] The design method of Patent Document 1 uses the characteristic quantity of the vulcanized rubber composition as the objective function and the combination of the raw material formulations as the design variables, and imposes constraints to optimize the combination of formulations. However, when the rubber compound consists of a large number of materials and the characteristic quantities are extracted based on the combinations and blending ratios of each material, the number of explanatory variables and the scale of the calculation increase, and it takes time for the learning of the calculation model and the prediction of physical properties, resulting in a problem of reduced efficiency. Also, if the number of explanatory variables is simply reduced to suppress the scale of the calculation model, there is a risk that the accuracy of the physical property prediction of the rubber compound will decrease.

[0006] The present invention has been made in view of such circumstances, and an object thereof is to provide a physical property prediction model generation method, a physical property prediction model generation device, a physical property prediction method, a physical property prediction device, a physical property prediction program, a formulation prediction method, a formulation prediction device, and a formulation prediction program that can extract characteristic quantities based on the combinations of each material and efficiently predict the physical properties of the rubber compound.

Means for Solving the Problems

[0007] One aspect of the present invention is a physical property prediction model generation method. The physical property prediction model generation method includes a data mart generation step of generating a data mart including the blending amounts of a plurality of materials used in the generation of a plurality of rubber compounds and the physical properties of the rubber compounds, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by using the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the characteristic quantity of the combination of each material reflecting the co-occurrence relationship, a blending vector generation step of dimensionally compressing the material vector and linearly combining it using the blending amount to generate a blending vector, and a learning step of learning a learning-type calculation model for predicting the physical properties of the rubber compound based on the data mart in which at least one of the components of the blending vector is added as an explanatory variable.

[0008] Another aspect of the present invention is a physical property prediction model generation device. The physical property prediction model generation device includes a data mart generation unit that generates a data mart including the compounding amounts of a plurality of materials used in the production of a plurality of rubber compounds and the physical properties of the rubber compounds, a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector linearly combined using the compounding amounts, and a learning processing unit that learns a learning-based arithmetic model for predicting the physical properties of rubber compounds based on the data mart with at least one of the components of the compounding vector added as an explanatory variable.

[0009] Another aspect of the present invention is a physical property prediction method. The physical property prediction method includes a data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the production of a rubber compound for which physical properties are to be predicted and a plurality of rubber compounds used for learning a learning-based arithmetic model for predicting the physical properties, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, and a physical property prediction step of adding at least one of the components of the compounding vector in the rubber compound for which physical properties are to be predicted as an explanatory variable and predicting the physical properties by the arithmetic model. It comprises.

[0010] Another aspect of the present invention is a physical property prediction device. The physical property prediction device includes a data mart generation unit that generates a data mart including a rubber compound for predicting physical properties and compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-based arithmetic model for predicting physical properties, a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship by using, as the co-occurrence relationship between materials, the number of rubber compounds in which any combination of two materials in the data mart occurs, a material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, a compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector linearly combined using the compounding amounts, and a physical property prediction processing unit that adds at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicts physical properties by the arithmetic model.

[0011] Another aspect of the present invention is a physical property prediction program. The physical property prediction program causes a computer to execute a data mart generation step of generating a data mart including a rubber compound for predicting physical properties and compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-based arithmetic model for predicting physical properties, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by using, as the co-occurrence relationship between materials, the number of rubber compounds in which any combination of two materials in the data mart occurs, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, and a physical property prediction step of adding at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicting physical properties by the arithmetic model.

[0012] Another aspect of the present invention is a compounding prediction method. The compounding prediction method includes a data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-type calculation model for predicting physical properties, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by using the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, and a compounding prediction step of predicting the compounding of a rubber compound by an optimization method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound, using the calculation model learned by adding at least one of the components of the compounding vector as an explanatory variable.

[0013] Another aspect of the present invention is a compounding prediction apparatus. The compounding prediction apparatus includes a data mart generation unit that generates a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-type calculation model for predicting physical properties, a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship by using the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector linearly combined using the compounding amounts, and a compounding prediction processing unit that predicts the compounding of a rubber compound by an optimization method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound, using the calculation model learned by adding at least one of the components of the compounding vector as an explanatory variable.

[0014] Another aspect of the present invention is a compounding prediction program. The compounding prediction program includes a data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-based calculation model for predicting physical properties, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, and a compounding prediction step of predicting the compounding of a rubber compound by an optimization method based on a target physical property and constraint conditions regarding the compounding amounts of materials in the rubber compound using the calculation model learned with at least one of the components of the compounding vector added as an explanatory variable. The above steps are executed by a computer.

Advantages of the Invention

[0015] According to the present invention, feature amounts based on combinations of each material can be extracted, and the physical properties of rubber compounds can be efficiently predicted.

Brief Description of the Drawings

[0016]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Embodiments for Carrying Out the Invention

[0017] Hereinafter, the present invention will be described based on preferred embodiments with reference to FIGS. 1 to 13. The same or equivalent components and members shown in each drawing are denoted by the same reference numerals, and redundant descriptions will be omitted as appropriate. Also, the dimensions of the members in each drawing are shown enlarged or reduced as appropriate for easy understanding. Further, some of the members that are not important for explaining the embodiments in each drawing are omitted.

[0018] (Embodiment 1) FIG. 1 is a block diagram showing the functional configuration of a physical property prediction model generation device 100 according to Embodiment 1. The physical property prediction model generation device 100 includes a storage unit 10, an operation unit 20, a display unit 30, and an arithmetic processing unit 40, and generates an arithmetic model for predicting the physical properties of a rubber compound used in, for example, a tire or the like. The physical property prediction model generation device 100 is an information processing device such as a PC (personal computer). Each unit in the physical property prediction model generation device 100 can be realized, in terms of hardware, by an electronic processing circuit composed of electronic elements including a computer CPU and mechanical parts, etc., and can be realized, in terms of software, by a computer program or the like. Here, however, functional blocks realized by their cooperation are depicted. Therefore, it is understood by those skilled in the art that these functional blocks can be realized in various forms by combinations of hardware and software.

