JP7610105B2 - Method and system for designing rubber compound - Google Patents

Description

The present invention relates to a method and system for designing rubber compounds, and more specifically, to a method and system for designing rubber compounds that can accurately and easily determine a rubber compound that meets the target rubber properties.

Vulcanized rubber is produced by vulcanizing unvulcanized rubber. Unvulcanized rubber is made by blending multiple types of compounding components in specific amounts. By varying the amounts of each compounding component, it is possible to change the rubber properties of the vulcanized rubber produced. In other words, to produce vulcanized rubber with the desired rubber properties, it is necessary to appropriately set the rubber compounding (the amount of each compounding component) according to the desired rubber properties.

Conventionally, to manufacture vulcanized rubber with the target rubber properties, for example, first, multiple types of unvulcanized rubber are molded with different rubber compositions, and then each unvulcanized rubber is vulcanized to manufacture vulcanized rubber. Next, the rubber properties of the multiple types of vulcanized rubber manufactured are measured to identify a vulcanized rubber with rubber properties that are close to the target rubber properties. By referring to the rubber composition of the identified vulcanized rubber, a rubber composition that matches the target rubber properties is identified. Identifying an appropriate rubber composition using this method requires a great deal of trial and error.

In recent years, a device and method have been proposed that uses machine learning to assist in the design of manufacturing recipes for products with specified characteristics (see, for example, Patent Document 1). The device and method proposed in Patent Document 1 are not specialized for rubber products, so rubber-specific considerations must be taken into account when applying them to rubber products. In addition, with this proposed device and method, there is a risk that the computational processing load will be excessive in order to accurately determine a rubber compounding that matches the target rubber physical properties, so it is necessary to reduce this load. Therefore, there is room for improvement in accurately and simply determining a rubber compounding that matches the target rubber physical properties.

JP 2019-86817 A

The object of the present invention is to provide a method and system for designing a rubber compound that can easily and accurately determine a rubber compound that meets the target rubber properties.

In order to achieve the above object, the rubber compounding design method of the present invention is a rubber compounding design method for grasping compounding data in which compounding amounts of each compounding component of vulcanized rubber that satisfies target values of multiple types of rubber property data are set, the multiple types of rubber property data of each vulcanized rubber produced by vulcanizing multiple types of unvulcanized rubber with different compounding data are obtained, and each of the compounding data, each of the obtained rubber property data, and each of the vulcanized rubber vulcanization condition data are used as learning data, and under constraint conditions set by predetermined constraint condition data, a compounding generation model constructed using an adversarial generative network based on the learning data is used to perform calculation processing by a calculation device to obtain the compounding data in which compounding amounts of each compounding component of vulcanized rubber that satisfies target values of multiple types of rubber property data are set, The method is characterized in that a plurality of candidates for compounding data that are determined to satisfy data constraints and satisfy the target values of each of the rubber physical property data and are different from the compounding data of the learning data are generated, and a prediction model constructed by machine learning based on the learning data is used to perform calculations using the compounding data and vulcanization condition data of each of the generated candidates by the arithmetic device, thereby calculating the rubber physical property data of each of the target vulcanized rubbers that are formed with the compounding data of each of the candidates and vulcanized based on the vulcanization condition data, and selecting from each of the candidates the compounding data of the target vulcanized rubber for which each of the calculated rubber physical property data is close to the target value.

The rubber compounding design system of the present invention has a calculation device, an input unit for inputting data to the calculation device, and an output unit communicably connected to the calculation device, and is a rubber compounding design system that grasps compounding data in which the compounding amounts of each compounding component of vulcanized rubber that satisfies target values of multiple types of rubber property data are set,
The rubber physical property data of each of the vulcanized rubbers produced by vulcanizing a large number of types of unvulcanized rubbers with different compounding data, each of the compounding data, and each of the vulcanized rubbers' vulcanization condition data are used as learning data, and a compounding generation model constructed using an adversarial generation network based on the learning data is used under constraint conditions set by predetermined constraint condition data to perform calculation processing by a calculation device, and it is determined that the constraints imposed by the constraint condition data are satisfied and that the target values of each of the rubber physical property data are satisfied. A plurality of candidates for compounding data different from the compounding data of the learning data are generated, and a prediction model constructed by machine learning based on the learning data is used to perform calculation processing based on the compounding data and vulcanization condition data of each of the generated candidates, thereby calculating each of the rubber physical property data of the target vulcanized rubber formed with the compounding data of each of the candidates and vulcanized based on the vulcanization condition data, and the compounding data of the target vulcanized rubber for which each of the calculated rubber physical property data is approximate to the target value is selected from each of the candidates and output by the output unit.

