Method for searching for sol-gel film including metal alkoxide, computer, and program

Abstract

To provide a method for searching for a sol-gel film which is a coating film including metal alkoxide, a computer, and a program.SOLUTION: A method for searching for a sol-gel film including metal alkoxide, executed by a computer, includes the steps of: generating a plurality of prediction models corresponding to a plurality of objective variables using result data relevant to a sol-gel film; and searching for a sol-gel film having a desired physical property balance by reverse analysis using a plurality of prediction models. The result data includes at least type and ratio of the metal alkoxide and synthesis information of the sol-gel film. The objective variables include at least one of the numerical values representing optical property, mechanical property, superficial, and chemical property.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to a method for searching for a sol-gel film containing a metal alkoxide, a computer, and a program.

Background Art

[0002] The sol-gel method is a technique for synthesizing glass, ceramics, etc. by a liquid-phase method, and can form various-shaped materials such as coating films, bulk bodies, fibers, and gels at relatively lower temperatures than the solid-phase method. Also, due to the characteristic of being a low-temperature synthesis method, it is possible to form a composite with an organic component, and a functional organic-inorganic hybrid material can be created.

[0003] By the way, in the case of glass, an information database (INTERGLAD) in which the composition and material characteristics are made into a database is known (see, for example, Non-Patent Document 1). INTERGLAD stores the characteristics and structure data of glass, and it is possible to search for the constituent components of glass having the characteristics specified by the user and the characteristics of glass having the composition specified by the user.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] The glass and ceramics obtained by the sol-gel method have a relationship between raw materials such as metal alkoxides, synthesis conditions such as treatment temperature, and material properties. On the other hand, the information database related to Non-Patent Document 1 does not record the synthesis conditions related to the sol-gel method. Also, although some components for forming sol-gel films such as tetraethoxysilane are recorded, the properties of sol-gel films depend not only on the macroscopic state such as the ratio of constituent components, but rather on the microscopic structure of the constituent molecules. For this reason, it was not possible to search for material properties based on the synthesis conditions related to the sol-gel method using this database. Similarly, it was not possible to search for the synthesis conditions of the sol-gel method based on material properties.

[0006] In addition, since the material properties of the glass and ceramics obtained by the sol-gel method vary greatly depending on the synthesis conditions and the synthesis requires a long time, a great deal of effort is required to optimize the synthesis conditions. Therefore, there were limitations in discovering new materials and developing new products through trial and error by utilizing past knowledge and experience.

[0007] Therefore, there is a need for a method for searching for sol-gel films, which are coating films containing metal alkoxides, a computer, and a program.

Means for Solving the Problems

[0008] The characteristic configuration of the method for searching for sol-gel films according to the present invention is a method for searching for sol-gel films containing metal alkoxides, which is executed by a computer, and includes a step of generating a plurality of prediction models each corresponding to a plurality of objective variables using performance data related to the sol-gel films, and a step of searching for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models. The performance data includes at least the type and ratio of the metal alkoxide and the synthesis information of the sol-gel film, and the objective variables include numerical values representing at least one of optical properties, mechanical properties, surface properties, and chemical properties.

[0009] The characteristic configuration of the computer according to the present invention is a computer that includes a control unit and searches for a sol-gel film containing a metal alkoxide. The control unit generates a plurality of prediction models each corresponding to a plurality of target variables using performance data related to the sol-gel film, searches for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models, the performance data includes at least the type and ratio of the metal alkoxide and the synthesis information of the sol-gel film, and the target variables include numerical values representing at least one of optical properties, mechanical properties, surface properties, and chemical properties.

[0010] The characteristic configuration of the program according to the present invention is to cause a computer to generate a plurality of prediction models each corresponding to a plurality of target variables using performance data related to a sol-gel film containing a metal alkoxide, and to search for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models. The performance data includes at least the type and ratio of the metal alkoxide and the synthesis information of the sol-gel film, and the target variables include numerical values representing at least one of optical properties, mechanical properties, surface properties, and chemical properties, and search for a sol-gel film.

[0011] The sol-gel film, which is a coating film synthesized by the sol-gel method, has various properties ranging from hydrophilicity, water repellency, anti-fouling, heat insulation, low reflectivity, antistatic, electromagnetic wave shielding, etc. These various properties vary greatly depending on the synthesis conditions of the sol-gel film. The synthesis conditions are synthesis information such as the type and ratio of metal alkoxides, which are precursor raw materials contained in the sol-gel film, and the film formation method of the sol-gel film. According to this configuration, among the various properties of the sol-gel film, a numerical value representing at least one of the optical properties, mechanical properties, surface properties, and chemical properties is used as the target variable, and by generating a plurality of prediction models corresponding to each of the target variables, it is possible to predict the various properties of the sol-gel film synthesized under predetermined conditions. Also, by performing inverse analysis using the prediction model, it is possible to narrow down the type of metal alkoxide, etc., that is likely to obtain the desired properties. In this way, by applying machine learning to material synthesis by the sol-gel method, it is possible to shorten the development speed of new materials and new products.

[0012] Another characteristic configuration of the method for searching for a sol-gel film according to the present invention is that the numerical value representing the optical property is the transmittance gain, the numerical value representing the mechanical property is a numerical value related to abrasion resistance, the numerical value representing the surface property is the contact angle, and the numerical value representing the chemical property is a numerical value related to chemical resistance or water resistance.

[0013] According to this configuration, by using a numerical value related to at least one of the transmittance gain, abrasion resistance, contact angle, chemical resistance, and water resistance as the target variable and generating a plurality of prediction models corresponding to each of the target variables, it is possible to predict the various properties of the sol-gel film synthesized under predetermined conditions.

[0014] Another characteristic configuration is that the metal alkoxide is a silane compound.

