JP2020030683A - Rubber material design method, rubber material design device, and program - Google Patents

Abstract

To efficiently and accurately extract the combination of mixture of raw materials achieving a target value of a feature quantity characterizing a vulcanized rubber composition, even from among an extremely large number of raw materials.SOLUTION: A rubber material design method includes: taking the feature quantity of a vulcanized rubber composition as an objective function and the combination of mixture of raw materials as a design variable by using a prediction module of a machine-learned computer, and performing optimization to extract the combination of mixture achieving a set target value of the objective function; and when performing the optimization, imposing a constraint condition, creating the combination of mixture of the raw materials to be the design variable, and inputting the combination to the prediction module. The constraint condition is to include at least one or more candidates for the combination of mixture with set raw materials as a pair. The candidates for the combination of mixture are a pair of raw materials in which the combination of mixture is created with the amount of blending of the raw materials in input data for learning.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a rubber material design method, a rubber material design apparatus, and a program for causing a computer to calculate a combination of raw materials of a vulcanized rubber composition in order to realize a target value of a characteristic amount that characterizes the vulcanized rubber composition.

2. Description of the Related Art In recent years, techniques for making a computer perform machine learning to predict various feature amounts from input data have been actively proposed. For the vulcanized rubber composition prepared by blending a plurality of rubber materials, fillers, oils, and the like as raw materials, it is conceivable to apply the above technique to predicting physical property data.
BACKGROUND ART Conventionally, a vulcanized rubber composition has been prepared by mixing a plurality of rubber materials, fillers, oils, and the like by trial and error, and physical property data has been measured. For this reason, a large number of data in which the compounding information of the vulcanized rubber composition and the physical property data are linked. By utilizing this accumulated data, a computer can make machine learning to predict physical property data.

  For example, a technology relating to a multicomponent material optimization analysis apparatus and a design method capable of facilitating material design of a material composed of multiple components and presetting a design range of a material composition ratio and a desired range of mechanical behavior has been developed. It is known (Patent Document 1).

  In the above technique, learning is performed so as to convert the composition ratio of the multi-component material composed of multiple components into the mechanical behavior of the multi-component material, and the composition ratio of the multi-component material composed of multiple components is determined. A conversion system using a neural network that has a nonlinear correspondence and determines the mechanical behavior of a multi-component material as an input and an output is defined. Thereafter, an objective function representing the mechanical behavior is determined, and a constraint condition for restricting at least one of the allowable range of the mechanical behavior and the composition ratio of the multi-component material is determined. Further, by using the determined conversion system, the optimization which gives the optimal value of the objective function based on the objective function and the constraint condition is performed to determine the composition ratio of the multi-component material, and the multi-component system is determined based on the composition ratio of the multi-component material. Design rubber material.

Japanese Patent No. 4393586

In the above-mentioned technology, since the constraint condition that restricts the allowable range of at least one of the mechanical behavior and the composition ratio of the multi-component material is determined, the material design of the material composed of the multi-component is facilitated, and the design of the composition ratio of the material is facilitated. It is stated that a desired range of the range and the mechanical behavior can be set in advance. The constraint on the constituent ratio of the material is, for example, the mass of at least one constituent material of the multi-component material.
However, the multi-component material is generally appropriately selected from an extremely large number of raw materials such as several tens of rubber materials, several tens of fillers including oil, etc. Therefore, it is extremely difficult to extract and determine the composition ratio of the multi-component material that gives the optimum value of the target mechanical behavior from an extremely large number of raw materials even if the above-mentioned constraints are imposed. In addition, in a conversion system using a neural network, the accuracy of the output of the mechanical behavior may be degraded depending on the learning data used for the machine learning, so the optimization with high accuracy corresponding to the actual mechanical behavior Often can not.

  Therefore, the present invention, when searching for the combination of the raw material composition of the vulcanized rubber composition that achieves the target value of the characteristic amount that characterizes the vulcanized rubber composition, even from among the combination of a large number of raw material compositions, It is an object of the present invention to provide a rubber material design method, a rubber material design device, and a program that can efficiently and accurately extract a combination of raw materials.

