JP7577934B2 - Physical property prediction device, physical property prediction method, and manufacturing method - Google Patents

Description

Applicable under Article 30, Paragraph 2 of the Patent Act. Published on April 13, 2020 on the website of the Research Association for Advanced Material Rapid Development Technology (ADMAT) (http://www.admat.or.jp/news) and four other websites.

The present invention relates to a polymer property prediction device, a property prediction method, and a manufacturing method.

In recent years, AI (Artificial Intelligence) technology has come to be used in a variety of technical fields. In the fields of chemistry and materials, AI technology is applied as materials informatics (MI) and is used in the research and development of new substances and materials.

For example, when synthesizing a polymer using multiple monomers, AI technology can be used to explore what ratio of these multiple monomers is necessary to synthesize a polymer with desired physical properties. The structural characteristics of compounds such as monomers are expressed according to any expression method suitable for calculation processing, such as ECFP (Extended Connectivity Fingerprint), and a regression model can be constructed in which the ECFP of each monomer and its content ratio are explanatory variables and a certain physical property of the polymer is the objective variable. By using the constructed regression model, it becomes possible to predict the physical properties of a polymer synthesized from the unknown content ratio of the monomers.

Sometimes it is necessary to evaluate a polymer based on multiple physical properties. In this case, it is possible to construct a regression model with each physical property as a response variable and predict multiple types of physical property values from the monomer content ratio. When predicting such multiple types of physical property values, for example, Bayesian optimization can be used.

David Rogers and Mathew Hahn, "Extended-Connectivity Fingerprints". J. Chem. Inf. Model, 50, 742 (2010) Fabian Pedregosa, et. al., "Scikit-learn: Machine Learning in Python", JMLR. 12, 2825 (2011)

However, the predicted values for each physical property generally have different scales or numerical ranges, and in conventional Bayesian optimization, in order to comprehensively evaluate a polymer based on multiple physical properties, it was necessary to rescale each objective variable based on the researcher's experience, etc.

In order to solve the above problems, the objective of the present invention is to provide a method for evaluating the physical properties of a polymer based on multiple physical property values.

[1] A property prediction device having a property prediction unit that obtains predicted values of multiple property types of a polymer using a model for each property type based on monomer information indicating multiple monomers and their content ratios, and a monomer determination unit that calculates a deviation value for the monomer information from the predicted values of the multiple property types and determines the monomers used in the synthesis of the polymer and their content ratios based on the deviation value.

[2] The property prediction device described in [1], in which the model for each property type is derived by Gaussian process regression using training data consisting of pairs of the monomer information and the property values of the polymer property type corresponding to the monomer information.

[3] The property prediction device according to [1] or [2], wherein the monomer determination unit calculates an output value of an acquisition function of Bayesian optimization from the predicted values of the plurality of property types and calculates a deviation value of the output value.

[4] The property prediction device described in any one of [1] to [3], wherein the monomer determination unit determines the monomer information for which the deviation value is maximum or minimum as the monomers and their content ratios used in the synthesis of the polymer.

[5] Further comprising a monomer information generating unit that generates the monomer information,
The property prediction device according to any one of [1] to [4], wherein the monomer information is expressed by a product of an ECFP (Extended Connectivity Fingerprint) of each monomer and a molar ratio.

[6] A method for predicting physical properties, comprising the steps of: one or more processors obtaining predicted values for multiple physical property types of a polymer using a model for each physical property type based on monomer information indicating multiple monomers and their content ratios; and the one or more processors calculating a deviation value for the monomer information from the predicted values for the multiple physical property types, and determining the monomers used in synthesizing the polymer and their content ratios based on the deviation value.

[7] The method for predicting physical properties described in [6], in which the model for each physical property type is derived by Gaussian process regression using training data consisting of pairs of the monomer information and the physical property values of the physical property type of the polymer corresponding to the monomer information.

[8] The physical property prediction method described in [6] or [7], in which the determining step calculates an output value of an acquisition function of Bayesian optimization from the predicted values of the multiple physical property types, and calculates a deviation value of the output value.

[9] The method for predicting physical properties described in any one of [6] to [8], in which the determining step determines the monomer information for which the deviation value is maximum or minimum as the monomers and their content ratios used in the synthesis of the polymer.

[10] The method further comprises the step of generating the monomer information,
The method for predicting physical properties according to any one of [6] to [9], wherein the monomer information is expressed by a product of an ECFP (Extended Connectivity Fingerprint) of each monomer and a molar ratio.

[11] A method for producing a polyurethane cured product by determining monomers and their content ratios using the method for predicting physical properties according to any one of [6] to [10].

The above aspect of the present invention provides a method for evaluating the physical properties of a polymer based on multiple physical property values.