According to the present invention, by using the compounding generation model constructed using a generative adversarial network based on the learning data under constraint conditions set by predetermined constraint condition data, multiple candidates of compounding data that are determined to satisfy the constraints set by the constraint condition data and to satisfy the target values of each of the rubber physical property data and are different from the compounding data of the learning data are generated. That is, since the candidate compounding data is generated under predetermined constraint conditions, the burden of the calculation process for generating the candidate compounding data is reduced. In addition, by performing calculation processing using the prediction model based on the compounding data of each of the candidates and the vulcanization condition data of the target vulcanized rubber, the rubber physical property data of each of the target vulcanized rubber can be calculated with high accuracy. As a result, the compounding data of the target vulcanized rubber in which each of the rubber physical property data is highly accurately close to the target value can be easily grasped.

FIG. 1 is an explanatory diagram illustrating a rubber compound design system of the present invention. FIG. 11 is an explanatory diagram illustrating learning data. An explanatory diagram of a combination generation model that generates candidates for combination data. 1 is a flow diagram illustrating a procedure for determining compounding data of a target vulcanized rubber whose rubber physical property data approximates a target value.

The rubber compounding design method and system of the present invention will be specifically explained below based on the embodiment shown in the figure.

The rubber compounding design method of the present invention is carried out using a rubber compounding design system 1 of the present invention illustrated in Fig. 1. Vulcanized rubber G is produced by vulcanizing unvulcanized rubber R. The vulcanized rubber G (unvulcanized rubber R) is formed by compounding a plurality of compounding components C ( _C1 , _C2 , ... _Cn ) in specific compounding amounts. The present invention has been devised and created in order to accurately and easily grasp the rubber compounding (the compounding amount of each compounding component) of vulcanized rubber G having the target rubber physical properties.

This design system 1 has a calculation device 2, an input unit 3, and an output unit 4. The calculation device 2, the input unit 3, and the output unit 4 are connected to each other so that they can communicate with each other via wires or wirelessly. In this embodiment, the design system 1 further has a measurement device 5. Data acquired by the measurement device 5 is input to the calculation device 2 through the input unit 3. The measurement device 5 may be independent and separate from the design system 1, and the data acquired by the measurement device 5 may be input to the calculation device 2.

The calculation device 2 receives and stores various data and performs calculations using these data. Various computers can be used as the calculation device 2. Therefore, the calculation device 2 has a memory in which various data are stored and a CPU that performs calculations.

The input unit 3 is an input means for inputting various data to the calculation device 2. Examples of the input unit 3 include a keyboard, a mouse, various terminal devices, and an interface provided in the calculation device 2.

The output unit 4 is an output means for displaying various data input to the calculation device 2 and the results of calculations (numerical values, tables, drawings, etc.) performed using the data. Examples of the output unit 4 include various monitors and printers.

The measuring device 5 is a means for measuring and acquiring rubber property data D2 of the vulcanized rubber G. Examples of the rubber property data D2 include various known rubber properties such as tan δ (loss factor), hardness, breaking strength, breaking elongation, 300% modulus, specific gravity, and vulcanization speed T95. In the present invention, multiple desired types of rubber property data D2 are used. The measuring device 5 uses a known device corresponding to the measurement of each rubber property data D2, and each rubber property data D2 is measured and acquired by a known method (such as the corresponding JIS standard). The vulcanization speed T95 is the vulcanization time up to 95% of the maximum torque obtained from a vulcanization curve of torque and vulcanization time obtained using a vibration disk vulcanization tester.