[0015] According to this configuration, it is possible to predict the film properties of the sol-gel film when using silicon alkoxide (silane compound), which is a typical sol-gel precursor, as the raw material.

[0016] Another characteristic configuration is that the silane compound is at least one of silicon alkoxide or silicon alkoxide modified with an organic group.

[0017] According to this configuration, it is possible to predict the film properties of sol-gel films made from different silane compounds or sol-gel films made from multiple types of silane compounds.

[0018] Another characteristic configuration is that in the step of generating the plurality of prediction models, feature quantities are calculated based on the performance data, and the feature quantities are used as explanatory variables of the plurality of prediction models.

[0019] According to this configuration, by calculating the feature quantities, factors contributing to the target characteristics can be obtained. Also, by using the feature quantities as explanatory variables, the accuracy of the prediction model can be improved.

[0020] Another characteristic configuration is that the sol-gel film further contains fine particles of an inorganic oxide, and the performance data further includes the diameter and ratio of the fine particles.

[0021] It is known that the various properties of the sol-gel film are greatly affected by the fine particles of the inorganic oxide. Therefore, according to this configuration, by calculating the fine particles contributing to the properties of the sol-gel film as feature quantities, the accuracy of the prediction model can be improved.

[0022] Another characteristic configuration is that the inorganic oxide is SiO2.

[0023] When SiO2 is contained in the sol-gel film, the low reflectivity of the sol-gel film can be exhibited. Therefore, according to this configuration, it becomes possible to search for a sol-gel film having low reflectivity.

[0024] Another characteristic configuration is that the feature quantity includes a molecular descriptor.

[0025] A molecular descriptor represents the structure and properties of a molecule numerically. According to this configuration, by calculating the molecular descriptor as a feature quantity and using it as an explanatory variable, it becomes possible to treat molecules not included in the performance data as input values and output values of the prediction model.

[0026] Another characteristic configuration is that the molecular descriptor is any one of the following (1), (2), (3), and (4). (1) The amount of a structure containing a specific organic group and a specific alkoxide group bonded to a Si atom (2) The ratio of molecules in which the number of specific elements other than Si in the silicon alkoxide or the silicon alkoxide modified with the organic group is a specific number or more (3) The ratio of trifunctional or higher-functional molecules (4) The number of organic rings in the molecule

[0027] According to this configuration, by calculating the molecular weight of the silane compound and the number of organic rings in the molecule of the silane compound as feature quantities and using them as explanatory variables, it becomes possible to treat silane compounds not included in the learning data as input values and output values of the prediction model.

[0028] Another characteristic configuration is that the molecular descriptor is the amount of a structure containing a methyl group and an ethoxy group bonded to a Si atom.

[0029] According to this configuration, it is possible to associate the number of bonds by which the silane compound undergoes polycondensation with the film properties of the sol-gel film.

[0030] Another characteristic configuration is that it is the ratio of molecules in which the number of oxygen atoms in the silicon alkoxide or the silicon alkoxide modified with the organic group is 4 or more.

[0031] According to this configuration, it is possible to associate the number of functional groups containing oxygen with high electronegativity with the film properties of the sol-gel film.

[0032] Another characteristic configuration is that the performance data includes performance data of a sol-gel film containing no predetermined silane compound and performance data of a sol-gel film containing the predetermined silane compound. In the step of generating the plurality of prediction models, after generating the plurality of prediction models using the performance data of the sol-gel film containing no predetermined silane compound, the film characteristics of the performance data of the sol-gel film containing the predetermined silane compound are predicted using the generated plurality of prediction models.

[0033] According to this configuration, after generating a prediction model using the performance data of a sol-gel film containing no predetermined silane compound, by predicting the film characteristics of the performance data containing the predetermined silane compound using the generated prediction model, it becomes possible to estimate the prediction accuracy of the prediction model for unknown data. Also, by adding data to the prediction model, the prediction model is refined and the optimization of the synthesis conditions of the sol-gel film becomes possible.

Brief Description of the Drawings

[0034]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Modes for Carrying Out the Invention

[0035] Hereinafter, embodiments of a method for searching for a sol-gel film, a computer, and a program according to the present invention will be described with reference to the drawings. However, the present invention is not limited to the following embodiments, and various modifications can be made without departing from the gist thereof.

[0036] The sol-gel film according to the present invention refers to a coating film synthesized by the sol-gel method. As shown in FIG. 1, the sol-gel method is a synthesis process in which a precursor solution in which a metal alkoxide, which is a sol-gel precursor raw material, is dissolved in a solvent is converted into a sol and a gel by hydrolysis and condensation reactions, and finally a metal oxide is obtained. By adding fine particles of inorganic oxides or organic compounds to the precursor solution, a material having a desired function can be synthesized. Further, by forming a gel under predetermined conditions, a thin film (coating film), an aerogel, or a fiber can be obtained. The sol-gel film may be crystalline or amorphous. That is, the sol-gel film may be glass or ceramics.

[0037] The system configuration used in the method for searching for a sol-gel film according to the present embodiment will be described with reference to FIGS. 2 and 3. A computer 1 that executes the search for the sol-gel film and an AI 5 (artificial intelligence) are connected to the Internet line 4. Here, the AI 5 may be provided on the Internet line 4 or may be provided in the computer 1. The computer 1 acquires, for example, performance data 6 related to sol-gel films stored in a server. The performance data 6 may be input to the computer 1 by a user. Further, at least one measuring device for measuring physical property information among the performance data 6 related to sol-gel films may be connected to the Internet line 4, and the computer 1 may acquire the performance data 6 related to physical property information via the Internet line 4.