One embodiment of the present invention relates to a combination of raw materials in the unvulcanized rubber composition in order to achieve a target value of a characteristic amount that characterizes a vulcanized rubber composition produced from the unvulcanized rubber composition. This is a rubber material design method to be calculated by a computer. The unvulcanized rubber composition is obtained by combining raw materials from a group of a plurality of raw materials set in advance.
The rubber material design method,
Learning data using the characteristic amount of each of a plurality of vulcanized rubber compositions produced by processing the unvulcanized rubber composition as learning output data, and using a combination of the raw materials as learning input data. Using the feature amount for the combination of the raw material combination machine learning a prediction module in the computer,
Using the machine learning-based prediction module, the characteristic amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization step in which the combination module is extracted by an optimization module of the computer.
The optimization step includes:
The optimization module, the learning input data in the combination of the raw materials in the combination of the raw material was created a combination of a plurality of combinations of combinations of raw materials to create a combination candidate,
The combination of the raw materials that is created in the optimization step is a combination of the raw material that includes at least one combination candidate under the constraint condition that includes at least one combination candidate. Created as, the prediction module calculates the output value of the feature amount for the design variable,
The optimization module, by using the combination of the raw material and the output value of the feature amount input as the design variable to the prediction module, extracting the combination of the combination to achieve the target value, Having.

  The optimization module assigns an identification number to each of the combination candidates, creates a combination of the raw materials as the design variables using the identification numbers, and the identification number is one of the combination candidates. It is preferable that the numbers are ranked according to the order of the total amount of the predetermined raw materials.

The group of raw materials includes at least a group of rubber raw materials, a group of fillers, and a group of vulcanization aids,
It is preferable that the raw materials that are candidates for the combination are the raw materials in each of the groups.

The group of raw materials includes at least a group of rubber raw materials, a group of fillers, and a group of vulcanization aids,
It is preferable that the raw materials that are the combination candidate of the combination are raw materials in which the groups are different from each other.

  Preferably, the step of optimizing extracts the combination of the combinations that achieves the target value using an evolutionary algorithm.

  It is preferable that, in the optimizing step, a combination of the combination that achieves the target value is extracted using a genetic algorithm.

A plurality of the objective functions are set,
The optimization module further comprises a step of visualizing a relationship between the combination of the blends used as the design variables and the plurality of the objective functions obtained by the optimization step and displaying the relationship on a screen. Is preferred.

Another embodiment of the present invention is a computer-based method of combining raw materials in the unvulcanized rubber composition to achieve a target value of a characteristic amount of the vulcanized rubber composition produced from the unvulcanized rubber composition. Is a program for calculating The unvulcanized rubber composition is obtained by combining raw materials from a group of a plurality of raw materials set in advance.
The program is
The physical property data relating to a plurality of vulcanized rubber compositions produced by combining and processing a plurality of blends of raw materials used in a vulcanized rubber composition is used as learning output data, and a combination of the blends of the raw materials is used as learning input data. Using the learning data to, the feature amount machine learning the feature amount for the combination of the raw material combination to the prediction module formed in the computer by the start of the program,
Using the machine learning-based prediction module, the characteristic amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization procedure in which an optimization module formed in the computer extracts the combination of the blends by activating the program, and
The optimization procedure is as follows:
The optimization module, in the learning input data, a procedure of creating a plurality of combination candidates of the combination of the raw materials are set as a combination of the combination of the mixture of the raw materials,
The combination of the raw materials to be created in the optimization procedure is a combination of the raw material combinations including at least one combination candidate under the constraint condition including at least one combination candidate. Creating, the prediction module calculates the output value of the feature amount for the design variable,
The optimization module, by using the combination of the raw material and the output value of the feature amount input as the design variable to the prediction module, the procedure of extracting the combination of the combination to achieve the target value, Having.

Another embodiment of the present invention calculates a combination of raw materials in the unvulcanized rubber composition in order to achieve a target value of the characteristic amount of the vulcanized rubber composition produced from the unvulcanized rubber composition. It is a rubber material design device. The unvulcanized rubber composition is obtained by combining raw materials from a group of a plurality of raw materials set in advance.
The rubber material design device,
The physical property data relating to a plurality of vulcanized rubber compositions produced by combining and processing a plurality of blends of raw materials used in a vulcanized rubber composition is used as learning output data, and a combination of the blends of the raw materials is used as learning input data. Using a learning data to be, a prediction unit for machine learning the feature amount for the combination of the raw material combination,
Using the machine-learned prediction unit, the feature amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization unit for extracting the combination of the blends,
The optimization unit, in the input data for learning, a portion that creates a plurality of combination candidates of the combination of the raw materials are set as a combination of the combination of the mixture of the raw materials,
The combination of the raw materials created by the optimization unit is a combination of the raw materials including at least one combination candidate under a constraint including at least one combination candidate as the design variable. And a part that causes the prediction module to calculate an output value of the feature amount with respect to the design variable;
A part that extracts the combination of the combinations that achieves the target value by using the combination of the raw materials and the output value of the feature amount that are input to the prediction module as the design variables.

  According to the rubber material design method, the rubber material design device, and the program described above, a combination of raw materials that achieves a target value of a characteristic amount that characterizes a vulcanized rubber composition can be selected from a combination of a large number of raw materials. Extraction can be performed efficiently and accurately.