[0019] The storage unit 10 is a storage device composed of, for example, an SSD (Solid State Drive), a hard disk, a CD-ROM, a DVD, or the like. The storage unit 10 stores a data mart 11 used for predicting the physical properties of a rubber compound, a computer program executed by the arithmetic processing unit 40, and data used for the execution of the computer program, etc.

[0020] The data mart 11 includes formulation data 12, physical property data 13, raw material data 14, and processing condition data 15 corresponding to each of a plurality of rubber compounds. In the data mart 11, for example, identification information is given to each of a plurality of rubber compounds by means of a serial number or the like, and the formulation, physical properties, raw materials, and processing conditions corresponding to each identification information are included in the formulation data 12, physical property data 13, raw material data 14, and processing condition data 15.

[0021] The compounding data 12 is a data group indicating the compounding amounts of each material in a plurality of rubber compounds. The compounding data 12 includes the names and compounding amounts of rubber materials, reinforcing agents, and various chemicals as the main components corresponding to each of the plurality of rubber compounds. The rubber materials are, for example, natural rubber (NR), butadiene rubber (BR), styrene-butadiene rubber (SBR), isobutylene-isoprene-styrene rubber (IIR), etc. The reinforcing materials are, for example, carbon black, silica, etc. The various chemicals are, for example, sulfur, wax, etc. Note that the compounding amounts are represented by the volume, weight, and ratios of the materials.

[0022] The physical property data 13 is a data group such as, for example, hardness Hs, tensile stress S, loss tangent tanδ, abrasion resistance, resilience, etc. for each of the plurality of rubber compounds. The raw material data 14 is a data group such as property values and categorical variables of each material used for each of the plurality of rubber compounds. The property values are, for example, carbon surface area, etc. The categorical variables are, for example, ASTM grades of carbon, etc.

[0023] The processing condition data 15 is a data group including conditions such as time and temperature in each processing process such as kneading, molding, and vulcanization during the production of each of the plurality of rubber compounds. The data mart 11 may acquire the compounding data 12, physical property data 13, raw material data 14, and processing condition data 15 of the plurality of rubber compounds from the external device 90. The external device 90 is an information processing device such as a PC, for example, and is assumed to have a storage device that stores the compounding data 12, physical property data 13, raw material data 14, and processing condition data 15 of the plurality of rubber compounds.

[0024] The operation unit 20 has an operable input device such as a touch panel, switch, keyboard, and mouse device, for example, and receives the user's operation input. The operation unit 20 receives the user's operation input in the generation of the physical property prediction model of the rubber compound, physical property prediction, and compounding prediction.

[0025] The display unit 30 has a display device such as a liquid crystal display, and displays a screen for receiving various data and user operation inputs in the process of generating a physical property prediction model of a rubber compound. Further, the display unit 30 displays a screen for receiving various data and user operation inputs in the physical property prediction and compounding prediction of the rubber compound.

[0026] The arithmetic processing unit 40 includes a data mart generation unit 41, a natural language processing unit 42, a physical property prediction processing unit 43, and a learning processing unit 44. The arithmetic processing unit 40 is an electronic circuit that executes arithmetic processing such as a CPU, and functions by reading and executing a computer program and data stored in the storage unit 10. The data mart generation unit 41, the natural language processing unit 42, the physical property prediction processing unit 43, and the learning processing unit 44 in the arithmetic processing unit 40 may be constructed as a plurality of program modules formed by a computer program.

[0027] The data mart generation unit 41 acquires compounding data 12, physical property data 13, raw material data 14, and processing condition data 15 of a plurality of rubber compounds from an external device 90 and generates a data mart 11. Further, the data mart generation unit 41 may set all or part of a database of compounding, physical properties, property values of raw materials, and category variables, and processing conditions of a plurality of rubber compounds stored in the storage unit 10 in advance as the data mart 11 used for generating a physical property prediction model. The data mart generation unit 41 may receive a user operation input for designating a rubber compound to be included in the data mart 11 by the operation unit 20 and generate the data mart 11.

[0028] The natural language processing unit 42 includes a co-occurrence matrix generation unit 42a, a material vector generation unit 42b, and a compounding vector generation unit 42c, and extracts feature amounts of combinations of materials in the compounding data 12 using techniques such as co-occurrence matrix generation in natural language processing.

[0029] The co-occurrence matrix generation unit 42a uses the number of rubber compounds in which any two combinations of materials in the data mart 11 occur as the co-occurrence relationship between the materials, and generates a co-occurrence matrix representing the co-occurrence relationship. FIG. 2(a) is a chart showing an example of formulation in a rubber compound, and FIG. 2(b) is a chart showing an example of the co-occurrence matrix to be generated. FIGS. 2(a) and 2(b) show simple formulation examples for the sake of brevity of explanation.

[0030] In FIG. 2(a), for each of the rubber compounds G1, G2, and G3, the amounts of materials A, B, and C formulated are shown. In FIG. 2(b), for materials A, B, and C, the number of rubber compounds in which any two combinations of materials occur is shown. For example, the combination of material A and material B occurs in two rubber compounds, G2 and G3, and the combination of material A and material C occurs in three rubber compounds, G1, G2, and G3.

[0031] FIG. 2(c) is a chart showing an example of a co-occurrence matrix using expressions by natural language processing. As shown in FIG. 2(c), each row in FIG. 2(b) is expressed as a material A document, a material B document, and a material C document, and each column is expressed as a material A word, a material B word, and a material C word. The co-occurrence matrix generation unit 42a generates a co-occurrence matrix based on the combination of materials for each material formulated in a plurality of rubber compounds according to the concept in natural language processing. The co-occurrence matrix generated in this way is also called a pairwise co-occurrence matrix.