1 is a schematic diagram illustrating a regression model generating device and a property predicting device according to an embodiment of the present invention. 1 is a block diagram showing a hardware configuration of a regression model generating device and a property predicting device according to an embodiment of the present invention. FIG. 1 is a block diagram showing a functional configuration of a regression model generating device according to an embodiment of the present invention. FIG. 2 illustrates an ECFP according to one embodiment of the present invention. 1 is a flowchart illustrating a regression model generation process according to an embodiment of the present invention. 1 is a block diagram showing a functional configuration of a property prediction apparatus according to an embodiment of the present invention. FIG. 13 is a diagram showing predicted values of multiple types of physical properties according to an embodiment of the present invention. FIG. 13 illustrates an index according to one embodiment of the present invention. FIG. 2 is a diagram showing a diol and its content ratio and a catalyst that maximizes the index according to an embodiment of the present invention. FIG. 10 is a diagram showing the measured values of reduced transmittance, breaking stress, and elongation of a cured polyurethane product obtained by curing a polyurethane precursor synthesized based on the diols and their contents and catalysts shown in FIG. FIG. 1 is a diagram showing diols and their content ratios, and catalysts of Comparative Examples (1) to (10). FIG. 12 is a graph showing the measured values of reduced transmittance, breaking stress, and elongation of a cured polyurethane product obtained by curing a polyurethane precursor synthesized based on the diols and their contents and catalysts shown in FIG. 11 . 1 is a flowchart showing a property prediction process according to an embodiment of the present invention.

The present invention will now be described in detail with reference to examples.

In the embodiment described below, a regression model generation device is disclosed that generates a regression model for each physical property type to predict the physical property values of each type of polymer from the monomers used in the synthesis of the polymer and their content ratio, and a physical property prediction device that uses the generated regression model for each physical property type to predict the physical property values of each type from the monomers and their content ratio, and calculates an index for comprehensively evaluating the physical properties of the polymer from the predicted physical property values for each type.

[Summary]
First, an outline of a regression model generating device and a physical property predicting device according to an embodiment of the present invention will be described with reference to Fig. 1. Fig. 1 is a schematic diagram showing a regression model generating device and a physical property predicting device according to an embodiment of the present invention.

As shown in FIG. 1, the regression model generation device 100 acquires from a substance database 50 teacher data or training data consisting of pairs of monomer information indicating the monomers and their content ratios for synthesizing a polymer, and physical property values of multiple types of polymers (N types in the illustrated example) synthesized based on the monomer information. The regression model generation device 100 then constructs regression models for each type (estimated regression models #1, #2, ..., #N) that predict the physical property values of each physical property type from the acquired monomer information, and generates regression models for each type for all types.

Next, the property prediction device 200 obtains a predicted value for each property type from the monomer information using the regression models (regression models #1, #2, ..., #N) constructed in advance by the regression model generation device 100, and calculates an index for evaluating the overall property of the polymer from the obtained multiple predicted values. The property prediction device 200 calculates an index for each monomer information, and determines the monomers and their content ratios (such as molar ratios) used in the synthesis of the polymer based on the index. For example, a deviation value may be used as the index, and the monomer and its content ratio for which the calculated deviation value is maximum or minimum may be selected.

The regression model generating device 100 and the property prediction device 200 may have a hardware configuration as shown in FIG. 2, for example. That is, the regression model generating device 100 and the property prediction device 200 have a drive device 101, an auxiliary storage device 102, a memory device 103, a CPU (Central Processing Unit) 104, an input/output device 105, and a communication device 106, which are interconnected via a bus B.

Various computer programs, including programs for implementing various functions and processes described below in the regression model generating device 100 and the physical property prediction device 200, may be provided by a recording medium 107 such as a CD-ROM (Compact Disk-Read Only Memory). When the recording medium 107 storing the program is set in the drive device 101, the program is installed from the recording medium 107 to the auxiliary storage device 102 via the drive device 101. However, the program does not necessarily have to be installed by the recording medium 107, and may be downloaded from any external device via a network or the like. The auxiliary storage device 102 stores the installed program as well as necessary files and data. When an instruction to start the program is given, the memory device 103 reads out and stores the program and data from the auxiliary storage device 102. The CPU 104, functioning as a processor, executes various functions and processes of the regression model generating device 100 and the property prediction device 200, as described below, in accordance with various data such as programs stored in the memory device 103 and parameters necessary to execute the programs. The input/output device 105 provides an interface between the regression model generating device 100 and the property prediction device 200 and the user, and may be composed of, for example, input devices such as a keyboard, mouse, touch screen, and output devices such as a display and speaker. The communication device 106 executes various communication processes for communicating with external devices.

However, the regression model generating device 100 and the property prediction device 200 are not limited to the hardware configuration described above, and may be realized by any other appropriate hardware configuration, such as a processing circuit wired to realize the functions and processing described below.

[Regression model generation device]
Next, a regression model generating device 100 according to an embodiment of the present invention will be described with reference to Fig. 3. Fig. 3 is a block diagram showing the functional configuration of the regression model generating device 100 according to an embodiment of the present invention.

As shown in FIG. 3, the regression model generation device 100 has a training data acquisition unit 110 and a regression model generation unit 120.

The training data acquisition unit 110 acquires teacher data or training data consisting of pairs of monomer information and multiple types of physical property values of a polymer synthesized based on the monomer information. The monomer information may indicate, without limitation, multiple monomers used to synthesize the polymer and their content ratios.