[0038] The performance data 6 related to the sol-gel film includes the synthesis information and property information of the sol-gel film. The synthesis information includes, for example, the types and ratios of precursor raw materials used in the synthesis, solvent types, catalysts, additive and other raw material information, heating conditions for promoting hydrolysis and condensation reactions, film-forming conditions of the sol-gel film, etc. As the metal alkoxide which is the precursor raw material, for example, metal alkoxides of Si, Ti, Al, etc. can be mentioned. In this embodiment, however, a silicon alkoxide which is a silane compound is used as the metal alkoxide. Therefore, the synthesis information includes the types and ratios of the silane compounds contained in the sol-gel film. Silane compounds are roughly classified into silicon alkoxides in which all the functional groups bonded to the Si atom are alkoxide groups and silicon alkoxides modified with organic groups to which organic groups other than alkoxide groups are bonded to the Si atom. The former is, for example, tetraethoxysilane, and the latter is, for example, methyltriethoxysilane, phenyltriethoxysilane, etc. At least one type of silane compound is used as the precursor raw material in the sol-gel film. Various chemical properties can be imparted to the sol-gel film by the functional groups of the silane compounds. The ratio of the silane compound represents the molar ratio of the silane compounds contained in the sol-gel film.

[0039] In addition, inorganic oxide fine particles are used as additives added to the precursor solution of the sol-gel film. Therefore, the synthesis information may further include the types, diameters, or ratios of the inorganic oxide fine particles. Inorganic oxides are, for example, SiO2, TiO2, ZnO, CeO2, ZrO2, ITO, SnO2, etc. The ratio of the fine particles represents the weight ratio of the fine particles contained in the sol-gel film. It is known that various properties such as hydrophilicity, water repellency, antifouling, heat insulation, low reflection, antistatic, and electromagnetic wave shielding can be imparted to the sol-gel film according to the types and particle shapes of the fine particles. In addition to the inorganic oxide fine particles, metal fine particles may be further added. Metals are Au, Ag, Cu, etc.

[0040] Among the synthesis information, the heating conditions may include the heating temperature, heating time, etc. of the precursor solution. The film formation conditions are film formation methods such as the roll coating method or the spin coating method. In the case of a sol-gel film obtained by the spin coating method, the film formation conditions such as the solid content concentration in the coating solution and the spin coating rotation speed may be included. Depending on the film formation conditions, the film thickness of the sol-gel film and the bonding property with the substrate on which the sol-gel film is formed change. The characteristics of the sol-gel film change depending on the film thickness.

[0041] The physical property information of the sol-gel film is, for example, a measurement value representing optical characteristics. The physical property information related to the performance data 6 includes numerical values representing at least one of the optical properties, mechanical properties, surface properties, and chemical properties among the film properties of the sol-gel film. More specifically, the numerical value representing the optical property is the transmittance gain, the numerical value representing the mechanical property is a numerical value related to abrasion resistance, the numerical value representing the surface property is the contact angle, the numerical value representing the chemical property is a numerical value related to chemical resistance and water resistance, and the physical property information includes at least one of these. The transmittance gain is the increment of the average transmittance of the coated body with the sol-gel film formed in the wavelength range of 380 to 850 nm with respect to the average transmittance of the coated body (for example, a transparent substrate) in the state where the sol-gel film is not formed in this wavelength range.

[0042] Next, the computer 1 that executes the search for the sol-gel film will be described. The computer 1 includes a prediction model generation device 2 and a film property prediction device 3. The prediction model generation device 2 inputs the feature amount extracted from the performance data 6 to the AI 5. The AI 5 performs machine learning and generates a prediction model 10 with the film properties of the sol-gel film as the target variable. The target variable related to the prediction model 10 includes numerical values representing at least one of the optical properties, mechanical properties, surface properties, and chemical properties among the respective film properties of the sol-gel film. Specifically, it includes at least one of the numerical values related to the transmittance gain, abrasion resistance, contact angle, chemical resistance, and water resistance.

[0043] The prediction model generation device 2 includes a first communication unit 21, a feature amount calculation unit 22, a model generation unit 23, a first storage unit 24, and a first control unit 25 (an example of a control unit).

[0044] The first communication unit 21 is an interface for transmitting and receiving data to and from a server or AI 5 via the Internet line 4. The first communication unit 21 acquires, as learning performance data 24a, performance data 61 in which the precursor raw material does not contain a predetermined silane compound among the performance data 6 stored in the server (step #1 in FIG. 4). Further, the first communication unit 21 acquires, as re-learning performance data 24c, performance data 62 in which the precursor raw material contains a predetermined silane compound (step #4 in FIG. 4). As will be described later, by generating the prediction model 10 using the performance data 61 that does not contain a predetermined silane compound and predicting the film properties using the prediction model 10 with the performance data 62 that contains a predetermined silane compound, it is possible to evaluate the extrapolation prediction accuracy of the prediction model 10.

[0045] The first storage unit 24 is composed of a non-temporary storage medium such as an HDD or SSD or a temporary storage medium such as a RAM, and stores programs and applications executed by the processor. This first storage unit 24 stores the learning performance data 24a and the re-learning performance data 24c acquired via the first communication unit 21. The learning performance data 24a and the re-learning performance data 24c have the synthesis information and physical property information of the sol-gel film described above.

[0046] The first storage unit 24 stores the prediction model 10 generated by the model generation unit 23. The prediction model 10 is a model that functions by a computer and is obtained by supervised learning of machine learning.

[0047] The first control unit 25 includes a processor and controls the operations of the prediction model generation device 2 such as the first communication unit 21. The feature quantity calculation unit 22 and the model generation unit 23 function by the first control unit 25. The processor includes other hardware for executing applications stored in an ASIC, FGPA, CPU, or the first storage unit 24 (the same applies hereinafter).