[0032] The material vector generation unit 42b calculates the term frequency tf and inverse document frequency idf of each material in the data mart 11 based on the co-occurrence matrix generated by the co-occurrence matrix generation unit 42a, and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship. FIG. 3(a) is a chart showing an example of the term frequency based on the co-occurrence matrix, FIG. 3(b) is a chart showing an example of the inverse document frequency based on the co-occurrence matrix, and FIG. 3(c) is a chart showing an example of the generated material vector.

[0033] The material vector generation unit 42b calculates the term frequency tf based on the following formula, where Ei is the number of occurrences of word i in document d and Et is the total number of occurrences of all words in document d. tf = Ei / Et ·····(1)

[0034] As shown in Fig. 3(a), for example, in the document of Material A, the total number of occurrences of all words Et is 5, the number of occurrences of the word of Material B Ei is 2, and based on Equation (1), the term frequency tf of the word of Material B is 0.4. For the word of Material A in the document of Material A, since it does not mean a combination of materials, the term frequency tf is treated as 0. Similarly, the term frequency tf of each material word in each material document is calculated.

[0035] The material vector generation unit 42b calculates the inverse document frequency idf based on the following formula, where Td is the total number of documents and Ti is the number of documents containing word i. idf = 1 + Ln(Td / Ti) ·····(2)

[0036] As shown in Fig. 3(b), the total number of documents Td is 3. For example, the number of documents Ti containing the word of Material C is 2, and the inverse document frequency idf is calculated by applying it to Equation (2). Similarly, the inverse document frequency idf of each material is calculated.

[0037] The material vector generation unit 42b multiplies the term frequency tf by the inverse document frequency idf of each material to generate a material vector indicating a feature amount reflecting the co-occurrence relationship. In Fig. 3(c), for example, for the word of Material A, the inverse document frequency idf of the word of Material A is multiplied by the values of the term frequency tf in the documents of Material B and Material C. The material vector is a vector having the words of Material A, the words of Material B, and the words of Material C as components for the document of Material A.

[0038] The compounding vector generation unit 42c performs dimensional compression on the material vectors generated by the material vector generation unit 42b using principal component analysis (PCA), and further generates a compounding vector by linearly combining them using the compounding ratios. By performing dimensional compression by principal component analysis, the compounding vector generation unit 42c synthesizes several principal components that best represent the overall variation with little correlation as variables. In addition to principal component analysis (PCA), known techniques such as latent semantic analysis (LSA), linear discriminant analysis (LDA), and t-distribution type stochastic neighbor embedding method (t-SNE) may be used as the dimensional compression method in the compounding vector generation unit 42c.

[0039] FIG. 4 is a chart showing an example of a material vector dimensionally compressed by principal component analysis. The compounding vector generation unit 42c dimensionally compresses the material vectors of three components corresponding to materials A, B, and C shown in FIG. 3(c) into material vectors represented by two principal components, dimension U1 and dimension U2, shown in FIG. 4.

[0040] FIG. 5 is a chart showing an example of a compounding vector generated by the compounding vector generation unit 42c. The compounding vector generation unit 42c linearly combines the material vectors indicating the feature amounts based on the information of the compounding amounts by adding the dimensionally compressed material vectors with weights according to the compounding amounts of the materials (for example, volume fraction) to generate a compounding vector. In FIG. 5, X, XX, and XXX represent the values in each dimension of the compounding vector and are calculated as specific numerical values.

[0041] The physical property prediction processing unit 43 predicts the physical properties by the calculation model 43a, in addition to using at least one of the components of the compounding vector generated by the compounding vector generation unit 42c as an explanatory variable that is an input to the calculation model 43a.

[0042] FIG. 6 is a chart for explaining explanatory variables with the blending vector added. In rubber blends G1, G2, and G3, the blending amounts of materials A, B, and C are included in data mart 11 as blending data 12, and are basic explanatory variables in predicting physical properties. For example, if it is empirically found that the influence of material C on the physical properties to be predicted is small and not important, the blending amount of material C may be excluded from the explanatory variables.

[0043] The physical property prediction processing unit 43 adds a blending vector composed of dimensions U1 and U2 as a feature amount of rubber blends G1, G2, and G3 to the explanatory variables. Similar to the exclusion of the above-described explanatory variables related to materials, among dimensions U1 and U2, for example, if it is found that the influence of dimension U2 on the physical properties to be predicted is small and not important, dimension U2 may be excluded from the explanatory variables.

[0044] Generally, in dimensionality reduction by principal component analysis, each dimension synthesized as a variable in sequence has a lower degree of variation as the order becomes higher, and it is considered that the influence on physical properties becomes smaller. Here, the higher order means that when principal components are synthesized like dimensions U1, U2, U3, and U4, the dimension obtained in the latter stage is called the higher order than the dimension obtained in the former stage. For example, dimension U4 is considered to be of a higher order than dimensions U1, U2, and U3. However, as will be described later, it is also inferred from the examples that higher-order dimensions are important.

[0045] In the physical property prediction processing unit 43, the explanatory variables that are input data to the arithmetic model 43a are blending data 12, the blending vector generated by the natural language processing unit 42, raw material data 14, and processing condition data 15. The arithmetic model 43a is a learning model of the random forest type, which is an ensemble learning algorithm using a decision tree.

[0046] The learning processing unit 44 compares the estimated values of the physical properties of the rubber compound as the output data from the calculation model 43a with the known physical property data 13 as the teacher data, and learns the calculation model 43a. Note that the calculation model 43a may be, for example, a generalized linear regression model, a DNN (Deep Neural Network) model, a gradient boosting type decision tree model, or the like. Also, for the verification of the calculation model 43a, known verification methods such as random data sampling and cross-validation can be used.