For example, when a diol and a diisocyanate are polyaddition-reacted using a catalyst to synthesize polyurethane, the monomers used in the monomer information may be various diols, and the content ratio may be the molar ratio of the diols used. Here, as shown in FIG. 4, each diol may be quantified according to any appropriate expression method such as ECFP (Extended Connectivity Fingerprint) based on the structural information extracted from the diol. The physical property type may be a film characteristic such as the converted transmittance, breaking stress, or elongation of polyurethane. Instead of the molar ratio, the number of moles may be given as training data. In this case, the sum of the molar ratios of the monomers is normalized to 1, and the molar ratio is calculated from the number of moles.

Here, the converted transmittance is calculated by the formula: 4×(refractive index)/[(refractive index)+1] ² (where refractive index ≧1).
It is calculated as follows.

For example, the monomer information of sample 1 in the training data may indicate diol 1, diol 2, and diol 3 as monomers, with a molar ratio of "0.41:0.01:0.11." Furthermore, the physical property values of polyurethane synthesized based on the monomer information of sample 1 may be reduced transmittance "0.971," breaking stress "20.1," and elongation "129.2."

Here, diols 1 to 3 are as follows:

同様に、サンプル２の単量体情報は、単量体としてジオール１、ジオール２及びジオール３を示し、モル比として"０．４２：０．０１：０．１０"を示すものであってもよい。また、サンプル２の単量体情報に基づき合成されたポリウレタンの物性値は、換算透過率"０．９７３"、破断応力"３１．２"及び伸び"３００．３"であってもよい。

Similarly, the monomer information of Sample 2 may indicate diol 1, diol 2, and diol 3 as monomers, with a molar ratio of "0.42:0.01:0.10." Furthermore, the physical property values of polyurethane synthesized based on the monomer information of Sample 2 may be reduced transmittance "0.973," break stress "31.2," and elongation "300.3."

Each sample of such training data is acquired in advance through experiments or the like, and is stored, for example, in the substance database 50. The training data acquisition unit 110 acquires the training data from the substance database 50 and passes it to the regression model generation unit 120.

The regression model generation unit 120 generates a regression model for each type from the training data. Specifically, the regression model generation unit 120 constructs a regression model for each type with the monomer information of the training data provided by the training data acquisition unit 110 as the explanatory variable and the physical property value of each physical property type as the objective variable.

For example, when the above-mentioned samples 1, 2, ... are given, the regression model generation unit 120 first quantifies the monomers, which are non-quantified data items. For example, the regression model generation unit 120 may convert the monomers into a vector representation according to ECFP. Specifically, the regression model generation unit 120 can extract the three partial structures shown in the center from the monomer structure shown on the left side of FIG. 4 according to a known ECFP derivation method, and obtain the vector representation shown on the right side.

After calculating the ECFP of each monomer in this manner, the regression model generating unit 120 multiplies the molar ratio of each sample by the vector of each ECFP. For example, when diol 1 is expressed by _X1 = ( _X11 , _X12 , ...), diol 2 is expressed by _X2 = ( _X21 , _X22 , ...), and diol 3 is expressed by _X3 = ( _X31 , _X32 , ...), the regression model generating unit 120 calculates ( _0.41X11 , _0.41X12 , ..., _0.01X21 , _0.01X22 , ..., _0.11X31 , _0.11X32 , ...) for the molar ratio of sample 1 "0.41:0.01:0.11". Similarly, the regression model generating unit 120 calculates (0.42X ₁₁ , 0.42X ₁₂ , . . . , 0.01X ₂₁ , 0.01X ₂₂ , . . . , 0.10X ₃₁ , 0.10X ₃₂ , . . . ) for the molar ratio of sample 2, which is “0.42:0.01:0.10”.

Next, the regression model generating unit 120 generates subsample 1-1" monomer information: (0.41X ₁₁ , 0.41X ₁₂ , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.11X ₃₁ , 0.11X ₃₂ , ...); physical property value: converted transmittance 0.971, breaking stress 20.1, elongation 129.2" for each physical property type from sample 1" monomer information: (0.41X ₁₁ , 0.41X ₁₂ , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.11X ₃₁ , 0.11X ₃₂ , ...); physical property value: converted transmittance 0.971, breaking stress _20.1 , elongation _129.2 " , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.11X ₃₁ , 0.11X ₃₂ , ...); physical property: break stress 20.1", and subsample 1-3" monomer information: (0.41X ₁₁ , 0.41X ₁₂ , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.11X ₃₁ , 0.11X ₃₂ , ...); physical property: elongation 129.2".