It is a figure showing an example of a flow of a rubber material design method of one embodiment. It is a figure showing an example of composition of a rubber material design device which performs a rubber material design method of one embodiment. It is a figure showing the example of the detailed composition of the optimization module of the rubber material design device of one embodiment. It is a figure explaining the combination combination candidate used by the rubber material design method of one embodiment. It is a figure showing an example of the flow at the time of optimizing using a Genetic Algorithm. (A)-(c) is a figure which shows the example of the self-organization map regarding the compounding quantity of each raw material which is a design variable, (d)-(e) is the self-organization of the characteristic amount which is an objective function. It is a figure showing an example of a map.

  Hereinafter, a rubber material design method, a rubber material design device, and a program according to an embodiment will be described in detail.

(Outline of rubber material design method)
FIG. 1 is a diagram illustrating an example of a flow of a rubber material design method according to one embodiment. In the example shown in FIG. 1, the combination of raw materials before vulcanization, which was previously set by trial and error in order to obtain new physical properties, and vulcanization data which are physical properties of a vulcanized rubber composition after vulcanization Several hundred to several thousand sets of accumulated data are prepared by setting the characteristic amount of the rubber composition and the set. The feature amount in the accumulated data is not limited to the physical property data. For example, it may be raw material cost data.
Using the accumulated data as learning data, the prediction module formed in the computer provides the characteristics of the vulcanized rubber composition, which is physical property data of the vulcanized rubber composition after vulcanization with respect to the combination of raw materials before vulcanization. Machine learning of quantity. The combination of raw materials is used as input data for learning, and the characteristic amount of the vulcanized rubber composition is used as output data for learning.
Next, using a prediction module that has undergone machine learning, a plurality of combinations are created using the combination of the raw materials as a design variable so that the feature amount becomes the target value, and the feature amount at that time becomes the target value. , To search for combinations of raw materials.
At this time, constraints are imposed on the combination of raw materials to be created as design variables. In other words, instead of creating unlimited combinations of raw materials, the raw materials for which the combinations of raw materials are combined in the input data for learning are extracted, and the combinations of the extracted raw materials are combined. A plurality of candidates are created, and a combination of raw materials containing at least one of the plurality of combination candidates is created. Hereinafter, a specific description will be given.

As shown in FIG. 1, a characteristic amount relating to each of a plurality of vulcanized rubber compositions produced by processing an unvulcanized rubber composition and a blending of raw materials in the corresponding unvulcanized rubber composition at this time. The combination is prepared as learning data (step S10). The unvulcanized rubber composition is a combination of a plurality of raw materials used in the vulcanized rubber composition.
Next, the feature amount is used as learning output data, and the corresponding combination of raw material combinations is used as learning input data to determine the feature amount (the difference between the combination of raw material combinations and the feature amount) for the combination of raw material combinations. Is machine-learned by a prediction module in the computer (step S12).

Next, a constraint condition used in creating a combination of raw materials is created (step S14).
Thereafter, the optimization module creates various combinations of the mixing of the raw materials in accordance with the above-described constraints (step S16). For example, a combination of a predetermined number of raw materials is created by using at least one of a plurality of created combinations of raw material combinations as one element of a combination of combinations to be created now. That is, the constraint condition is that the combination of the raw material compositions created for the optimal calculation includes at least one combination candidate. As for the combination candidates, a set of raw materials in which the combination of the raw materials is formed based on the amounts of the raw materials in the input data for learning is determined as the combination candidates. In addition, in creating a combination of raw material combinations, an experiment design method or a Latin hypercube method can be used depending on the optimization calculation method.
The optimization module obtains a predicted value (output value) of the feature value by the prediction module by causing the prediction module to predict the feature value using the created combination of the raw materials (step S18).

Next, the optimization module performs an optimization operation using the combination of the raw material mixture created as the design variable and the predicted value of the feature amount corresponding to the combination (Step S20). In the optimization operation, it is preferable to use an evolutionary algorithm. The evolutionary algorithm includes Genetic Algorithm (Genetic Algorithm), Differential Evolution, Particle Swarm Optimization, Ant Colony Optimization, and the like.
It is also preferable to use the experimental design method or the Latin hypercube method by combining the combination of the raw materials and the optimization calculation. In this case, since the combination of the mixing of the raw materials is obtained by using the response surface method, it is preferable that the combination is created so as to cover the space of the design variable comprehensively.
Next, the optimization module extracts an optimal combination of raw materials that achieves the target value of the characteristic amount (step S22).