[0047] Next, the operation of the physical property prediction model generation device 100 will be described. FIG. 7 is a flowchart showing the procedure of the generation process of the calculation model 43a. The data mart generation unit 41 of the arithmetic processing unit 40 generates a data mart 11 for learning the calculation model 43a (S1). The data mart 11 includes at least the compounding data 12 and the physical property data 13. Also, the data mart 11 may include at least one of the raw material data 14 and the processing condition data 15 in addition to the compounding data 12 and the physical property data 13.

[0048] The co-occurrence matrix generation unit 42a generates, based on the compounding data 12 of the data mart 11, the number of rubber compounds in which a combination of two materials occurs as the co-occurrence relationship between the materials, and generates a co-occurrence matrix representing the co-occurrence relationship (S2). The material vector generation unit 42b calculates the term frequency tf and the inverse document frequency idf of each material in the data mart 11 based on the co-occurrence matrix generated in step S2 (S3). The material vector generation unit 42b multiplies the term frequency tf by the inverse document frequency idf of each material to generate a material vector indicating a feature amount reflecting the co-occurrence relationship (S4).

[0049] The compounding vector generation unit 42c performs dimensionality reduction on the material vector generated in step S4 using, for example, principal component analysis (S5). The compounding vector generation unit 42c linearly combines the material vectors indicating the feature amounts based on the information on the compounding amounts by adding the material vectors dimensionally reduced in step S5 with weights according to the compounding amounts of the materials to generate a compounding vector (S6).

[0050] The physical property prediction processing unit 43 adds at least one of the components of the compounding vector generated in step S6 to the explanatory variable that is the input to the calculation model 43a (S7). The learning processing unit 44 compares the estimated value of the physical property of the rubber compound calculated by the calculation model 43a with the known physical property data 13 as teacher data, learns the calculation model 43a (S8), and ends the process.

[0051] FIG. 8 is a chart showing the importance of the compounding vector in the example in which the calculation model 43a is learned. In the example shown in FIG. 8, 207 types of materials are used in a plurality of rubber compounds, and 10 from dimension U1 to U10 as components of the compounding vector generated by the compounding vector generation unit 42c are added to the explanatory variables. The number of explanatory variables is 217, which is the sum of the compounding amounts of 207 types of materials and 10 from dimension U1 to U10 in the compounding vector. Further, the calculation model 43a in the example estimates each physical property of the peak temperature T of the loss tangent Tanδ obtained by the dynamic viscoelasticity test, the hardness Hs (instantaneous value at 23° C.), the tensile stress S (300% tensile stress), the wear amount F obtained by the Lambourne wear tester, and the rebound resilience rate R (value at 23° C.).

[0052] In FIG. 8, with respect to the peak temperature T of the loss tangent Tanδ, the dimensions U1 to U10, which are the components of the compounding vector, are in the range from the 7th to the 29th in terms of importance, the highest rank is the 7th, and the average rank is the 17th. The highest rank of 7th is at the upper 3.2% position of all 217 types of explanatory variables, and the average rank of 17th is at the upper 7.6% position of all 217 types of explanatory variables. In each physical property estimated by the calculation model 43a, the dimensions U1 to U10, which are the components of the compounding vector, have the highest rank within the upper 5% and the average rank within the upper 10%, indicating that they are important explanatory variables for estimating the physical properties of the rubber compound.

[0053] The method for generating a physical property prediction model according to this embodiment generates a data mart 11 including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds and the physical properties of the rubber compounds in the data mart generation step. The method for generating a physical property prediction model generates a co-occurrence matrix representing a co-occurrence relationship by using, as the co-occurrence relationship between materials, the number of rubber compounds in which any two-material combination in the data mart 11 occurs in the co-occurrence matrix generation step. The method for generating a physical property prediction model calculates the term frequency tf and inverse document frequency idf of each material in the data mart 11 based on the co-occurrence matrix in the material vector generation step, and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship.

[0054] The method for generating a physical property prediction model generates a compounding vector obtained by linearly combining the dimensionally compressed material vectors using the compounding amounts in the compounding vector generation step. The method for generating a physical property prediction model learns a learning-type arithmetic model 43a for predicting the physical properties of a rubber compound based on the data mart 11 in which at least one of the components of the compounding vector is added as an explanatory variable in the learning step. According to this method for generating a physical property prediction model, feature amounts based on each combination of materials can be extracted, and the physical properties of rubber compounds can be predicted efficiently.

[0055] In the method for generating a physical property prediction model, the data mart 11 may include raw material data 14 of the materials in the rubber compound, that is, data such as the property values and categorical variables of each material. Thereby, the method for generating a physical property prediction model can learn the arithmetic model 43a reflecting the influence of the raw material data 14.

[0056] In the method for generating a physical property prediction model, the data of the property values in the raw material data 14 includes data calculated based on the compounding amounts of a plurality of materials. FIG. 9(a) is a chart showing an example of the raw material data 14, and FIG. 9(b) is a chart showing examples of the property value data and categorical variables in the data mart 11. Regarding FIG. 9(a), for carbons A and B as materials, it has carbon surface area and the like as data of property values, and has ASTM grade of carbon and the like as categorical variables.

[0057] As shown in FIG. 9(b), for the rubber compound, the carbon surface area is calculated as a weighted average value based on the compounding amounts of Carbon A and Carbon B. The categorical variable represents whether it corresponds by 0 and 1. For example, the carbon surface area in the rubber compound G1 is calculated as (125×50 + 70×25) / (50 + 25) = 106. Thus, the physical property prediction model generation method can synthesize the property values common to a plurality of materials and use them as explanatory variables of the calculation model 43a.