Similarly, the regression model generating unit 120 generates sub-sample 2-1" monomer information: (0.42X ₁₁ , 0.42X ₁₂ , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.10X ₃₁ , 0.10X ₃₂ , ...); physical property values: converted transmittance 0.973, breaking stress 31.2, elongation 300.3" for each physical property type from sample 2" monomer information: (0.42X ₁₁ , 0.42X ₁₂ , ..., 0.01X ₂₁ , 0.01X ₂₂ , ..., 0.10X ₃₁ , 0.10X ₃₂ , ...); physical property values: converted transmittance 0.973, breaking stress _31.2 , elongation _300.3 ". , ..., _0.01X21 , _0.01X22 , ..., _0.10X31 , _0.10X32 , ...); physical property: breaking stress 31.2", and subsample 2-32-1" monomer information: ( _0.42X11 , _0.42X12 , ..., _0.01X21 , _0.01X22 , ..., _0.10X31 , _0.10X32 , ...); physical property: elongation "300.3". In a similar manner, the regression model generation unit 120 generates subsamples i-1, i-2, and i-3 for each sample i.

Then, the regression model generation unit 120 extracts subsamples 1-1, 2-1, ... and performs regression analysis on these subsamples i-1 in order to construct a regression model #1 for each type with the monomer information as the explanatory variable and the converted transmittance as the objective variable. Specifically, the regression model generation unit 120 constructs a regression model #1 for each type so as to reproduce subsamples 1-1, 2-1, ... according to Gaussian process regression.

Here, Gaussian process regression is one of the models that estimates a function y _i =f(x _i ) from an input variable x to a real-valued output variable y, and one of its features is its nonlinearity, making it effective even when linear regression cannot provide a good fit. Another important feature is the use of Bayesian estimation. The estimated function is not a single function, but is obtained as a distribution of functions, making it possible to express the uncertainty of the estimation.

The relationship between each _yi is described by the covariance. The normal distribution of each _yi and the covariance describing the relationship between each _yi are obtained from the training data. Therefore, since there is a normal distribution of the value of _yi with respect to the uncertainty, for example, if a 95% confidence interval is taken, it is possible to predict the value of _yi with an uncertainty of 2σ.

In the same manner, the regression model generation unit 120 extracts subsamples 1-2, 2-2, ... and performs regression analysis on these subsamples i-2 in order to construct a regression model #2 for each type with monomer information as the explanatory variable and rupture stress as the objective variable. Specifically, the regression model generation unit 120 constructs a regression model #2 for each type so as to reproduce subsamples 1-2, 2-2, ... according to Gaussian process regression.

Furthermore, the regression model generation unit 120 extracts subsamples 1-3, 2-3, ... and performs regression analysis on these subsamples i-3 in order to construct a regression model #3 for each type with monomer information as the explanatory variable and elongation as the objective variable. Specifically, the regression model generation unit 120 constructs a regression model #3 for each type so as to reproduce subsamples 1-3, 2-3, ... according to Gaussian process regression.

In this way, after constructing regression models #1, #2, and #3 for each type, the regression model generation unit 120 passes the constructed regression models #1, #2, and #3 to the property prediction device 200.

[Regression model generation process]
Next, a regression model generation process according to an embodiment of the present invention will be described with reference to Fig. 5. The regression model generation process may be executed by, for example, the above-described regression model generation device 100, and more specifically, may be realized by a processor of the regression model generation device 100. Fig. 5 is a flowchart showing the regression model generation process according to an embodiment of the present invention.

As shown in FIG. 5, in step S101, the regression model generating device 100 acquires training data. For example, when polyurethane is synthesized as a polymer, the training data is composed of pairs of monomer information indicating monomers (e.g., various diols, etc.) and their content ratios (e.g., number of moles, molar ratio, etc.) and multiple types of physical property values of the polymer (e.g., reduced transmittance, breaking stress, elongation, etc.), and is acquired in advance by experiments, etc.

In step S102, the regression model generating device 100 calculates the partial structure and the molar ratio from the monomer information for each sample of the acquired training data. Specifically, the regression model generating device 100 calculates the ECFP of each monomer in the monomer information and converts each monomer into a vector representation. In addition, the regression model generating device 100 calculates the molar ratio from the number of moles.

In step S103, the regression model generating device 100 calculates explanatory variable values of the regression model from the calculated partial structure and molar ratio. Specifically, the regression model generating device 100 calculates a product vector of the ECFP and molar ratio calculated in step S102, and sets the calculated product vector as the explanatory variable value of the regression model.

In step S104, the regression model generating device 100 associates the explanatory variable values calculated in step S103 with the physical property values of each type of training data acquired in step S101, and constructs a regression model for each type.

Specifically, the regression model generating device 100 associates the vector of each sample of monomer information as the explanatory variable value calculated in step S104 with the converted transmittance for that sample, and constructs a regression model #1 for each type that predicts the converted transmittance from the vector of monomer information.

The regression model generating device 100 also associates the vector of each sample of monomer information as the explanatory variable value calculated in step S104 with the rupture stress for that sample, and constructs a regression model #2 for each type that predicts the rupture stress from the vector of monomer information.

Furthermore, the regression model generating device 100 associates the vector of each sample of monomer information as the explanatory variable value calculated in step S104 with the elongation for that sample, and constructs a regression model #3 for each type that predicts elongation from the vector of monomer information.

In one embodiment, the regression models #1, #2, and #3 for each type may be constructed using Gaussian process regression, as described above.