  As described above, before creating a combination of raw material combinations serving as design variables, a plurality of combination candidate combinations in which the raw material combinations are made based on the amounts of the raw materials in the learning input data are set. Create it. When creating a combination of raw material combinations that will be design variables, create as a design variable a combination of raw materials that contains, as an element, at least one combination combination candidate selected from multiple created combination combinations. Therefore, a combination of raw materials that achieves a target value of the characteristic amount that characterizes the vulcanized rubber composition can be extracted based on the combination combination candidate. For this reason, the optimal combination of combinations can be efficiently and accurately extracted from the combination of a large number of raw materials.

(Specific description of rubber material design method and rubber material design device)
FIG. 2 is a diagram illustrating an example of a configuration of a rubber material design apparatus that performs the rubber material design method according to one embodiment.
The rubber material design apparatus 10 shown in FIG. 2 is a computer having a CPU 12, a ROM (Read Only Memory) 14, a RAM (Random Access Memory) 16, and an input / output unit 18. The input / output unit 18 is connected to an input operation system 30 including a mouse and a keyboard, and a display 32.
The computer stored in the ROM 14 is called a computer, which is the rubber material design apparatus 10, and is called and activated, whereby the prediction module 20 and the optimization module 22 are formed. The prediction module 20 and the optimization module 22 correspond to a prediction unit and an optimization unit in the rubber material design device 10.

The CPU 12 calls up a huge amount of accumulated data in which a combination of the raw material composition recorded in the RAM 16 and a characteristic amount (for example, a physical property value) of the actual vulcanized rubber composition in the corresponding combination are called, and each of them is learned. To be supplied to the prediction module 20 as input data for learning and output data for learning.
The prediction module 20 is configured to use the prepared learning input data and learning output data to machine-learn a feature amount for a combination of raw material combinations. The prediction module 20 machine-learns a feature amount for a combination of raw material combinations using a preset prediction model. Predictive models include models using neural networks represented by well-known deep learning, models using the well-known random forest method that performs “classification” or “regression” using multiple decision trees, LASSO regression Includes models using. Further, as a prediction model, a non-linear function using a polynomial, Kriging, or RBF (Radial Base Function) can also be used.

FIG. 3 is a diagram illustrating an example of a detailed configuration of the optimization module 22. The optimization module 22 includes a constraint condition setting unit 22a, a combination combination creation / feature amount acquisition unit 22b, and an optimization calculation unit 22c.
The constraint condition setting unit 22a is configured to set a constraint condition for restricting the raw materials to be combined in order to create a plurality of combinations of the raw materials to be created for the optimization operation. Raw materials are all the raw materials used in the input data for learning.

FIG. 4 is a diagram illustrating a combination candidate of a rubber material among raw materials including a group of a rubber material, a filler, a resin material, and a vulcanization aid.
In the example shown in FIG. 4, two combinations of the raw materials in which the combination of the raw materials is made based on the amounts of the raw materials in the learning input data are created. One combination of raw materials is rubber materials 1, 2 and 3, and another combination of raw materials is rubber materials 10 and 11. The rubber materials 1, 2, 3 and the rubber materials 10, 11 are rubber materials that are not used simultaneously in the input data for learning. In such a rubber material, a plurality of combination candidates of the mixing of the raw materials are created. For example, in the rubber materials 1, 2, and 3, a combination of 50 parts by weight, 38 parts by weight, and 12 parts by weight is a candidate for one combination. In the rubber materials 1, 2, and 3, a combination of 75 parts by weight, 18 parts by weight, and 7 parts by weight is a candidate for one more combination. Further, in the rubber materials 1, 2, and 3, a combination of 88 parts by weight, 7 parts by weight, and 5 parts by weight is a further combination candidate. Similarly, in the rubber materials 10 and 11, the compounding amounts of the rubber materials 10 and 11 are sequentially 25 parts by weight and 75 parts by weight, and the compounding amounts of the rubber materials 10 and 11 are 63 parts by weight and 37 parts by weight, respectively. In the rubber materials 10 and 11, combinations of 80 parts by weight and 20 parts by weight are candidates for the combination.
Although the example shown in FIG. 4 is an example of a combination candidate of a rubber material, a combination candidate of a filler and the like can be created.

  Such a combination combination candidate is identified by an identification number. For example, the combination combination candidates shown in FIG. 4 are sequentially assigned identification numbers 1, 2, and 3. Since the optimization calculation is performed using the identification number, for example, when the optimization is performed using the response surface function, the blending amount of the predetermined raw material is reduced or increased according to the order of the identification number. By assigning the identification numbers to the IDs, the identification numbers can be given a ranking. The predetermined raw material is, for example, at least one or more raw materials that affect physical property data of the vulcanized rubber composition.