[0058] In the physical property prediction model generation method, the data mart 11 may include the processing condition data 15 in the rubber compound. Thus, the physical property prediction model generation method can learn the calculation model 43a reflecting the influence of the processing condition data 15. In the physical property prediction model generation method, the property values and categorical variables of the raw material data 14, and the processing conditions in the processing condition data 15 are included in the data mart 11, and in addition to the explanatory variables of the calculation model 43a, the physical properties can be predicted.

[0059] In the physical property prediction model generation method, the compounding vector generation step compresses the dimensionality of the material vector by principal component analysis. Thus, the physical property prediction model generation method can adopt, as explanatory variables, a part of a plurality of dimensions that are components of the material vector obtained by principal component analysis, and not adopt the remaining dimensions as explanatory variables, and can suppress an increase in the amount of calculation.

[0060] In the physical property prediction model generation method, the calculation model 43a is any one of, for example, a random forest model, a gradient boosting decision tree model, and a generalized linear regression model. Thus, the physical property prediction model generation method can construct a calculation model utilizing the characteristics of each model.

[0061] (Embodiment 2) FIG. 10 is a block diagram showing the functional configuration of the physical property prediction apparatus 110 according to Embodiment 2. The physical property prediction apparatus 110 has a configuration obtained by removing the learning processing unit 44 from the physical property prediction model generation apparatus 100 in Embodiment 1. Further, as the arithmetic model 43a in the physical property prediction processing unit 43 of the physical property prediction apparatus 110, a learned arithmetic model by the physical property prediction model generation apparatus 100 is used.

[0062] The data mart generation unit 41 of the physical property prediction apparatus 110 generates a data mart 11 by adding a new rubber compound Gn for predicting physical properties to a plurality of rubber compounds used for learning the arithmetic model 43a. Since the data mart 11 generated here does not learn the arithmetic model 43a, it suffices to have the compounding data 12, the raw material data 14, and the processing condition data 15, and it does not have to include the physical property data 13.

[0063] Similar to Embodiment 1, the natural language processing unit 42 of the physical property prediction apparatus 110 uses methods such as co-occurrence matrix generation in natural language processing to extract feature amounts of combinations of materials in the compounding data 12. The natural language processing unit 42 includes a co-occurrence matrix generation unit 42a, a material vector generation unit 42b, and a compounding vector generation unit 42c, and each of these units functions in the same manner as in Embodiment 1.

[0064] Specifically, the co-occurrence matrix generation unit 42a of the natural language processing unit 42 uses the number of rubber compounds in which an arbitrary combination of two materials in the data mart 11 occurs as the co-occurrence relationship between the materials, and generates a co-occurrence matrix representing the co-occurrence relationship. The material vector generation unit 42b calculates the term frequency tf and the inverse document frequency idf of each material in the data mart 11 based on the co-occurrence matrix generated by the co-occurrence matrix generation unit 42a, and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship. The compounding vector generation unit 42c performs dimensionality reduction on the material vector generated by the material vector generation unit 42b using principal component analysis or the like, and further generates a compounding vector obtained by linearly combining using the compounding ratio.

[0065] The physical property prediction processing unit 43 predicts the physical properties of a new rubber formulation Gn by using, as explanatory variables for input to the calculation model 43a, the compounding vector generated for the new rubber formulation Gn, in addition to the compounding data and, referring to the example in the above-described Embodiment 1, the components of the compounding vector from dimension 1 to dimension 10 as explanatory variables, and predicts the physical properties by using the learned calculation model 43a.

[0066] Further, in the physical property prediction processing unit 43, the explanatory variables that are input data to the calculation model 43a may include the raw material data 14 and the processing condition data 15 in addition to the compounding data 12 and the compounding vector generated by the natural language processing unit 42.

[0067] Next, the operation of the physical property prediction device 110 will be described. FIG. 11 is a flowchart showing the procedure of the physical property prediction process. The data mart generation unit 41 of the arithmetic processing unit 40 generates a data mart 11 including a new rubber formulation Gn (S11). The data mart 11 adds a new rubber formulation Gn for predicting the physical properties to the plurality of rubber formulations used for learning the calculation model 43a, and includes the compounding data 12 of these rubber formulations. Further, the data mart 11 may include at least one of the raw material data 14 and the processing condition data 15 in addition to the compounding data 12.

[0068] The processing from step S12 to step S16 is equivalent to the processing from step S2 to step S6 in FIG. 7, and the description thereof is omitted for simplicity. The physical property prediction processing unit 43 adds at least one of the components of the compounding vector generated in step S16 as an explanatory variable for the new rubber formulation Gn (S17). The physical property prediction processing unit 43 inputs the explanatory variable for the new rubber formulation Gn to the calculation model 43a, estimates the physical properties (S18), and ends the processing.

[0069] In the processes of steps S12 to S15, the same natural language processing as the generation process of the arithmetic model 43a in Embodiment 1 is performed again on the data mart 11 including the compounding data of the rubber compound Gn for which physical properties are to be predicted, and the physical properties of the rubber compound Gn are predicted in the processes of steps S16 to S18. On the other hand, as the material vector and the dimensionally compressed material vector, those already created in the generation process of the arithmetic model 43a in Embodiment 1 may be used. In this case, in step S16, since a compounding vector for the rubber compound Gn for which physical properties are to be predicted can be created by linearly combining material vectors based on the compounding data (material ratio) of the rubber compound Gn for which physical properties are to be predicted, steps S12 to S15 may be omitted.

[0070] The physical property prediction method in this embodiment generates a data mart 11 including the compounding amounts of a plurality of materials used in the generation of a rubber compound Gn for which physical properties are newly predicted and a plurality of rubber compounds used for learning a learning-type arithmetic model for predicting physical properties in the data mart generation step. The physical property prediction method generates a co-occurrence matrix representing a co-occurrence relationship by using, as the co-occurrence relationship between materials, the number of rubber compounds in which an arbitrary combination of two materials in the data mart 11 occurs in the co-occurrence matrix generation step. The physical property prediction method calculates the term frequency tf and the inverse document frequency idf of each material in the data mart 11 based on the co-occurrence matrix in the material vector generation step, and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship.