[0048] Furthermore, the first memory unit 24 stores learning feature amounts 24b and relearning feature amounts 24d calculated by the feature amount calculation unit 22 for machine learning. The feature amount calculation unit 22 calculates a plurality of learning feature amounts 24b and relearning feature amounts 24d from the composite information included in the learning performance data 24a and the relearning performance data 24c. The learning feature amounts 24b and the relearning feature amounts 24d may include, for example, the diameter and ratio of the fine particles of the inorganic oxide contained in the sol-gel film. Also, the learning feature amounts 24b and the relearning feature amounts 24d may include molecular descriptors.

[0049] A molecular descriptor is a numerical representation of the structure and properties of a molecule. Molecular descriptors can be obtained by using a known molecular descriptor generation library such as RDKit or commercially available software such as alvaDesc from the information of the sol-gel film. Examples of such molecular descriptors include, as 0-dimensional descriptors, molecular weight, the number of atoms in the molecule, the number of bonds, etc. Examples of 1-dimensional descriptors include the number of functional groups in the molecule, the number of alicyclic structures, the number of aromatic rings, etc. Examples of 2-dimensional descriptors include topological descriptors that represent a molecule based on graph theory such as Wiener index, Chiconnectivity index, Balaban index, BCUT descriptors, etc. Examples of 3-dimensional descriptors include those calculated based on the 3-dimensional structure of the molecule such as 3D-MoRSE, WHIM, CoMFA, CoMSIA, etc. Examples of 4-dimensional descriptors include those calculated taking interactions into account such as Hopfinger, etc., but they can be appropriately selected and used. Thus, by calculating molecular descriptors as the learning feature amounts 24b or the relearning feature amounts 24d, molecules not included in the learning performance data 24a and the relearning performance data 24c can also be treated as input items of the machine learning model.

[0050] The molecular descriptors considered to be particularly useful in the present invention are roughly classified as follows. (1) The amount of a structure containing a specific organic group and a specific alkoxide group bonded to an Si atom (2) The proportion of molecules in silicon alkoxide or silicon alkoxide modified with an organic group in which the number of specific elements other than Si is equal to or greater than a specific number (3) The proportion of trifunctional or higher molecules, that is, molecules in silicon alkoxide or silicon alkoxide modified with an organic group in which the number of alkoxide groups bonded to the Si atom is three or more (4) The number of organic rings in the molecule

[0051] Here, examples of the organic ring include, but are not limited to, saturated alkyl rings such as cycloalkane, unsaturated alkyl rings such as benzene, and heterocyclic rings such as pyridine. More specifically, for (1) among the molecular descriptors, examples include the amount of the structure containing a methyl group and an ethoxy group bonded to the Si atom. More specifically, for (2) among the molecular descriptors, examples include the proportion of molecules in silicon alkoxide or silicon alkoxide modified with an organic group in which the number of oxygen atoms is 4 or more. Different from the fused glass in which the physical properties depend only on the proportion of the components of the composition, the physical properties of the sol-gel film are greatly affected by the local structure of the constituent molecules. Therefore, it can be seen that these molecular descriptors are very important factors for predicting the physical properties from the synthesis information.

[0052] In this embodiment, only one type of silane compound may be used as the precursor raw material, or two or more types of silane compounds may be used. Since one vector can be obtained for one molecule with respect to the molecular descriptor, it is necessary to convert the molecular descriptor representing one molecule into the molecular descriptor representing the mixture. Therefore, the molecular descriptor of the mixture may be used as the weighted average, weighted dispersion, maximum value, and minimum value of the molecular descriptors of the constituent molecules. For example, if one of the items of the molecular descriptor is the molecular weight, four values of the weighted average, load dispersion, maximum value, and minimum value of the two types of molecular weights can be treated as the descriptors of the molecular weight of the mixture according to the mixing ratio of the two types of silane compounds.

[0053] The model generation unit 23 uses the learning feature quantity 24b calculated by the feature quantity calculation unit 22 as input data, and uses the learning result data 24a as teacher data to input to the AI 5 for machine learning to generate the prediction model 10 (step #3 in FIG. 4). At this time, the model generation unit 23 generates a plurality of prediction models 10 corresponding to various target variables. For example, when there are five target variables, the model generation unit 23 generates five prediction models 10.

[0054] Further, the model generation unit 23 uses the re-learning feature quantity 24d calculated by the feature quantity calculation unit 22 as input data, and uses the re-learning result data 24c as teacher data to input to the AI 5 for machine learning to re-learn the prediction model 10 (step #6 in FIG. 4).

[0055] The prediction model 10 is, for example, linear regression, deep learning, or a decision tree. Note that the prediction model 10 may be an ensemble model such as XGBoost, Random Forest, LightGBM, CatBoost, or AdaBoost in which a plurality of decision trees are provided in association with each other.

[0056] Next, the film property prediction device 3 will be described. As shown in FIG. 3, the film property prediction device 3 includes a second communication unit 31, a film property prediction unit 32, a synthesis condition prediction unit 33, a second storage unit 34, and a second control unit 35 (an example of a control unit).

[0057] The second control unit 35 includes a processor and controls the operations of the film property prediction device 3 such as the second communication unit 31. The film property prediction unit 32 and the synthesis condition prediction unit 33 function by the second control unit 35.

[0058] The second communication unit 31 is an interface for transmitting and receiving data to and from a server via the Internet line 4 or to and from the first communication unit 21. The second communication unit 31 acquires the re-learning result data 24c and the re-learning feature quantity 24d stored in the first storage unit 24. Further, the second communication unit 31 acquires the property information 34a having a desired physical property balance input by the user to an input unit (not shown) (step #11 in FIG. 5).