The combination combination creation / feature amount acquisition unit 22b creates, as a design variable, a combination of raw materials containing at least one combination candidate under the constraint set by the constraint setting unit 22a. A plurality of combinations of raw materials are created. The combination combination creation / feature acquisition unit 22b gives this combination to the prediction module to predict the feature. Thereby, the combination combination creation / feature amount acquisition unit 22b acquires the feature amount predicted by the prediction module. The constraint condition is that the combination of raw materials created by the combination combination creation / feature amount acquisition unit 22b includes at least one combination combination candidate described above.
For example, in rubber materials 1, 2, and 3, the combination candidates of the compounding amounts of 50 parts by weight, 38 parts by weight, and 12 parts by weight are set to identification number 1, and in rubber materials 10 and 11, the compounding amount is 25 parts by weight in order. When the combination candidate of 75 parts by weight is identified as the identification number 101, the combination combination creating / feature amount acquiring unit 22b determines the combination of the identification number 1 and the identification number 101 as the combination of the raw materials created for the optimal calculation. Create a combination of the combination candidates. In this case, the combination ratio between the combination candidate of the combination with the identification number 1 and the combination combination candidate with the identification number 101 can be set. That is, the combination of the raw material combinations created by the combination combination creation / feature amount acquisition unit 22b for the optimal calculation includes a combination combination candidate as one element.
In addition, although the combination combination candidate is set by an integer numerical value, such as 50 parts by weight, 38 parts by weight, and 12 parts by weight, for example, 48 to 52 parts by weight is regarded as 50 parts by weight, and 48 parts by weight is considered. Or 52 parts by weight, the same combination of 50 parts by weight can be determined as a combination candidate. Further, even when the numerical value of parts by weight includes a decimal number, rounding or the like can be performed to determine a combination combination candidate as an integer numerical value.

  The method of setting the blending of the raw materials (including the blending ratios of the above-mentioned blending combination candidates) in the blending combination created by the blending combination creating / feature amount acquiring unit 22b differs depending on the optimization calculation method. In the experimental design method or the Latin hypercube method, a preset blending amount or blending ratio level is prepared, and the blending amount and the blending ratio are given. In the case of the Genetic Algorithm, the compounding amount and the compounding ratio are randomly added. Note that, in the case of the Genetic Algorithm, prediction is performed by the iterative prediction module every time a generation is switched between the first generation, the second generation, the third generation, and so on.

The optimal operation unit 22c uses the combination of the raw material mixture created by the combination combination creation / feature amount acquisition unit 22b and the predicted value of the feature amount to combine the combination of the raw materials that achieves the target value of the feature amount. Specifically, the amount of each raw material is extracted. In the combination of combinations created by the combination combination creation / feature amount acquisition unit 22b, a combination combination candidate is set by an identification number as one element of the combination. Is converted into the blending amount of the raw material corresponding to. In the case of the combination candidate of the rubber materials 1, 2 and 3 having the identification number 1, the rubber material 1 is converted into 20 parts by weight, the rubber material 2 is converted into 15 parts by weight, and the rubber material 3 is converted into 5 parts by weight.
For example, when the experimental design method or the Latin hypercube method is used, the number of combinations of raw materials is increased, and the data of the combinations is exhaustively distributed in the space of the design variables. Is created, and the data that is closest to the target value of the feature value is extracted by performing interpolation or the like, thereby extracting the optimum blending amount.

FIG. 5 is a diagram illustrating an example of a flow in a case where the optimal operation unit 22c optimizes using a Genetic Algorithm.
First, before the optimization operation by the optimum operation unit 22c, the constraint condition setting unit 22a sets a plurality of combinations of combination candidates in which the combination of the raw materials formed by the mixture amounts of the raw materials is set in the learning input data. It is created (step S100).

Next, the combination combination creation / feature amount acquisition unit 22b sets the combination of the mixture of the raw materials to be the design variables in advance under the constraints included in the combination of the combinations as the design variables. The created number is created (step S102). An identification number is assigned to each combination candidate, and the combination combination creation / feature amount acquisition unit 22b creates a combination of raw materials as design variables using the identification number. In creating a combination of raw materials, a combination of combinations is randomly created, or a combination is created by a Latin hypercube method.
As a result, the combination combinations created by the combination combination creation / feature amount acquisition unit 22b include combination combination candidates.
The combination of the raw materials obtained in step S102 is a combination of the first generation.

  Next, the combination combination creating / feature amount acquiring unit 22b gives the created combination of the raw material combinations to the prediction module 20 to calculate the predicted value of the feature amount. Thereby, the combination combination creation / feature amount acquisition unit 22b acquires the predicted value of the feature amount corresponding to the combination of the raw material combinations (step S104).