[0071] The physical property prediction method generates a compounding vector obtained by dimensionally compressing a material vector and linearly combining it using the compounding amount in the compounding vector generation step. The physical property prediction method predicts physical properties by adding at least one of the components of the compounding vector in the rubber compound Gn for which physical properties are to be predicted as an explanatory variable and using an arithmetic model. According to this physical property prediction method, feature amounts based on each combination of materials used in the generation of a plurality of rubber compounds can be extracted, and the physical properties of the rubber compound can be efficiently predicted.

[0072] There are several cases to consider depending on the range of explanatory variables of the calculation model 43a that outputs the physical properties of the rubber compound in the physical property prediction method. The same applies to the method for generating the calculation model 43a in Embodiment 1. In any case, in Embodiments 1 and 2, the explanatory variables of the calculation model 43a include at least one of the components of the compounding vector generated in the compounding vector generation step. The first case is to use the compounding of the rubber compound as the explanatory variable of the calculation model 43a. The second case is to use the compounding of the rubber compound and the processing conditions as the explanatory variables of the calculation model 43a. The third case is to use the compounding of the rubber compound and the property values of the raw materials as the explanatory variables of the calculation model 43a. The fourth case is to use the property values of the raw materials as the explanatory variables of the calculation model 43a. The fifth case is to use the property values of the raw materials and the processing conditions as the explanatory variables of the calculation model 43a. Note that instead of the property values of the raw materials, the categorical variables of the raw materials may be used as the explanatory variables of the calculation model 43a. Also, in addition to the property values of the raw materials, the categorical variables of the raw materials may be included in the explanatory variables of the calculation model 43a.

[0073] (Embodiment 3) FIG. 12 is a block diagram showing the functional configuration of the compounding prediction device 120 according to Embodiment 3. The compounding prediction device 120 includes a compounding prediction processing unit 45. The compounding prediction processing unit 45 has a calculation model 43a that has been learned by the physical property prediction model generation device 100.

[0074] The data mart generation unit 41 in the compounding prediction device 120 generates a data mart 11 including a plurality of rubber compounds used for learning the calculation model 43a. Since the data mart 11 generated here does not learn the calculation model 43a, it only needs to have the compounding data 12, the raw material data 14, and the processing condition data 15, and does not necessarily include the physical property data 13. The natural language processing unit 42 uses methods such as co-occurrence matrix generation in natural language processing to extract the feature amounts of the material combinations in the compounding data 12, similar to Embodiments 1 and 2. The natural language processing unit 42 includes a co-occurrence matrix generation unit 42a, a material vector generation unit 42b, and a compounding vector generation unit 42c, and each of these units functions in the same manner as in Embodiments 1 and 2.

[0075] The compounding prediction processing unit 45 uses the calculation model 43a to predict, as an inverse problem, the compounding of the rubber compound by means of a genetic algorithm method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound. The constraint conditions may include the raw material data 14 and the processing condition data 15. Further, the compounding prediction processing unit 45 may predict the compounding of the rubber compound as an inverse problem using a known optimization method, for example, using the gradient descent method, grid search, or the like.

[0076] The compounding prediction processing unit 45 may add the compounding data, raw material data, and processing condition data of the rubber compound assumed in the prediction process to the data mart 11, and the natural language processing unit 42 may extract the feature amounts of the material combinations and reflect them in the compounding prediction. Further, the compounding prediction processing unit 45 may be configured to predict, for the rubber compound having the target physical properties, at least any one of the property values and category variables of the raw materials and the processing conditions in addition to the compounding (material ratio).

[0077] FIG. 13 is a flowchart showing the procedure of the compounding prediction processing. The compounding prediction processing unit 45 acquires the physical properties of the target rubber compound (S21), and acquires the constraint conditions of the rubber compound (S22). The compounding prediction processing unit 45 uses the learned calculation model 43a to predict, as an inverse problem, the compounding of the rubber compound by means of an optimization method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound (S23), and ends the processing.

[0078] The compounding prediction method in the present embodiment generates a data mart 11 including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning the learning-type calculation model 43a for predicting physical properties in the data mart generation step. The compounding prediction method generates a co-occurrence matrix representing the co-occurrence relationship by using, as the co-occurrence relationship between materials, the number of rubber compounds in which an arbitrary combination of two materials in the data mart 11 occurs in the co-occurrence matrix generation step.

[0079] The compounding prediction method calculates the frequency of occurrence and rarity of each material in the data mart 11 of each material based on the co-occurrence matrix in the material vector generation step, and generates a material vector indicating the feature amount of the combination of each material reflecting the co-occurrence relationship. The compounding prediction method generates a compounding vector by dimensionally compressing the material vector and linearly combining it using the compounding amount in the compounding vector generation step. The compounding prediction method uses an arithmetic model 43a in which at least one of the components of the compounding vector is added to the explanatory variable and learned in the compounding prediction step, and predicts the compounding of the rubber compound by an optimization method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound. According to this compounding prediction method, it is possible to extract the feature amount based on the combination of each material and predict the compounding of the rubber compound that can obtain the target physical properties.

[0080] The compounding prediction processing unit 45 may predict a rubber compound that satisfies the physical properties and the compounding unit price by adding the target compounding unit price in addition to the target physical properties. In this case, the arithmetic model 43a outputs the compounding unit price, and the compounding unit price as physical property data is learned in advance as teacher data. For example, by including the unit price of each material in the raw material data 14 and taking the weighted average using the compounding amount of the material and the unit price of the material as in FIG. 9(b), the compounding unit price is calculated and registered in the physical property data 13, so that the compounding unit price can be treated in the same way as the physical properties.

[0081] Generalizing the technical idea embodied by the above embodiment, it can be said that the technical idea described in the following items is included.