[0059] The second storage unit 34 is composed of a non-temporary storage medium such as an HDD or an SSD, or a temporary storage medium such as a RAM, and stores programs and applications executed by the processor. This second storage unit 34 stores the re-learning performance data 24c acquired via the first communication unit 21, the re-learning feature amount 24d, at least one prediction model 10 generated by the prediction model generation device 2, and the object property information 34a input by the user.

[0060] The film property prediction unit 32 inputs the re-learning feature amount 24d into the prediction model 10 corresponding to the target variable to be predicted, and outputs the predicted values of various physical properties related to the film properties of the sol-gel film predicted by the prediction model 10. The predicted value is a numerical value representing at least one of optical properties, mechanical properties, surface properties, and chemical properties. Specifically, it is any value of transmittance gain, wear resistance, contact angle, chemical resistance, and water resistance. These estimated values are stored in the second storage unit 34. The user may perform accuracy verification based on this predicted value and the known object property information included in the performance data 62. As a result of the verification, if the accuracy is within the practical range, the search process for the sol-gel film using the prediction model 10 may be performed. Note that the film property prediction unit 32 may perform accuracy verification and may output a prediction error. As a result of the verification, if the accuracy is within the practical range, the search process for the sol-gel film using the prediction model 10 may be performed.

[0061] The synthesis condition prediction unit 33 performs inverse analysis using the prediction model 10 based on the property information 34a input by the user, and searches for the synthesis conditions of the sol-gel film having the desired property balance. The property information 34a may be the target value of each film property indicating the desired property balance of the sol-gel film. The desired property balance may be, for example, the target value related to at least one film property. The synthesis condition prediction unit 33 predicts the prediction feature quantity 34b capable of expressing the desired property balance, executes the optimization process, and outputs the search result (steps #12, #13, #14 in FIG. 5). The synthesis condition prediction unit 33 predicts, as the synthesis conditions related to the sol-gel film having the desired property balance, for example, the type and ratio of the precursor raw material, the type and diameter of the fine particles, the heating conditions, or the film formation conditions. Alternatively, the synthesis condition prediction unit 33 may predict the feature quantity related to at least one sol-gel film having the desired property balance. The optimization process may be performed using a method such as the steepest descent method or Bayesian optimization.

[0062] Subsequently, the method for generating the prediction model 10 according to the present embodiment and the search method for searching for the sol-gel film having the desired property balance using the prediction model 10 will be described with reference to FIGS. 4 and 5. Note that the generation of the prediction model 10 and the search for the sol-gel film using the prediction model 10 may be executed by a program. The program may be stored in an arbitrary recording device of the first storage unit 24 or the second storage unit 34, and the processor included in the first control unit 25 or the second control unit 35 may execute this program.

[0063] First, the first communication unit 21 of the prediction model generation device 2 acquires the learning performance data 24a from the performance data 6 that does not contain a predetermined silane compound in the precursor raw material, and stores it in the first storage unit 24 (step #1).

[0064] Subsequently, the feature quantity calculation unit 22 calculates the learning feature quantity 24b based on the learning performance data 24a and stores it in the first storage unit 24 (step #2). Specifically, the feature quantity calculation unit 22 calculates the learning feature quantity 24b based on the type and ratio of the silane compound, which is the composite information included in the learning performance data 24a, the particle diameter and ratio of the inorganic oxide fine particles, the film formation conditions, and the molecular descriptor. For the calculation of the learning feature quantity 24b, the feature quantity calculation unit 22 may refer to a database, a library, etc. as appropriate, and these databases, etc. may be stored in the first storage unit 24.

[0065] The model generation unit 23 inputs the learning feature quantity 24b as input data and the learning performance data 24a as teacher data into the AI 5. The AI 5 generates at least one prediction model 10 by machine learning, using the learning feature quantity 24b as an explanatory variable and a numerical value representing at least one of the optical properties, mechanical properties, surface properties, and chemical properties, more specifically, a numerical value related to at least one of the film characteristics of the transmittance gain, wear resistance, contact angle, chemical resistance, and water resistance as the objective variable (step #3).

[0066] Next, the first communication unit 21 acquires the re-learning performance data 24c from the performance data 6, specifically, the performance data 62 that contains a predetermined silane compound in the precursor raw material, and stores it in the first storage unit 24 (step #4).

[0067] The feature quantity calculation unit 22 calculates the re-learning feature quantity 24d based on the re-learning performance data 24c and stores it in the first storage unit 24 (step #5). Similar to the learning feature quantity 24b, the re-learning feature quantity 24d includes, for example, the type and ratio of the silane compound, the particle diameter and ratio of the inorganic oxide fine particles, the film formation conditions, and the molecular descriptor.

[0068] The model generation unit 23 inputs the re-learning feature quantity 24d as input data and the re-learning performance data 24c as teacher data into the AI 5. The AI 5 re-learns at least one prediction model 10 by machine learning (step #6).

[0069] Subsequently, the film property prediction device 3 searches for a sol-gel film having a desired physical property balance. The second communication unit 31 acquires physical property information 34a representing the desired physical property balance (step #11). The physical property information 34a is stored in the second storage unit 34.

[0070] The synthesis condition prediction unit 33 searches for a prediction feature amount 34b related to a sol-gel film having a desired physical property balance based on the physical property information 34a by means of the prediction model 10 (step #12). Further, the synthesis condition prediction unit 33 executes an optimization process on the search result obtained in step #12 (step #13). The optimized search result is output, for example, as the type and ratio of the silane compound, the type of fine particles, or the film formation conditions (step #14). Note that the synthesis condition prediction unit 33 may output the prediction feature amount 34b of the sol-gel film having a desired physical property balance.

[0071] 〔Example〕 Hereinafter, examples of the present invention will be described, but the present invention is not limited to the descriptions of these examples.