In one embodiment, the regression models #1, #2, and #3 for each type may use the type of catalyst as an explanatory variable in addition to the monomer information. Specifically, the type of catalyst may be quantified using any appropriate method, such as binary notation, and added to the explanatory variables of the regression models #1, #2, and #3 for each type.

After constructing the regression models #1, #2, and #3 for each type, the regression model generation device 100 provides the constructed regression models #1, #2, and #3 for each type to the property prediction device 200.

In the above-described embodiment, the regression model for each type was constructed according to Gaussian process regression, but the regression model for each type according to the present invention is not limited to this and may be constructed according to any other appropriate regression analysis method, such as multiple regression analysis.

[Physical property prediction device]
Next, a physical property prediction device 200 according to an embodiment of the present invention will be described with reference to Figures 6 to 8. The physical property prediction device 200 uses the regression models for each type generated by the regression model generation device 100, calculates predicted values for multiple physical property types of a polymer from monomer information as test data, and evaluates the overall physical properties of the polymer based on the calculated predicted values of multiple physical properties. Figure 6 is a block diagram showing the functional configuration of the physical property prediction device 200 according to an embodiment of the present invention.

As shown in FIG. 6, the property prediction device 200 has a monomer information generation unit 210, a property prediction unit 220, and a monomer determination unit 230.

The monomer information generation unit 210 generates monomer information. Specifically, the monomer information generation unit 210 comprehensively generates combinations of monomers and the content ratios of each monomer as test data for the monomer information, and passes the generated monomer information to the property prediction unit 220.

For example, if the polymer is polyurethane, the monomer information generating unit 210 selects a predetermined number of monomers from a number of candidate monomers, and further sets the molar ratio of each monomer in the combination of the selected monomers. Here, the monomer information may be expressed by setting the molar ratio of unselected monomers to 0, or the monomer information may be expressed only by the molar ratio of the selected monomers.

The molar ratio may also be comprehensively selected according to the constraint conditions. For example, the molar ratio may be comprehensively set in increments of 0.01 so that the sum of the molar ratios of each diol and diisocyanate is 1.0. However, the molar ratio may be selected so that (total number of moles of OH groups in diol)/(total number of moles of NCO groups in diisocyanate)=1.1 to 1.2. However, the present invention is not limited to such constraint conditions, and appropriate constraint conditions may be set depending on the embodiment.

Then, the monomer information generation unit 210 calculates the ECFP of the selected monomer, multiplies the vector calculated by the ECFP by the content ratio, and passes the calculated product vector to the property prediction unit 220 as test data.

The physical property prediction unit 220 inputs monomer information indicating multiple monomers and their content ratios into a model for each physical property type, and obtains predicted values for multiple physical property types. Specifically, the physical property prediction unit 220 inputs vectors representing the monomer information generated as test data by the monomer information generation unit 210 into the regression model for each type as explanatory variable values, and obtains corresponding physical property predicted values from each regression model for each type as objective variable values. For example, when three regression models for each type corresponding to three objective variables, reduced transmittance, breaking stress, and elongation, are available, the physical property prediction unit 220 obtains the predicted value of reduced transmittance, predicted value of breaking stress, and predicted value of elongation for each vector.

In one embodiment, the physical property prediction unit 220 uses an acquisition function for Bayesian optimization. Specifically, the physical property prediction unit 220 may input the predicted value of each physical property type to the acquisition function and determine whether the physical property is good or bad based on the output value of the acquisition function. In this embodiment, the acquisition function a may be of the UCB (Upper Confidence Bound) type,
a(m,s)=m+2×s
may be used. Here, m is the predicted value, and s is the standard deviation of each sample of the test data indicating the uncertainty of the prediction. Gaussian process regression may be used to obtain the predicted value m and the uncertainty s. Gaussianprocessregressor of scikit-learn may be used to execute the Gaussian process regression. The kernel is not particularly limited, but may be a product of ConstantKernel and Matern plus WhiteKernel. When Gaussian process regression is used, a predicted distribution is output for each sample. The mean and standard deviation of the predicted distribution may be referred to as the predicted value m and the uncertainty s, respectively. The prediction uncertainty s is used when calculating the acquisition function a of each physical property.

Note that the acquisition function a is not limited to the above, and may be an EI (Expected Improvement) type or a PI (Probability of Improvement) type.

For example, as shown in FIG. 7(a), when test data 1 has diols 1, 2, and 3 as monomers with a content ratio of 0.41:0.01:0.11, the physical property prediction unit 220 uses a regression model to obtain a predicted value of 0.41 for the converted transmittance, a predicted value of 33 for the breaking stress, and a predicted value of 500 for the elongation, as shown in FIG. 7(b). Furthermore, the physical property prediction unit 220 inputs each predicted value into the acquisition function a, and obtains an output value of 0.49 for the converted transmittance, an output value of 33.6 for the breaking stress, and an output value of 504 for the elongation.