In the present invention, the learning data shown in FIG. 2 is used. This learning data includes compounding data D1 for many types of vulcanized rubber G (G1, G2, ...) manufactured with different compounding data D1, multiple types of desired rubber property data D2, and vulcanization condition data D3 for each vulcanized rubber G. Although FIG. 2 shows data for only four types of vulcanized rubber, in reality, data for a very large number of types of vulcanized rubber G becomes the learning data. The compounding data D1 is data that specifies the compounding amount of each compounding component C. The learning data includes parameters (α, β, ...) that define each compounding data D1. The parameter values are numerical values that can be converted into the compounding data D1, and if the parameter values are known, the compounding data D1 is also known. These parameter values are used in the calculation.

To obtain the learning data, unvulcanized rubber R is molded based on the compounding data D1 in which the compounding amount of each compounding component C is set. At that time, many types of unvulcanized rubber R with different compounding data D1 are molded. Varying the compounding data D1 means varying the compounding amount of at least one type of each compounding component C. There are cases where the compounding amount of a particular compounding component C is zero, in which case the type of compounding component C used will be different.

Next, each molded unvulcanized rubber R is vulcanized under specified vulcanization conditions to produce vulcanized rubber G. In other words, many types of vulcanized rubber G with different compounding data D1 are produced. The compounding amount of each compounding component C affects the rubber property data D2 of the vulcanized rubber G, but the extent of this effect differs depending on the type and compounding amount of compounding component C.

Examples of vulcanization conditions include vulcanization temperature, vulcanization pressure, and vulcanization time. An index indicating these vulcanization conditions is the vulcanization condition data D3. The vulcanization conditions are set to appropriate conditions for the unvulcanized rubber R so that there is neither too much nor too little vulcanization. The vulcanization condition data D3 also affects the rubber physical property data D2. When manufacturing vulcanized rubber G at the laboratory level, for example, vulcanization can be performed under standardized vulcanization conditions, and the vulcanization conditions for all vulcanized rubbers G can be set to be constant. This makes it possible to substantially eliminate the influence of differences in the vulcanization condition data D3 on the rubber physical property data D2 of the vulcanized rubber G.

Next, the rubber property data D2 of each vulcanized rubber G produced is measured and obtained using a measuring device 5. From the many types of rubber property data D2 exemplified above, desired types of rubber property data D2 are measured and obtained. A manufacturer of rubber products can effectively use the vast amount of data D1, D2, and D3 that has been accumulated in the past as learning data.

In the present invention, a compounding generation model GM constructed using a conditional generative adversarial net (CGAN) is used. This compounding generation model GM is a type of computer program that determines that vulcanized rubber G satisfies target values D2a for multiple desired rubber properties and generates candidates for compounding data D1 (data with different combinations and amounts of compounding components C) that differ from the training data.

Furthermore, the present invention utilizes a prediction model PM constructed by machine learning based on the learning data. This prediction model PM is a type of computer program for predicting multiple desired types of rubber property data D2 of a target vulcanized rubber Ga produced by vulcanizing a target unvulcanized rubber Ra under specified vulcanization conditions.

Next, we will explain an example of the procedure for constructing the breeding generation model GM.

As shown in FIG. 3, the generative mix model GM can be generalized and constructed using a known CGAN having a generation unit and an identification unit. The generation unit generates a fake sample (x*) for a real sample (x) of the training data. A random vector Z and predetermined constraint condition data Dy are input to the generation unit. In response to this input, the generation unit outputs a fake sample (x*|y, y) that satisfies the constraints imposed by the constraint condition data Dy, with the constraints imposed by the constraint condition data Dy being label y. In other words, the generation unit outputs a combination of the fake sample (x*) and label y. The identification unit receives input of this combination (x*|y, y) generated by the generation unit and the combination (x, y) of the training data sample (x) and the constrained label y of the training data sample (x).