[0072] 〔Example 1〕 As the performance data 6, synthesis information and physical property information related to known sol-gel films, and synthesis information and physical property information of sol-gel films newly created by experiments were collected. From the known data, physical property information indicating each film property regarding transmittance gain, abrasion resistance, contact angle, and chemical resistance was collected. The composition and film formation conditions of the precursor solution of the sol-gel film newly created by experiments were determined so that the D-optimality criterion would be maximized. The D-optimality criterion means the determinant det(X when the matrix with each feature amount in the column direction and each data in the row direction is X. T·X). A set of experimental parameters with a larger D-optimality criterion results in fewer candidates for similar experimental parameters and makes it easier to construct an accurate machine learning model. Physical property information indicating each film property such as transmittance gain, wear resistance, contact angle, chemical resistance, and water resistance was collected from the newly created sol-gel film by experiments. As the performance data 6, among the sol-gel films prepared by the roll coating method and the sol-gel films prepared by the spin coating method, a sol-gel film prepared by changing the spin coating rotation speed in increments of 500 rpm between 1000 and 3500 rpm and having the largest transmittance gain value was used. That is, by using the film formation conditions of the sol-gel film from which the data with the largest transmittance gain value was obtained as the optimal values, a prediction model 10 that does not use the film thickness (film formation conditions) as an explanatory variable was generated.

[0073] The wear resistance was evaluated by the change in the average transmittance of light with wavelengths in the range of 380 to 850 nm on the substrate coated with the sol-gel film before and after performing a reciprocating wear test (load: 4 N, number of reciprocations: 150 times) by bringing the abrasive CS-10F into contact with the sol-gel film. The water resistance was evaluated by the change in reflectance after immersing in water at 80 °C for 10 hours and then rubbing with a cloth. The chemical resistance was evaluated by the change in the average transmittance of light with wavelengths in the range of 380 to 850 nm on the substrate coated with the sol-gel film before and after immersing the sol-gel film in an alkaline aqueous solution (pH of about 8) maintained at 65 °C for 24 hours.

[0074] Among the performance data 6, using the performance data 61 that does not contain a predetermined silane compound, a plurality of prediction models 10 with the transmittance gain, abrasion resistance, contact angle, chemical resistance, or water resistance as the target variables were generated by machine learning, and the relearning of the prediction model 10 and the evaluation of the extrapolation prediction accuracy were performed using the performance data 62 that contains the predetermined silane compound. The predetermined silane compound was tetraethoxysilane. The evaluation of the extrapolation prediction accuracy was performed based on the prediction error between the measured values of the physical property information included in the performance data 62 and the predicted values of each film property predicted by the prediction model 10, and the distribution of the data in the plot of the measured values against the predicted values. Note that since a plurality of prediction errors are obtained when evaluating the prediction accuracy for each of the plurality of silane compounds used as the precursor raw materials, the average value of the plurality of prediction errors was used as an index of the extrapolation prediction accuracy of the prediction model 10.

[0075] First, for each film property, the extrapolation prediction error was evaluated for each prediction model 10 created using different molecular descriptors as explanatory variables. The prediction error of the prediction model 10 with the best prediction accuracy among the prediction models 10 related to the transmittance gain was 0.415, the prediction error of the prediction model 10 with the best prediction accuracy among the prediction models 10 related to the abrasion resistance was 0.501, the prediction error of the prediction model 10 with the best prediction accuracy among the prediction models 10 related to the contact angle was 20.085, the prediction error of the prediction model 10 with the best prediction accuracy among the prediction models 10 related to the chemical resistance was 1.091, and the prediction error of the prediction model 10 with the best prediction accuracy among the prediction models 10 related to the water resistance was 1.069. Except for the chemical resistance and water resistance, the obtained molecular descriptors of the prediction models 10 with good prediction accuracy were different, and it was speculated that the types of optimal molecular descriptors differed for each film property. Therefore, a plot of the measured values against the predicted values was made using the prediction model 10 in which the optimal molecular descriptor was used for each film property.

[0076] The plot of measured values against predicted values for each film property in FIG. 6 is a plot when phenyltriethoxysilane is used as the silane compound contained in the precursor raw material. As shown in FIG. 6, it was found that for the transmittance gain, abrasion resistance, contact angle, and chemical resistance, the measured values and the predicted values match and the prediction accuracy is good, while for the water resistance, the measured values and the predicted values deviate and the prediction accuracy is poor. However, when the measured values against predicted values were plotted for other silane compounds as the precursor raw material and a plurality of measured values against predicted values plots were evaluated as a whole, results were obtained such that the prediction accuracy for the transmittance gain, abrasion resistance, contact angle, and water resistance is good, and the prediction accuracy for the chemical resistance is poor. Thereby, it was found that a prediction model 10 having practical prediction accuracy can be generated for the transmittance gain, abrasion resistance, contact angle, and water resistance.

[0077] Example 2 As performance data 6, data of all the sol-gel films prepared by the spin coating method were used. Thereby, a prediction model 10 having the film thickness (film formation conditions) as an explanatory variable was generated. Note that since there are no parameters such as the spin coating rotation speed for the sol-gel films prepared by the roll coating method, it is difficult to handle them with the same prediction model 10. Therefore, the performance data 6 in Example 2 does not include data of the sol-gel films prepared by the roll coating method. The generation of the prediction model 10, the evaluation of the extrapolation prediction error, the plot of the actual values against the predicted values, etc. were performed in the same manner as in Example 1.