The optimization calculation unit 22c determines whether or not the obtained predicted value of the feature value includes a combination of combinations that achieves the target value of the feature value, and determines whether or not the optimization calculation ends. Is determined (step S106). In this determination, when it is determined that there is a combination of combinations that achieves the target value of the feature value, and when it is determined that the optimization calculation may be terminated, the optimization calculation ends, and the target value of the feature value is changed. The realized combination of the raw materials is displayed on the display 32 as the optimum combination.

  In the above determination, when there is no combination of combinations that achieve the target value of the feature amount and it is determined that the optimization calculation is to be continued, the optimization calculation unit 22c sets the Pareto rank of the combination of the created raw material combinations. The calculation is performed (step S108). In the illustrated example, there are a plurality of feature amounts. The Pareto rank ranks how close prediction values corresponding to a plurality of objective functions are to target values of a plurality of feature amounts as an objective function. Even in the case of the same rank, it is possible to give priority to the degree of the distance in the space between the target value and the predicted value. Instead of the Pareto rank, a plurality of feature amounts, which are objective functions, may be weighted and added as one feature amount, and the combination of raw materials may be ranked using the objective function of the integrated feature amounts. In addition, when there is one feature amount, ranking can be performed based on the magnitude of the difference between the predicted value of the feature amount and the target value of the feature amount for each combination of the raw materials.

  Next, the optimization calculation unit 22c divides, for example, into upper ranks (closer to the target value) or lower ranks (far from the target value) according to the Pareto rank, Two combinations are selected, and information of the two combinations is exchanged (a part of the information is exchanged), and at least one new combination is selected for the two selected combinations. It is created (step S110).

  Next, the optimization calculation unit 22c determines whether or not to mutate the new combination of combinations obtained by the crossover (step S112). Mutations are generated randomly, for example. When the mutation is not performed, the new combination of the combinations obtained by the crossover is maintained (step S114), and the process proceeds to step S118. On the other hand, when mutating, the amount of a part of the new combination obtained by the crossover is changed. It is possible to randomly determine which raw material should be changed in the blending amount. Thereby, the optimization calculation unit 22c corrects the new combination of the combinations obtained by the crossover (step S116).

Furthermore, the optimization calculation unit 22c determines whether the combination of the combination maintained in step S114 or the combination of the combination modified in step S116 satisfies the above-described constraint condition. The modified combination of combinations is further modified so as to satisfy the constraint condition (step S118). Here, satisfying the constraint condition means that the corrected combination of combinations can be represented using at least one combination combination candidate.
Thus, the combination of the second generation combination is set. The number of second generation combination combinations remains the same as the first generation combination combinations. Thus, also in the second generation, a combination of combinations including at least one combination combination candidate is created.

  In this way, until the predicted value of the feature value in each generation reaches a combination of combinations that realizes the target value of the feature value in step S106, steps S104 to S104 are performed until the optimization calculation can be completed. S118 is repeated.

Lastly, it is preferable that the optimization calculation unit 22c visualizes the relationship between the combination of the composition adopted as the design variable and the feature amounts as a plurality of objective functions and displays the relationship on the display 32. FIGS. 6A to 6C are diagrams showing examples of a self-organizing map relating to the blending amount of each raw material which is a design variable, and FIGS. 6D to 6E show feature amounts which are objective functions. 5 is a diagram showing an example of a self-organizing map of FIG. 6 (a) to 6 (c) are color-coded according to the amount of each raw material. FIGS. 6D to 6E color-code according to the magnitude of each feature amount. By paying attention to the same position in each rectangular area, it is possible to know the value of each feature amount and the blending amount that realizes this feature amount. In this way, it is possible to know the relationship between the value of the feature quantity as the objective function and the blending amount of the raw material as the design variable.
The example shown in FIG. 6 is a self-organizing map. Instead of the self-organizing map, a scatter diagram is used to determine the relationship between the combination of the blends adopted as design variables and the feature amounts as a plurality of objective functions. It is also preferable to visualize the relationship.

As described above, in the optimization module 22, before creating the combination of the raw materials used for the optimization, the raw materials in which the combination of the raw materials are formed based on the amount of the raw materials in the learning input data are combined with each other. When a plurality of combinations of raw materials are created and combinations of raw materials are created, combinations of raw materials are created as design variables so as to include the created combinations. For this reason, the combination of the raw materials can be restricted as compared with the related art. Therefore, it is possible to efficiently and accurately extract a combination of raw materials that achieves the target value of the characteristic amount that characterizes the vulcanized rubber composition, from an extremely large number of raw materials.
In addition, for the raw materials that are not included in the combination candidates, a combination of the blends is created using the respective blend amounts, and the optimization calculation is performed.