The discrimination unit judges whether the input (x*|y,y) and (x,y) are a combination (x,y) of real sample x and label y, and the generation unit and discrimination unit are trained (weights are changed) by backpropagation based on the judgment result. When training the generator, the weights of the discrimination unit are fixed, and when training the discrimination unit, the weights of the generator are fixed. The discrimination unit repeats the process of judging the combination (x*|y,y) generated by the generation unit and the combination (x,y) of the training data, and the discrimination unit and the generation unit are trained alternately for one set of mini-batches. By repeating this process, the discrimination unit improves the judgment accuracy so that it can make a correct judgment even if the generation unit generates a combination (x*|y,y) that is likely to exist based on the combination (x,y) of the training data. The generation unit improves the generation accuracy so that it can generate a combination (x*|y,y) that is judged by the discrimination unit with a higher probability to be the combination (x,y) of the training data. In this way, the generation section and the identification section are fully trained to construct the blending and molding model GM.

The constraint condition data Dy can be set arbitrarily, and examples include the number of compounding components C and the number of types of rubber components. In the present invention, the range that the compounding data D1 of the learning data can take becomes a sample space, and candidates for the compounding data D1 are generated as many times as the number of occurrences of the random vector Z. The generation unit receives the numerical values (physical properties) of the desired multiple types of rubber physical property data D2 from the learning data, the constraint condition data Dy, and the number of occurrences of the random vector Z. Based on the input of the numerical values of each rubber physical property data D2, the generation unit calculates and generates the compounding data D1 (parameter value) according to the set weight. The generated compounding data D1 is data that satisfies the constraint condition data Dy. The identification unit receives the compounding data D1 (parameter value) generated by the generation unit from the generation unit, and the compounding data D1 (parameter value) held by the learning data from the learning data. The identification unit determines whether the compounding data D1 generated by the generation unit (i.e., a fake sample) is similar to the compounding data D1 held by the learning data (i.e., a real sample). This process is repeated to construct a blending and molding model GM with fully trained generating and identifying sections, as described above.

The constructed blend generation model GM is stored in the calculation device 2. Using the blend generation model GM constructed using CGAN, candidates for blend data D1 are generated that satisfy the constraints set in the predetermined constraint condition data Dy and are determined to satisfy the target values D2a of the desired multiple types of rubber physical property data D2. The generated candidate blend data D1 is different from the blend data D1 present in the learning data.

Next, we will explain an example of the procedure for constructing the prediction model PM.

A prediction model PM is constructed to predict the rubber physical property data D2 of the target vulcanized rubber Ga by machine learning using the above-mentioned learning data. Specifically, the above data D1, D2, and D3 are input to the calculation device 2, and machine learning is performed using these data D1, D2, and D3. That is, the relationship between each rubber physical property data D2 and the compounding data D1 and vulcanization condition data D3 is linked, and the influence of the compounding data D1 and the vulcanization condition data D3 on each rubber physical property data D2 of the vulcanized rubber G to be manufactured is analyzed and evaluated to construct the prediction model PM. In terms of the compounding data D1, the influence of the type and amount of compounding component C on the rubber physical property data D2 is analyzed and evaluated.

The factors of the data D1 and D3 described above affect each other in a complex way, causing the rubber physical property data D2 of the vulcanized rubber Ga to change. Therefore, by using the data D1, D2, and D3 as learning data, it is possible to build a prediction model PM by machine learning the degree of difference in the data D2 (rubber physical property data) due to the combination of the data D1 and D3, the combination of each compounding component C in the data D1, the difference in the compounding amount of each compounding component C, etc., and the characteristics of the data D2 in each case. As a machine learning method, a well-known method such as deep learning using a neural network can be exemplified. In deep learning, a network is used in which a multi-layer structure is formed by an input layer, multiple intermediate layers, and an output layer using a well-known method, and weights are set between nodes between each layer to link them. Then, the compounding data D1 and vulcanization condition data D3 that affect each rubber physical property data D2 are input from the input layer, and the predicted value of the rubber physical property data D2 is output to the output layer. The calculated predicted value is compared with the actual measured value of each rubber physical property data D2, and the weights are changed so as to reduce the error between the two. This allows the construction of a prediction model PM with improved prediction accuracy. The constructed prediction model PM is stored in the calculation device 2.

Next, an example of a procedure for determining the rubber compounding D1 of a target vulcanized rubber Ga having target rubber physical properties will be described. That is, an example of a procedure for determining compounding data D1 of a target vulcanized rubber Ga having rubber physical properties that satisfy the respective target values D2a set for multiple desired types of rubber physical property data D2 will be described.