[0078] For each film property, the extrapolation prediction error was evaluated for each of the 10 prediction models created using different molecular descriptors as explanatory variables. Among the 10 prediction models related to the transmittance gain, the prediction error of the prediction model 10 with the best prediction accuracy was 0.349. Among the 10 prediction models related to the abrasion resistance, the prediction error of the prediction model 10 with the best prediction accuracy was 0.603. Among the 10 prediction models related to the contact angle, the prediction error of the prediction model 10 with the best prediction accuracy was 21.738. Among the 10 prediction models related to the chemical resistance, the prediction error of the prediction model 10 with the best prediction accuracy was 1.188. Among the 10 prediction models related to the water resistance, the prediction error of the prediction model 10 with the best prediction accuracy was 0.927. For any film property, the obtained molecular descriptors of the prediction model 10 with good prediction accuracy were different, and it was speculated that the optimal types of molecular descriptors were different for each film property. Also, compared with Example 1, it was found that the prediction error was large and the prediction accuracy of the prediction model 10 was slightly inferior. This is considered to be because the number of explanatory variables such as the spin coating rotation speed increased.

[0079] The plot of the measured values against the predicted values for each film property in Fig. 7 is the plot when phenyltriethoxysilane is used as the silane compound contained in the precursor raw material. It was found that for the transmittance gain, abrasion resistance, contact angle and water resistance, the measured values and the predicted values were in agreement and the prediction accuracy was good, while for the chemical resistance, the measured values and the predicted values deviated and the prediction accuracy was poor. Similar to Example 1, when the plots of the measured values against the predicted values were made for other silane compounds as the precursor raw materials and the plurality of plots of the measured values against the predicted values were evaluated as a whole, the same results were obtained. From this, it was found that even when the film formation conditions were added as explanatory variables, it was possible to generate a prediction model 10 having practical prediction accuracy for the transmittance gain, abrasion resistance, contact angle and water resistance. Also, since sufficient extrapolation prediction accuracy was obtained without using the data in which the film formation conditions of the sol-gel film were optimized, it was confirmed that there was no problem in constructing the prediction model 10 even if the film formation conditions were not optimized when collecting the performance data 6 and the data was collected by experimenting under arbitrary film formation conditions.

[0080] [Example 3] In order to analyze the parameters contributing to each film property of the sol-gel film, without dividing the performance data 6 into performance data 61 and performance data 62, the prediction model 10 was generated using all of the performance data 6. The generation of the prediction model 10 was performed in the same manner as in Example 1.

[0081] FIG. 8 is an example showing the contribution degrees of the respective explanatory variables to each film property in this example. The contribution of each explanatory variable to each film property was evaluated by determining the variable importance using a method called CVPFI. FIG. 8 shows the contribution of each explanatory variable to each film property. As the film properties, a transmittance gain, wear resistance, contact angle, chemical resistance, and water resistance are shown. Also, as the explanatory variables, the ratio of fine particles which are non-molecular descriptors, molecular descriptor A, molecular descriptor B, and molecular descriptor C are shown. Regarding the contribution of each explanatory variable to each film property, an explanatory variable with a variable importance of 1.0 or more is indicated as ◎, an explanatory variable with a variable importance of 0.2 or more is indicated as 〇, and an explanatory variable with a variable importance of 0.1 or less is indicated as ×.

[0082] As shown in FIG. 8, the variable importance of the ratio of fine particles which are non-molecular descriptors with respect to the transmittance gain is 1.0 or more, and it can be seen that the ratio of fine particles dominantly contributes to the transmittance gain. Also, the variable importance of the ratio of fine particles and molecular descriptor A with respect to the wear resistance is 0.2 or more, and it can be seen that the ratio of fine particles and molecular descriptor A contribute to the wear resistance. Molecular descriptor A is, for example, a partial structure in which an ethoxy group and a methyl group are bonded to Si. It is presumed that when a large amount of methyl groups are included, the number of bonds by which the precursor raw material undergoes polycondensation decreases, and the strength of the film decreases. Therefore, it is understandable that the more the partial structure represented by molecular descriptor A is included, the worse the wear resistance becomes.

[0083] The variable importance of the ratio of fine particles with respect to the contact angle was 0.1 or less, and the variable importance of molecular descriptor B was 0.2 or more. Therefore, it can be seen that for the contact angle, the contribution of the ratio of fine particles, which is a non-molecular descriptor, is small, and the contribution of molecular descriptor B is large. From this, it can be understood that the contact angle is greatly affected by the properties of the precursor raw materials. Molecular descriptor B is, for example, a partial structure containing four or more oxygen atoms. It is considered that the number of oxygen atoms in the molecule affects the contact angle because when highly electronegative oxygen is included, strongly polarized functional groups such as Si-OH are easily formed on the film surface, resulting in the expression of hydrophilicity.

[0084] The variable importance of the ratio of fine particles with respect to chemical resistance and the variable importance of molecular descriptor C were 0.2 or more. Therefore, it can be seen that for chemical resistance, the contribution of the ratio of fine particles, which is a non-molecular descriptor, and the contribution of molecular descriptor C are large. Molecular descriptor C is, for example, a partial structure in which three or more oxygen atoms are bonded to a silicon atom, that is, trifunctional Si or tetrafunctional Si in the precursor raw material. The reason why the more this partial structure is contained, the higher the chemical resistance is that trifunctional Si or tetrafunctional Si has more bonding numbers for polycondensation than bifunctional Si, so it can be understood that the strength of the film increases.

[0085] The variable importance of the ratio of fine particles with respect to water resistance was 1.0 or more, and it was found that for water resistance, the ratio of fine particles, which is a non-molecular descriptor, dominantly contributes. It was also found that the contribution of the molecular descriptor is small.

[0086] From the above, it was found that the prediction accuracy of prediction model 10 can be improved by using the ratio of fine particles as an explanatory variable. It was also found that the prediction accuracy of prediction model 10 can be improved by using a molecular descriptor with high variable importance as an explanatory variable for each film property.