Similarly, when test data 2 has diols 1, 2, and 3 as monomers with a content ratio of 0.42:0.01:0.10, the property prediction unit 220 uses a regression model to obtain a predicted converted transmittance value of 0.42, a predicted breaking stress value of 21, and a predicted elongation value of 501, as shown in FIG. 7(b). Furthermore, the property prediction unit 220 inputs each predicted value into acquisition function a, and obtains an output value of converted transmittance of 0.48, an output value of breaking stress of 21.4, and an output value of elongation of 507.

The advantage of using such an acquisition function is that it is easier to reach the global optimum solution. It is known that in the search for the optimum solution based on the predicted value, it is easy to recommend data around the data with good performance in the past data (teaching data) as candidates, and there is a possibility of falling into a local optimum solution. If the truly best performing data is different from the known teaching data, the search for the optimum solution based on the predicted value will be trapped in a local optimum solution, making it difficult to reach the global optimum solution. On the other hand, by using an acquisition function, it is possible to prevent being trapped in a local optimum solution. In the acquisition function, not only the predicted value m but also the prediction uncertainty s is taken into account. The prediction uncertainty s has the tendency to become larger the more different the features of the data are from the teaching data. Therefore, by recommending as candidates those with large predicted values m and prediction uncertainty s, it is possible to recommend "things that look good and have not been explored", and it is possible to efficiently find the optimal one from various variations, making it easier to reach the global optimum solution.

The property prediction unit 220 passes the predicted values and/or acquisition function values for each property type calculated for each test data to the monomer determination unit 230.

The monomer determination unit 230 calculates the deviation value for the content ratio from the predicted values of multiple physical property types, and determines the content ratio of the monomers used to synthesize the polymer to be synthesized based on the deviation value. Specifically, the monomer determination unit 230 first indexes the predicted values and/or acquisition function values of each physical property type obtained from the physical property prediction unit 220 using the deviation value in order to evaluate the overall physical properties of the polymer. For example, when indexing the physical property predicted values using the standard deviation of the acquisition function values, the monomer determination unit 230

によって偏差値Ｉ_ｉを算出する。ここで、Ｎは物性種別数であり、ｉはテストデータのインデックスであり、ｊは物性種別のインデックスであり（１≦ｊ≦Ｎ）、Ｘ_ｊは訓練データにおける物性ｊの平均であり、σ_ｊは訓練データにおける物性ｊの標準偏差であり、ｘ_ｉｊは物性ｊのテストデータｉに対する獲得関数値である。

Here, N is the number of physical property types, i is the index of the test data, j is the index of the physical property type (1≦j≦N), _Xj is the average of the physical property _j in the training data, _σj is the standard deviation of the physical property j in the training data, and _xij is the acquisition function value of the physical property j for the test data i.

For example, for the above-mentioned test data 1 and 2, the monomer determination unit 230 can calculate indices with standard deviations of 0.64 and 0.71, respectively, as shown in FIG. 8.

The deviation value is not limited to being calculated based on the acquisition function value, but may be calculated from a predicted value.

Once the deviation value for evaluating the overall physical properties of the polymer is obtained in this manner, the monomer determination unit 230 determines the monomers and their content ratios for producing the polymer based on the obtained deviation value. For example, if the physical properties can be evaluated to be better the larger the various physical property values used to calculate the deviation value, the monomer determination unit 230 may determine that the polymer produced based on the monomers with the largest deviation value and their content ratio has the best physical properties. On the other hand, if the physical properties can be evaluated to be better the smaller the various physical property values used to calculate the deviation value, the monomer determination unit 230 may determine that the polymer synthesized based on the monomers with the smallest deviation value and their content ratio has the best physical properties.

The monomer determination unit 230 outputs the monomer with the maximum or minimum deviation value and its content ratio, or the index of the corresponding test data (see, for example, FIG. 9).

Here, FIG. 9(a) shows the diol and its content ratio and catalyst that maximize the deviation value. Here, diol 4 is 2-methyl-2,4-pentanediol, and catalyst 1 is a silica-supported molybdenum catalyst. Also, FIG. 9(b) shows the predicted values, prediction uncertainty, and acquisition function of the reduced transmittance, breaking stress, and elongation of a polyurethane cured product obtained by curing a polyurethane precursor synthesized based on the diol and its content ratio and catalyst in FIG. 9(a). Furthermore, FIG. 9(c) shows the deviation value calculated based on the acquisition function value in FIG. 9(b).

Figure 10 shows the measured values of the reduced transmittance, breaking stress, and elongation of the cured polyurethane product obtained by curing the polyurethane precursor synthesized based on the diol and its content ratio and catalyst shown in Figure 9(a).

Here, the polyurethane precursor was synthesized as follows, using dicyclohexylmethane 4,4'-diisocyanate as the diisocyanate.

A 100 mL three-neck flask equipped with a Dimroth condenser, thermometer, and stirring blade, placed in an oil bath, was charged with the monomers (diol 1, diol 2, diol 4) in the content ratio shown in Figure 9 (a) and butyl acetate measured so that the final concentration of the entire reaction liquid was 20 mass%, and stirred to dissolve. At this point, it was confirmed that the above monomers had dissolved and become a homogeneous solution. Next, dicyclohexylmethane 4,4'-diisocyanate was measured and added so that the molar ratio of [OH]/[NCO] was 1.2, and the mixture was heated with continued stirring until the liquid temperature reached 90°C.