The main steps are as shown in FIG. 4. First, the necessary data is input to the calculation device 2 using the input unit 3. The data to be input is the target value D2a of the rubber property data D2 for the target vulcanized rubber Ga shown in Table 1 below, the constraint condition data Dy for the target vulcanized rubber Ga shown in Table 2 below, and the vulcanization condition data D3 for the target vulcanized rubber Ga shown in Table 3 below.

As shown in Table 1, the target value D2a of the rubber physical property data D2 is set for multiple desired types of rubber physical property data D2. In this example, five types of desired types of rubber physical property data D2 are set: tan δ at 60°C (initial strain 10%, amplitude ±2%, frequency 20Hz), hardness HS at 0°C, 300% modulus, vulcanization speed T95, and specific gravity, but the desired type of rubber physical property data D2 is adopted without being limited to these.

As shown in Table 2, the constraint condition data Dy sets the number of compounding components C of the target vulcanized rubber Ga and the number of types of rubber components in the compounding components C. The constraint condition data Dy is not limited to the types shown here, and may be of any type as needed.

As shown in Table 3, the vulcanization condition data D3 includes the vulcanization temperature and the type of vulcanization time. If the vulcanization condition data D3 of the learning data is unified and the blending data D1 is understood under these unified vulcanization conditions, the vulcanization condition data D3 can be omitted.

In addition to the above-mentioned input data, the calculation device 2 is input with the number of times (the desired number of times) to generate the random noise Z. The calculation device 2 uses a compounding generation model GM constructed based on the input constraint condition data Dy and the target value D2a of the rubber physical property data D2 to generate compounding data D1 that is determined to satisfy the constraints of the constraint condition data Dy and to satisfy the target value D2a of the desired multiple types of rubber physical property data D2, the number of times the random noise Z is generated. Each of the generated compounding data D1 is compounding data D1 that is different from the learning data.

Then, the compounding data D1 of each of the generated candidates and the vulcanization condition data D3 of the target vulcanized rubber Ga (the vulcanization condition data D3 when vulcanizing the target unvulcanized rubber Ra) are input to the prediction model PM, and a calculation process is performed to execute the prediction model PM. The prediction model PM predicts and calculates the rubber property data D2 of the target vulcanized rubber Ga formed with the compounding data D1 based on the input compounding data D1 and vulcanization condition data D3.

The calculated rubber physical property data D2 is compared with the target value D2a set for each data, and the calculation device 2 judges whether the data falls within the tolerance range set as being close. This tolerance range can be set arbitrarily, but is, for example, a range of about ±5% or ±2% of the target value D2a. From all the generated candidate compounding data D1, the compounding data D1 for which the rubber physical property data D2 is judged to be close to the target value D2a is selected by the calculation device 2. The compounding data D1 selected is not limited to one, and may be multiple.

As shown in Table 4 below, the compounding data D1 of the selected target vulcanized rubber Ga is output to the output 4. If the output unit 4 is a monitor, the compounding data D1 is displayed, and if the output unit 4 is a printer, the compounding data D1 is printed.

The compounding data D1 listed in Table 4 satisfies the constraint condition Dy, and the target vulcanized rubber Ga formed with this compounding data D1 and vulcanized under the vulcanization conditions of the input vulcanization condition data D3 is predicted to satisfy the target values D2a of each of the rubber physical property data D2 listed in Table 1.

The data shown in Tables 1 to 4 are based on actual examples, and the target unvulcanized rubber Ra with the compounding data D1 in Table 4 was vulcanized based on the vulcanization condition data D2 in Table 3 to produce the target vulcanized rubber Ga. The rubber property data D2 of the target vulcanized rubber Ga produced was measured, and the results of comparing the actual measured values with the target values D2a shown in Table 1 are shown in Table 5. As shown in Table 5, it was confirmed that by using the present invention, it is possible to determine compounding data D1 for the target vulcanized rubber Ga that generally satisfies multiple desired rubber properties.

As described above, in the present invention, the candidate combination data D1 is generated under the constraints set by the constraint condition data Dy. Therefore, the computational burden for generating the candidate combination data D1 is small. In addition, since the candidate combination data D1 generated does not exist in the learning data, new combination data D1 can be identified.