  The optimization module assigns an identification number to each combination combination candidate, and uses the identification number to create a combination combination of raw materials that is a design variable. At this time, according to one embodiment, it is preferable that the identification number is a number that is ranked among the candidate combinations according to the order of the total amount of the predetermined raw materials. In the case of the optimization operation, the order of the identification application can also have technical significance (the order of the magnitude of the total amount of the predetermined raw materials according to the identification number), and the identification number is used in the optimization operation. It can be used effectively.

The group of raw materials includes at least a group of rubber raw materials, a group of fillers, and a group of vulcanization aids. It is preferable that the raw materials that are the combination candidate of the combination are the raw materials in the above group. Thereby, a combination of raw materials can be extracted more efficiently and more accurately.
Further, according to one embodiment, in consideration of the compatibility between the rubber material and the filler, it is also preferable that the raw materials that are candidates for the combination of the compounds are raw materials having different groups. That is, a combination combination candidate may be created from raw materials in different groups.

In the above-described embodiment, when a combination of raw materials is created in the optimization step, a constraint that the combination of raw materials to be created includes at least one combination candidate is imposed. In addition, a constraint that limits the number of raw materials used as design variables may be imposed. Further, a constraint condition that a combination of two set raw materials is not created or a constraint condition that a combination is created such that two set raw materials are always set may be imposed.
When the number of raw materials is limited, the number of raw materials is limited to the maximum number of raw materials used in each of the combinations of raw materials used as input data for learning. Is preferably limited.
When a combination of two raw materials is not created, it is preferable that two raw materials having no combination among the combinations of the raw materials used as the input data for learning are targets for which no combination is created.
Further, when a combination is created such that two raw materials are always set, a set that creates a combination of two raw materials that are always set in a combination of raw material combinations used as input data for learning is used. Is preferred.
In the past accumulated data used as input data for learning, the combination of raw materials is often limited because raw materials are not preferably used as a set or raw materials are preferably used as a set. It is not preferable to create a combination of raw materials without considering such circumstances in designing a realistic rubber material.

As described above, when the rubber material design method is executed by a computer, the method is executed by calling and starting a program. In this case, the following program is started.
The program is, in order to achieve a target value of a feature amount characterizing a vulcanized rubber composition made from an unvulcanized rubber composition prepared by combining raw materials from a plurality of predetermined raw material groups, and This is a program that allows a computer to calculate the combination of raw materials.
This program is
The physical property data relating to a plurality of vulcanized rubber compositions produced by combining and processing the blends of a plurality of raw materials used for the vulcanized rubber composition are used as learning output data, and the combination of the blending of the raw materials is used as learning input data. Using the learning data to perform a machine learning process on a prediction module 20 formed on a computer by launching a program, based on a combination of raw materials,
Using the machine learning-based prediction module 20, the combination of the vulcanized rubber composition is used as an objective function, the combination of raw materials is used as a design variable, and the combination of the combinations that achieves the set target value of the objective function is defined in the program. And an optimization procedure extracted by the optimization module 22 formed in the computer upon activation.
The optimization procedure is
A procedure in which the optimization module 22 creates, in the learning input data, a plurality of combination combination candidates each of which includes a pair of raw materials in which the combination of the raw materials is formed by the mixing amount of the raw materials;
The combination of raw materials created by the optimization procedure is a combination of raw materials containing at least one combination candidate as a design variable under a constraint condition that includes at least one combination candidate as a design variable. 20: a procedure for calculating an output value of a feature quantity with respect to a design variable;
A step in which the optimization module 22 extracts a combination of the combinations that achieves the target value by using the combination of the raw materials and the output value of the characteristic amount input as the design variables to the prediction module 20.

  As described above, the rubber material design method, the rubber material design device, and the program of the present invention are not limited to the above embodiments, and various improvements and changes may be made without departing from the gist of the present invention. is there.

10 Rubber material design equipment 12 CPU
14 ROM (Read Only Memory)
16 RAM (Random Access Memory)
18 Input / output unit 20 Prediction module 22 Optimization module 22a Constraint condition setting unit 22b Mixing combination creation / feature amount acquisition unit 22c Optimization operation unit 30 Input operation system 32 Display

Claims (9)