When the liquid temperature was stabilized at 90°C, 5% by mass of catalyst 1 was added to the substrate (diol and diisocyanate) to initiate the reaction. Every hour, 0.1 mL of the reaction liquid in the system was collected through a Pasteur pipette, filtered through a syringe filter with a pore size of 0.25 μm, and then analyzed by a liquid chromatograph set to GPC (gel permeation chromatogram) mode. Here, the system was heated until the peak derived from the diol disappeared and the elution time of the polyurethane peak top position became constant. After confirming the end point of the reaction by the above operation, the reaction liquid was cooled to room temperature. Next, the catalyst was removed through a membrane filter with a pore size of 50 μm, and then the mixture was poured into 300 mL of n-hexane to precipitate polyurethane. Next, the precipitate was washed with 50 mL of n-hexane, and the solvent was removed in a vacuum drying device to obtain a polyurethane precursor.

Catalyst 1 (silica-supported molybdenum catalyst) was prepared from ammonium molybdate (manufactured by Nippon Inorganic Chemical Industry Co., Ltd.) and silica gel (manufactured by Fuji Silysia Chemical Co., Ltd.).

The reduced transmittance, breaking stress, and elongation of the cured polyurethane were measured as follows.

(Converted refractive index of cured polyurethane)
2 to 2.5 g of polyurethane precursor was placed in a 50 mL glass screw-cap bottle, and then about 6 g of ethyl acetate was added and dissolved by applying to a rotor-type stirring device. Next, in accordance with JIS K1557-6:2009, an isocyanurate-type hexamethylene diisocyanate trimer was added as a 50% by mass ethyl acetate solution so that the [NCO] amount [mol] was the same as the [OH] amount [mol] calculated from the hydroxyl value of the polyurethane precursor previously determined, and mixed for about 1 hour. The solution obtained here was spread in a TPX ^TM resin petri dish, air-dried for 15 hours, and then heated for 10 hours in a thermostatic chamber maintained at 90 ° C. to obtain a polyurethane cured product plate. The obtained polyurethane cured product plate was processed into a shape of 10 mm x 20 mm using a super straight cutter (made by dumbbell) to obtain a test piece.

The refractive index of the cured polyurethane was measured using an Abbe refractometer (Atago, DR-A1-Plus) with 1-bromonaphthalene as the intermediate liquid.

After measuring the refractive index of the cured polyurethane, the refractive index was calculated using the formula: 4×(refractive index)/[(refractive index)+1] ² (wherein the refractive index is ≧1).
The converted transmittance of the cured polyurethane product was calculated based on the above.

(Stress at break and elongation of cured polyurethane)
After obtaining a cured polyurethane plate in the same manner as above, a dumbbell-shaped No. 7 test piece according to JIS K 6251:2010 was prepared using a super straight cutter (manufactured by Dumbbell Co., Ltd.), and the breaking stress and elongation of the cured polyurethane were measured using a tabletop tensile compression tester (manufactured by A&D Co., Ltd., MCT-2150).

Figure 11 shows the diols and their content ratios and catalysts of Comparative Examples (1) to (10). Here, diol 5 is 2-butyl-2-ethyl-1,3-propanediol, catalyst 2 is dibutyltin dilaurate, and OH [mmol] means the number of moles of OH groups in each monomer. Figure 12 also shows the measured values of the reduced transmittance, breaking stress, and elongation of the cured polyurethane synthesized based on the diols and their content ratios in Figure 11.

The polyurethane precursors of Comparative Examples (1) to (9) were synthesized in the same manner as described above, except that dicyclohexylmethane 4,4'-diisocyanate was used as the diisocyanate and catalyst 2 was used as the catalyst.

The polyurethane precursor of Comparative Example (10) was synthesized in the same manner as described above, except that dicyclohexylmethane 4,4'-diisocyanate was used as the diisocyanate.

Furthermore, in [Equation 1], the actual measured value was used instead of x _ij (the acquisition function value) to calculate the index of the actual measured value.

[Property Prediction Processing]
Next, a property prediction process according to an embodiment of the present invention will be described with reference to Fig. 13. The property prediction process may be executed by, for example, the above-described property prediction apparatus 200, and more specifically, may be realized by a processor of the property prediction apparatus 200. Fig. 13 is a flowchart showing the property prediction process according to an embodiment of the present invention.

As shown in FIG. 13, in step S201, the property prediction device 200 acquires explanatory variable values to be input to the regression model for each type. Specifically, the property prediction device 200 generates monomer information as test data to be input to the regression model for each type provided by the regression model generation device 100. For example, the property prediction device 200 selects a predetermined number of monomers from a plurality of monomers, and further comprehensively sets various molar ratios for the selected monomers. For each test data generated in this manner, the property prediction device 200 calculates the ECFP of each monomer, multiplies the calculated ECFP vector by the molar ratio, and calculates a product vector.