[0082] The first item is a data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds and the physical properties of the rubber compounds, and a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and linearly combining it using the compounding amount to generate a compounding vector, and a learning step of learning a learning-based arithmetic model for predicting the physical properties of a rubber compound based on the data mart with at least one of the components of the compounding vector added as an explanatory variable. It is a method for generating a physical property prediction model.

[0083] The second item is the method for generating a physical property prediction model according to the first item, wherein the data mart includes data on the property values of materials in the rubber compound.

[0084] The third item is the method for generating a physical property prediction model according to the second item, wherein the data on the property values includes data calculated based on the compounding amounts of a plurality of materials.

[0085] The fourth item is the method for generating a physical property prediction model according to any one of the first to third items, wherein the data mart includes data on processing conditions in the rubber compound.

[0086] The fifth item is the method for generating a physical property prediction model according to any one of the first to fourth items, wherein the compounding vector generation step dimensionally compresses the material vector by principal component analysis.

[0087] The sixth item is the method for generating a physical property prediction model according to any one of the first to fifth items, wherein the arithmetic model is any one of a random forest model, a gradient boosting decision tree model, and a generalized linear regression model.

[0088] The seventh item is a data mart generation unit that generates a data mart including the compounding amounts of multiple materials used in generating multiple rubber compounds and the physical properties of the rubber compounds; a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship between materials, with the number of rubber compounds in which a combination of any two materials occurs in the data mart being the co-occurrence relationship between the materials; a material vector generation unit that calculates the occurrence frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship; a compounding vector generation unit that compresses the dimensions of the material vector and generates a compounding vector linearly combined using the compounding amounts; and a learning processing unit that trains a learning-type computational model that predicts the physical properties of a rubber compound based on the data mart to which at least one component of the compounding vector is added as an explanatory variable.

[0089] The eighth item is a data mart generation step for generating a data mart including the rubber compound for which the physical properties are to be predicted, and the compounding amounts of a plurality of materials used in generating a plurality of rubber compounds used in learning a learning-type computational model for predicting the physical properties; a co-occurrence matrix generation step for generating a co-occurrence matrix representing the co-occurrence relationship between materials, with the number of rubber compounds in which a combination of any two materials occurs in the data mart being regarded as a co-occurrence relationship between materials; a material vector generation step for calculating the frequency and rarity of occurrence of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship; a compounding vector generation step for compressing the dimensions of the material vector and generating a compounding vector linearly combined using the compounding amounts; and a property prediction step for adding at least one component of the compounding vector in the rubber compound for which the physical properties are to be predicted to an explanatory variable, and predicting the physical properties using the computational model.

[0090] Item 9 is a data mart generation unit that generates a data mart including a rubber compound for predicting physical properties and the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for training a learning-based arithmetic model for predicting physical properties, a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector linearly combined using the compounding amounts, and a physical property prediction processing unit that adds at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicts physical properties by the arithmetic model.

[0091] Item 10 is a physical property prediction program that causes a computer to execute a data mart generation step of generating a data mart including a rubber compound for predicting physical properties and the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for training a learning-based arithmetic model for predicting physical properties, a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship by taking the number of rubber compounds in which any two-material combination in the data mart occurs as the co-occurrence relationship between the materials, a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, a compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, and a physical property prediction step of adding at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicting physical properties by the arithmetic model.

[0092] The 11th item is a data mart generation step of generating a data mart including the blending amounts of a plurality of materials used in the generation of a plurality of rubber formulations used for learning a learning-type arithmetic model for predicting physical properties, and the number of rubber formulations in which any two-material combination in the data mart occurs is used as a co-occurrence relationship between the materials, and a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, and calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, and a blending vector generation step of dimensionally compressing the material vector and linearly combining it using the blending amount, and using the arithmetic model in which at least one of the components of the blending vector is added as an explanatory variable and learned, and predicting the blending of the rubber formulation by an optimization method based on the target physical properties and the constraint conditions regarding the blending amounts of the materials in the rubber formulation.

[0093] The 12th item is a data mart generation unit that generates a data mart including the blending amounts of a plurality of materials used in the generation of a plurality of rubber formulations used for learning a learning-type arithmetic model for predicting physical properties, and a co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship by using the number of rubber formulations in which any two-material combination in the data mart occurs as a co-occurrence relationship between the materials, and a material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship, and a blending vector generation unit that dimensionally compresses the material vector and linearly combines it using the blending amount, and a blending prediction processing unit that uses the arithmetic model in which at least one of the components of the blending vector is added as an explanatory variable and learned, and predicts the blending of the rubber formulation by an optimization method based on the target physical properties and the constraint conditions regarding the blending amounts of the materials in the rubber formulation.

[0094] The 13th item is a compound prediction program that causes a computer to execute: a data mart generation step of generating a data mart including compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-type arithmetic model for predicting physical properties; a co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any two-material combination in the data mart occurs is defined as the co-occurrence relationship between the materials; a material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship; a compound vector generation step of dimensionally compressing the material vector and linearly combining it using the compounding amounts to generate a compound vector; and a compound prediction step of predicting the compounding of a rubber compound by an optimization method based on the target physical properties and the constraint conditions regarding the compounding amounts of the materials in the rubber compound, using the arithmetic model learned by adding at least one of the components of the compound vector as an explanatory variable.

[0095] As described above, the embodiments of the present invention have been described based on the embodiments. These embodiments are examples, and it is understood by those skilled in the art that various modifications and changes are possible within the scope of the claims of the present invention, and such modifications and changes are also within the scope of the claims of the present invention. Therefore, the descriptions and drawings in this specification should be treated as illustrative rather than restrictive.