A rubber material design in which a computer calculates a combination of raw materials in the unvulcanized rubber composition in order to achieve a target value of a characteristic amount that characterizes a vulcanized rubber composition produced from the unvulcanized rubber composition. The method
The unvulcanized rubber composition is a composition obtained by combining raw materials from a group of a plurality of raw materials set in advance,
Learning data using the characteristic amount of each of a plurality of vulcanized rubber compositions produced by processing the unvulcanized rubber composition as learning output data, and using a combination of the raw materials as learning input data. Using the feature amount for the combination of the raw material combination machine learning a prediction module in the computer,
Using the machine learning-based prediction module, the characteristic amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization step in which the combination module is extracted by an optimization module of the computer.
The optimization step includes:
The optimization module, the learning input data in the combination of the raw materials in the combination of the raw material was created a combination of a plurality of combinations of combinations of raw materials to create a combination candidate,
The combination of the raw materials that is created in the optimization step is a combination of the raw material that includes at least one combination candidate under the constraint condition that includes at least one combination candidate. Created as, the prediction module calculates the output value of the feature amount for the design variable,
The optimization module, by using the combination of the raw material and the output value of the feature amount input as the design variable to the prediction module, extracting the combination of the combination to achieve the target value, A rubber material design method comprising:   The optimization module assigns an identification number to each of the combination candidates, creates a combination of the raw materials as the design variables using the identification numbers, and the identification number is one of the combination candidates. The rubber material design method according to claim 1, wherein the numbers are ranked in accordance with the order of the total amount of the predetermined raw materials. The group of raw materials includes at least a group of rubber raw materials, a group of fillers, and a group of vulcanization aids,
The rubber material design method according to claim 1, wherein the raw materials that are candidates for the combination are the raw materials in each of the groups. The group of raw materials includes at least a group of rubber raw materials, a group of fillers, and a group of vulcanization aids,
The rubber material designing method according to any one of claims 1 to 3, wherein the raw materials that are candidates for the combination of the compounds are raw materials in which the groups are different from each other.   The rubber material design method according to any one of claims 1 to 4, wherein the step of optimizing extracts a combination of the compound that achieves the target value using an evolutionary algorithm.   The rubber material design method according to claim 5, wherein the optimizing step extracts a combination of the compound that achieves the target value using a genetic algorithm. A plurality of the objective functions are set,
The optimization module further comprises a step of visualizing a relationship between the combination of the blends used as the design variables and the plurality of the objective functions obtained in the optimization step and displaying the relationship on a screen. Item 7. The rubber material designing method according to any one of Items 1 to 6. A program for causing a computer to calculate a combination of raw materials in the unvulcanized rubber composition in order to achieve a target value of the characteristic amount of the vulcanized rubber composition produced from the unvulcanized rubber composition,
The unvulcanized rubber composition is a composition obtained by combining raw materials from a group of a plurality of raw materials set in advance,
The physical property data relating to a plurality of vulcanized rubber compositions produced by combining and processing a plurality of blends of raw materials used in a vulcanized rubber composition is used as learning output data, and a combination of the blends of the raw materials is used as learning input data. Using the learning data to, the feature amount machine learning the feature amount for the combination of the raw material combination to the prediction module formed in the computer by the start of the program,
Using the machine learning-based prediction module, the characteristic amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization procedure in which an optimization module formed in the computer extracts the combination of the blends by activating the program, and
The optimization procedure is as follows:
The optimization module, in the learning input data, a procedure of creating a plurality of combination candidates of the combination of the raw materials are set as a combination of the combination of the mixture of the raw materials,
The combination of the raw materials to be created in the optimization procedure is a combination of the raw material combinations including at least one combination candidate under the constraint condition including at least one combination candidate. Created as, the prediction module calculates the output value of the feature amount for the design variable,
A step of extracting the combination of the combinations that achieves the target value by using the combination of the raw materials and the output value of the feature amount input as the design variables to the prediction module, A program characterized by having: A rubber material design apparatus that calculates a combination of raw materials in the unvulcanized rubber composition in order to achieve a target value of the characteristic amount of the vulcanized rubber composition produced from the unvulcanized rubber composition,
The unvulcanized rubber composition is a composition obtained by combining raw materials from a group of a plurality of raw materials set in advance,
The physical property data relating to a plurality of vulcanized rubber compositions produced by combining and processing a plurality of blends of raw materials used in a vulcanized rubber composition is used as learning output data, and a combination of the blends of the raw materials is used as learning input data. Using a learning data to be, a prediction unit for machine learning the feature amount for the combination of the raw material combination,
Using the machine-learned prediction unit, the feature amount of the vulcanized rubber composition is used as an objective function, and a combination of the raw materials is used as a design variable to realize a set target value of the objective function. An optimization unit for extracting the combination of the blends,
The optimization unit, in the input data for learning, a portion that creates a plurality of combination candidates of the combination of the raw materials are set as a combination of the combination of the mixture of the raw materials,
The combination of the raw materials created by the optimization unit is a combination of the raw materials including at least one combination candidate under a constraint including at least one combination candidate as the design variable. And a part that causes the prediction module to calculate an output value of the feature amount with respect to the design variable;
A part for extracting the combination of the combinations that achieves the target value by using the combination of the raw materials and the output value of the feature amount input as the design variables to the prediction module. Rubber material design equipment.