In step S202, the property prediction device 200 uses a regression model for each type to calculate a predicted value of each property as a target variable from the monomer information. Specifically, the property prediction device 200 inputs the product vector acquired in step S201 into the regression model for each type and calculates the corresponding property predicted value. At this time, the property prediction device 200 may input the calculated property predicted value into a predetermined acquisition function and calculate an acquisition function value for the predicted value of each property type. For example, if the acquisition function is of UCB type, the uncertainty s may be set as the standard deviation of each sample of the test data.

In step S203, the property prediction device 200 calculates a deviation value as an index for comprehensively evaluating the polymer from the predicted value and/or the acquired function value for each property type. Specifically, when the property prediction value is indexed by the standard deviation of the acquired function value, the property prediction device 200

によって偏差値Ｉ_ｉを算出する。ここで、Ｎは物性種別数であり、ｉはテストデータのインデックスであり、ｊは物性種別のインデックスであり（１≦ｊ≦Ｎ）、Ｘ_ｊは訓練データにおける物性ｊの平均であり、σ_ｊは訓練データにおける物性ｊの標準偏差であり、ｘ_ｉｊは物性ｊのテストデータｉに対する獲得関数値である。

Here, N is the number of physical property types, i is the index of the test data, j is the index of the physical property type (1≦j≦N), _Xj is the average of the physical property _j in the training data, _σj is the standard deviation of the physical property j in the training data, and _xij is the acquisition function value of the physical property j for the test data i.

In step S204, the property prediction device 200 determines the monomers and their content ratios used in the synthesis of the polymer based on the deviation value. Specifically, the property prediction device 200 may output the monomers and their content ratios that maximize or minimize the deviation value, or an index of the test data.

In the above-mentioned examples, the focus was on diol as a monomer of the polyurethane precursor, and only the diol was changed. However, the present invention is not limited to this. For example, the present invention can be applied to cases where the focus is on diisocyanate as a monomer of the polyurethane precursor, only the diisocyanate is changed, or both the diol and the diisocyanate are changed.

In addition, in the above-mentioned embodiment, the physical property type of the cured thermosetting polyurethane is used as the physical property type of the polymer, but the present invention is not limited to this, and can be applied, for example, to cases where the physical property type of not only thermosetting resin but also non-thermosetting resin such as thermoplastic resin is used.

Although the examples of the present invention have been described in detail above, the present invention is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present invention as described in the claims.

50 Material database 100 Regression model generating device 110 Training data acquisition unit 120 Regression model generating unit 200 Physical property prediction device 210 Monomer information generating unit 220 Physical property prediction unit 230 Monomer determining unit

Claims (10)

a property prediction unit that obtains predicted values of a plurality of property types of a polymer by a model for each property type based on monomer information indicating a plurality of monomers and their content ratios;
a monomer determination unit that calculates a deviation value for the monomer information from the predicted values of the plurality of physical property types, and determines the monomers to be used in synthesizing the polymer and their content ratios based on the deviation value;
A physical property prediction device having the above structure. The property prediction device according to claim 1, wherein the model for each property type is derived by Gaussian process regression using training data consisting of pairs of the monomer information and the property values of the polymer property type corresponding to the monomer information. The property prediction device according to claim 1 or 2, wherein the monomer determination unit calculates output values of an acquisition function of Bayesian optimization from the predicted values of the plurality of property types, and calculates a deviation value of the output values. The physical property prediction device according to any one of claims 1 to 3, wherein the monomer determination unit determines the monomer information for which the deviation value is maximum or minimum as the monomers and their content ratios used in the synthesis of the polymer. Further comprising a monomer information generating unit that generates the monomer information,
5. The physical property prediction apparatus according to claim 1, wherein the monomer information is expressed by a product of an Extended Connectivity Fingerprint (ECFP) of each monomer and a molar ratio. A step in which one or more processors obtain predicted values of a plurality of physical property types of a polymer using a model for each physical property type based on monomer information indicating a plurality of monomers and their content ratios;
The one or more processors calculate deviation values for the monomer information from the predicted values of the plurality of physical property types, and determine the monomers and their content ratios used in the synthesis of the polymer based on the deviation values;
The physical property prediction method has the following features: The method for predicting physical properties according to claim 6, wherein the model for each physical property type is derived by Gaussian process regression using training data consisting of pairs of the monomer information and the physical property values of the physical property type of the polymer corresponding to the monomer information. The physical property prediction method according to claim 6 or 7, wherein the determining step calculates an output value of an acquisition function of Bayesian optimization from the predicted values of the plurality of physical property types, and calculates a deviation value of the output value. The method for predicting physical properties according to any one of claims 6 to 8, wherein the determining step determines the monomer information for which the deviation value is maximum or minimum as the monomers and their content ratios used in the synthesis of the polymer. generating said monomer information;
10. The method for predicting physical properties according to claim 6, wherein the monomer information is expressed by a product of an Extended Connectivity Fingerprint (ECFP) of each monomer and a molar ratio.

Images