In addition, by performing calculations using the prediction model PM described above based on the compounding data D1 of each candidate generated using the compounding generation model GM and the vulcanization condition data D3 of the target vulcanized rubber Ga, the rubber physical property data D2 of each of the target vulcanized rubbers Ga can be calculated with high accuracy. As a result, it is possible to easily grasp the compounding data D1 of the target vulcanized rubber Ga in which each of the rubber physical property data D2 closely approximates the target value D2a with high accuracy. Moreover, since the compounding data D1 that can be grasped is new compounding data D1 that does not exist in the learning data, it greatly contributes to the efficiency of development and improvement work of vulcanized rubber having the target rubber physical properties.

When generating candidates for each of the compounding data D1, it is desirable to correct the total compounding amount of the rubber components in the compounding data D1 to 100 phr (100 parts by mass). In Table 4, the total compounding amount of the rubber components (NR, BR) is corrected to 100 phr, and the compounding amounts of each of the other compounding components C are listed. By correcting the compounding data D1 in this way when generating candidates for the compounding data D1, it becomes easier to grasp the compounding amount of each compounding component C.

When there are multiple compounding data D1 for the target vulcanized rubber Ga, each of which is selected because its rubber physical property data D2 is determined to be close to the target value D2a, the compounding data D1 can be output by the output unit 4 in the following order:

For example, the selected compounding data D1 is output by the output unit 4 in the order of approximation of the rubber property data D2 of the target vulcanized rubber Ga formed by each compounding data D1 to the target value D2a for each type of rubber property data D2. In detail, each rubber property data D2 of the target vulcanized rubber Ga formed by each compounding data D1 is compared with the target value D2a for each type of rubber property data D2. The compounding data D1 of the target vulcanized rubber Ga having the rubber property data D2 with the smallest difference between the two compared is ranked first. The compounding data D1 of the target vulcanized rubber Ga having the rubber property data D2 with the second and third smallest difference between the two are ranked second and third, respectively. This ranking may change depending on the type of rubber property data D2. Each compounding data D1 is output by the output unit 4 in the order ranked in this way. By ranking in this way, it becomes easier to understand the superiority of each blend data D1 by using the degree of approximation of each type of rubber physical property data D2 to the target value D2a as an index.

Alternatively, each selected compounding data D1 is output by the output unit 4 in order of the degree of approximation of each rubber physical property data D2 of the target vulcanized rubber Ga formed with each compounding data D1 determined based on the priority weighting set for the type of rubber physical property data D2 to the target value D2a. For example, a case will be described in which a coefficient k1 with a large priority weighting is set for the rubber physical property data D2 of tan δ@60°C, and a coefficient k2 with a small priority weighting is set for the rubber physical property data D2 of hardness HC@20°C. The coefficients k1 and k2 mean that the higher the numerical value, the higher the priority. Then, when the rubber physical property data D2 of tan δ@60°C is an approximation coefficient n1 indicating the degree of approximation to the target value D2a, and the rubber physical property data D2 of hardness HC@20°C is an approximation coefficient n2 to the target value D2a, the evaluation of the degree of approximation of the target vulcanized rubber Ga having these rubber physical property data D2 is calculated by the coefficient k1 x approximation coefficient n1 + coefficient k2 x approximation coefficient n2. The larger the approximation coefficients n1 and n2 are, the closer they are to the target value D2a. Then, each selected compounding data D1 is ranked in order of the degree of approximation of the target vulcanized rubber Ga formed by each compounding data D1, and output by the output unit 4. By ranking in this way, it becomes easier to grasp the superiority of each compounding data D1 taking into account the weighting of the priority of each rubber physical property data D2.

The selected combination data D1 can also be output by the output unit 4 in order of lowest material cost per unit weight calculated based on each combination data D1. In detail, the cost per unit weight of each combination component C is grasped, and the material cost per unit weight of each combination data D1 is calculated based on the grasped cost of each combination component C and the combination amount of each combination component C in each combination data D1. Based on this calculation result, each combination data D1 is ranked and output to the output unit 4. By ranking in this way, it becomes easier to grasp the superiority of the material cost of each combination data D1.