Explanation of Signs

[0096] 11 Data mart, 41 Data mart generation unit, 42a Co-occurrence matrix generation unit, 42b Material vector generation unit, 42c Compound vector generation unit, 43 Physical property prediction processing unit, 43a Arithmetic model, 44 Learning processing unit, 45 Compound prediction processing unit, 100 Physical property prediction model generation device, 110 Physical property prediction device, 120 Compound prediction device.

Claims

1. A data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the production of a plurality of rubber compounds and the physical properties of the rubber compounds, A co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any two-material combination in the data mart occurs is defined as the co-occurrence relationship between the materials, A material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, A compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amounts, A learning step of learning a learning-based arithmetic model for predicting the physical properties of rubber compounds based on the data mart with at least one of the components of the compounding vector added as an explanatory variable, A method for generating a physical property prediction model comprising the above steps.

2. The method for generating a physical property prediction model according to claim 1, wherein the data mart includes data on the property values of materials in the rubber compound.

3. The method for generating a physical property prediction model according to claim 2, wherein the data on the property values includes data calculated based on the compounding amounts of a plurality of materials.

4. The method for generating a physical property prediction model according to claim 1, wherein the data mart includes data on processing conditions in the rubber compound.

5. The method for generating a physical property prediction model according to claim 1, wherein the compounding vector generation step dimensionally compresses the material vector by principal component analysis.

6. The method for generating a physical property prediction model according to claim 1, wherein the arithmetic model is any one of a random forest model, a gradient boosting decision tree model, and a generalized linear regression model.

7. A data mart generation unit that generates a data mart including the compounding amounts of a plurality of materials used in the production of a plurality of rubber compounds and the physical properties of the rubber compounds, A co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any two-material combination in the data mart occurs is defined as the co-occurrence relationship between the materials, A material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, A compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector obtained by linearly combining using the compounding amounts; A learning processing unit that learns a learning-based arithmetic model for predicting the physical properties of a rubber compound based on the data mart obtained by adding at least one of the components of the compounding vector as an explanatory variable; A physical property prediction model generation device comprising the above.

8. A data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the production of a rubber compound for predicting physical properties and a plurality of rubber compounds used for learning a learning-based arithmetic model for predicting physical properties; A co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any two-material combination in the data mart occurs is defined as the co-occurrence relationship between the materials; A material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship; A compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector obtained by linearly combining using the compounding amounts; A physical property prediction step of adding at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicting the physical properties by the arithmetic model; A physical property prediction method comprising the above.

9. A data mart generation unit that generates a data mart including the compounding amounts of a plurality of materials used in the production of a rubber compound for predicting physical properties and a plurality of rubber compounds used for learning a learning-based arithmetic model for predicting physical properties; A co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any two-material combination in the data mart occurs is defined as the co-occurrence relationship between the materials; A material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generates a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship; A compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector obtained by linearly combining using the compounding amounts; A physical property prediction processing unit that adds at least one of the components of the compounding vector in the rubber compound for predicting physical properties as an explanatory variable and predicts the physical properties by the arithmetic model; A physical property prediction device comprising the above.

10. A data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for predicting physical properties and in the learning of a learning-type calculation model for predicting physical properties, A co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any combination of two materials in the data mart occurs, as the co-occurrence relationship between the materials, A material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, A compounding vector generation step of compressing the dimension of the material vector and generating a compounding vector linearly combined using the compounding amount, A physical property prediction step of adding at least one of the components of the compounding vector in a rubber compound for predicting physical properties as an explanatory variable, and predicting the physical properties by the calculation model, A physical property prediction program for causing a computer to execute.

11. A data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for predicting physical properties and in the learning of a learning-type calculation model for predicting physical properties, A co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, where the number of rubber compounds in which any combination of two materials in the data mart occurs, as the co-occurrence relationship between the materials, A material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each combination of materials reflecting the co-occurrence relationship, A compounding vector generation step of compressing the dimension of the material vector and generating a compounding vector linearly combined using the compounding amount, A compounding prediction step of predicting the compounding of a rubber compound by an optimization method based on a target physical property and constraint conditions regarding the compounding amounts of materials in the rubber compound, using the calculation model learned by adding at least one of the components of the compounding vector as an explanatory variable, A compounding prediction method comprising.

12. A data mart generation unit that generates a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for predicting physical properties and in the learning of a learning-type calculation model for predicting physical properties, A co-occurrence matrix generation unit that generates a co-occurrence matrix representing the co-occurrence relationship, taking the number of rubber compounds in which any two material combinations in the data mart occur as the co-occurrence relationship between materials; A material vector generation unit that calculates the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generates a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship; A compounding vector generation unit that dimensionally compresses the material vector and generates a compounding vector linearly combined using the compounding amount; A compounding prediction processing unit that predicts the compounding of a rubber compound by an optimization method based on a target physical property and constraints on the compounding amounts of materials in the rubber compound, using the calculation model learned by adding at least one component of the compounding vector as an explanatory variable; A compounding prediction device comprising the above.

13. A data mart generation step of generating a data mart including the compounding amounts of a plurality of materials used in the generation of a plurality of rubber compounds used for learning a learning-type calculation model for predicting physical properties; A co-occurrence matrix generation step of generating a co-occurrence matrix representing the co-occurrence relationship, taking the number of rubber compounds in which any two material combinations in the data mart occur as the co-occurrence relationship between materials; A material vector generation step of calculating the appearance frequency and rarity of each material in the data mart based on the co-occurrence matrix, and generating a material vector indicating the feature amount of each material combination reflecting the co-occurrence relationship; A compounding vector generation step of dimensionally compressing the material vector and generating a compounding vector linearly combined using the compounding amount; A compounding prediction step of predicting the compounding of a rubber compound by an optimization method based on a target physical property and constraints on the compounding amounts of materials in the rubber compound, using the calculation model learned by adding at least one component of the compounding vector as an explanatory variable; A compounding prediction program for causing a computer to execute the above.

Images