Reference Signs List 1 Rubber compound design system 2 Calculation device 3 Input unit 4 Output unit 5 Measurement device PM Prediction model GM Compound generation model D1 Compounding data Dy Constraint condition data D2 Rubber property data D2a Target value D3 Vulcanization condition data R Unvulcanized rubber Ra Target unvulcanized rubber G Vulcanized rubber Ga Target vulcanized rubber C (C1, C2, ... Cn) Compounding components

Claims (7)

In a method for designing a rubber compound, compounding data is obtained in which compounding amounts of each compounding component of a vulcanized rubber that satisfies target values of multiple types of rubber property data are set,
The rubber physical property data of each of the vulcanized rubbers produced by vulcanizing a large number of types of unvulcanized rubbers having different compounding data is obtained, and each of the compounding data, each of the obtained rubber physical property data, and each of the vulcanized rubbers' vulcanization condition data are used as learning data,
A computing device performs computational processing using a blend generation model constructed using a generative adversarial network based on the learning data under constraint conditions set by predetermined constraint condition data, thereby generating a plurality of candidates for blend data that are determined to satisfy the constraints set by the constraint condition data and to satisfy the target values of each of the rubber physical property data, and that are different from the blend data of the learning data;
A rubber compounding design method, characterized in that: using a predictive model constructed by machine learning based on the learning data, calculations are performed by the arithmetic device based on the generated compounding data and vulcanization condition data of each of the candidates, thereby calculating the rubber property data of each of the target vulcanized rubbers formed from the compounding data of each of the candidates and vulcanized based on the vulcanization condition data; and selecting from each of the candidates the compounding data of the target vulcanized rubber for which the calculated rubber property data of each of the target vulcanized rubbers approximates the target value. The rubber compounding design method according to claim 1, wherein the constraint data is the number of compounding components or the number of types of rubber components in the compounding components. The rubber compounding design method according to claim 1 or 2, wherein when a plurality of compounding data are selected, each of the selected compounding data is output by an output unit for each type of rubber physical property data in order of approximation of the rubber physical property data of the target vulcanized rubber formed with each compounding data to the target value. The rubber compounding design method according to claim 1 or 2, wherein when a plurality of compounding data are selected, the selected compounding data are output by an output unit in order of the degree of approximation of the rubber physical property data of the target vulcanized rubber formed with each compounding data, which is determined based on the priority weighting set for the type of rubber physical property data, to the target value. The rubber compounding design method according to claim 1 or 2, wherein when a plurality of compounding data are selected, each of the selected compounding data is output by an output unit in order of lowest material cost per unit weight calculated based on each of the compounding data. The rubber compounding design method according to any one of claims 1 to 5, wherein the total compounding amount of the rubber components in the compounding data is corrected to 100 phr when generating the candidates for the compounding data. A rubber compounding design system includes a calculation device, an input unit for inputting data to the calculation device, and an output unit communicably connected to the calculation device, and obtains compounding data in which compounding amounts of compounding components of vulcanized rubber that satisfy target values of multiple types of rubber physical property data are set,
The rubber property data of each of the vulcanized rubbers produced by vulcanizing a large number of types of unvulcanized rubbers having different compounding data, the compounding data of each of the vulcanized rubbers, and the vulcanization condition data of each of the vulcanized rubbers are used as learning data,
A calculation device performs calculation processing using a compounding generation model constructed using a generative adversarial network based on the learning data under constraint conditions set by predetermined constraint condition data, thereby generating a plurality of candidates for compounding data that are determined to satisfy the constraints set by the constraint condition data and to satisfy the target values of each of the rubber physical property data, and that are different from the compounding data of the learning data;
A rubber blend design system characterized in that, using a predictive model constructed by machine learning based on the learning data, the arithmetic device performs calculations based on the compounding data and vulcanization condition data of each of the generated candidates, thereby calculating the rubber property data of each of the target vulcanized rubbers formed from the compounding data of each of the candidates and vulcanized based on the vulcanization condition data, and the compounding data of the target vulcanized rubber for which the calculated rubber property data of each of the target vulcanized rubbers approximates the target value is selected from each of the candidates and output by the output unit.

Images