[0087] 〔Other Embodiments〕 (a) In the above-described embodiment, the second communication unit 31 acquired the performance data 24c for relearning and the feature quantity 24d for relearning via the first communication unit 21. However, the second communication unit 31 may directly acquire the performance data 62 from the server as the performance data 24c for relearning. Further, the film property prediction unit 32 may calculate the feature quantity 24d for relearning. Furthermore, the film property prediction unit 32 may be provided in the prediction model generation device 2.

[0088] (b) In the above-described embodiment, the AI 5 may be provided in the computer 1. Further, the computer 1 may not be connected to the Internet line 4.

[0089] (c) In the above-described embodiment, part or all of the method for generating the prediction model 10 by the prediction model generation device 2, the method for relearning the prediction model 10, and the method for searching for the sol-gel film by the film property prediction device 3 are not limited to the above-described device configuration and can be realized with any configuration.

[0090] (d) In the above-described example, the prediction model 10 was generated with the transmittance gain, wear resistance, contact angle, chemical resistance, and water resistance as the target variables. However, the target variables are not limited to these, and for example, heat insulation properties or the like may be used as the target variables.

[0091] (e) In the above-described embodiment, the prediction model 10 was generated using the precursor raw material as a silane compound. However, the prediction model 10 may be generated using a metal alkoxide other than the silane compound.

[0092] (f) In the above-described embodiment, after calculating the learning feature quantity 24b, the explanatory variables may be selected using an algorithm (Boruta) that can extract variables with significantly high importance from noise, for example. When using molecular descriptors, the number of explanatory variables increases, and the model construction and interpretability of the model deteriorate. Therefore, a useful prediction model 10 can be generated by selecting the explanatory variables.

[0093] (g) In the above-described embodiment, after generating a plurality of prediction models 10 using performance data 61 that does not contain a predetermined silane compound, the prediction models 10 are re-learned using performance data 62 that contains a predetermined silane compound. However, instead of performing re-learning, in the step of generating the prediction models 10 (step #3 in FIG. 4), after generating a plurality of prediction models 10 using performance data 61 that does not contain a predetermined silane compound, the film properties may be predicted using performance data 62 that contains a predetermined silane compound.

Industrial Applicability

[0094] The present invention can be used in a method for searching for a sol-gel film which is a coating film containing a metal alkoxide, a computer, and a program.

Explanation of Symbols

[0095] 1: Computer 6: Performance data 10: Prediction model 24b: Feature quantity for learning (feature quantity) 24d: Feature quantity for re-learning (feature quantity) 25: First control unit (control unit) 35: Second control unit (control unit) 61: Performance data 62: Performance data

Claims

1. A method for searching for a sol-gel film containing a metal alkoxide, which is executed by a computer, comprising: generating a plurality of prediction models each corresponding to a plurality of target variables using performance data related to the sol-gel film; searching for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models; and the performance data includes at least the type and ratio of the metal alkoxide and the synthesis information of the sol-gel film; the target variable includes a numerical value representing at least one of optical properties, mechanical properties, surface properties, and chemical properties, a method for searching for a sol-gel film.

2. The numerical value representing the optical property is transmittance gain, the numerical value representing the mechanical property is a numerical value related to wear resistance, the numerical value representing the surface property is contact angle, and the numerical value representing the chemical property is a numerical value related to chemical resistance or water resistance. The method for searching for a sol-gel film according to Claim 1.

3. The method for searching for a sol-gel film according to Claim 1, wherein the metal alkoxide is a silane compound.

4. The method for searching for a sol-gel film according to Claim 3, wherein the silane compound is at least one of silicon alkoxide or silicon alkoxide modified with an organic group.

5. In the step of generating the plurality of prediction models, a feature amount is calculated based on the performance data, and the feature amount is used as an explanatory variable of the plurality of prediction models. The method for searching for a sol-gel film according to Claim 1.

6. The sol-gel film further contains inorganic oxide fine particles, and the performance data further includes the diameter and ratio of the fine particles. The method for searching for a sol-gel film according to Claim 1.

7. The inorganic oxide is SiO ２ The method for searching a sol-gel film according to claim 6, wherein the inorganic oxide is SiO

8. The method for searching for a sol-gel film according to Claim 5, wherein the feature amount includes a molecular descriptor.

9. The method for searching for a sol-gel film according to Claim 8, wherein the molecular descriptor is any one of the following (1), (2), (3), and (4). (1) The amount of a structure containing a specific organic group and a specific alkoxide group bonded to an Si atom (2) The ratio of molecules in which the number of specific elements other than Si in the silicon alkoxide or the silicon alkoxide modified with the organic group is a specific number or more (3) The ratio of trifunctional or higher-functional molecules (4) The number of organic rings in the molecule

10. The method for searching for a sol-gel film according to Claim 9, wherein the molecular descriptor is the amount of a structure containing a methyl group and an ethoxy group bonded to an Si atom.

11. The method for searching a sol-gel film according to claim 9, wherein the molecular descriptor is the ratio of molecules having 4 or more oxygen atoms in the silicon alkoxide or the silicon alkoxide modified with the organic group.

12. The performance data includes performance data of a sol-gel film not containing a predetermined silane compound and performance data of a sol-gel film containing the predetermined silane compound. In the step of generating the plurality of prediction models, After generating the plurality of prediction models using the performance data not containing the predetermined silane compound, the film properties of the performance data containing the predetermined silane compound are predicted using the generated plurality of prediction models. The method for searching a sol-gel film according to any one of claims 1 to 11.

13. A computer including a control unit for searching a sol-gel film containing a metal alkoxide, The control unit generates a plurality of prediction models respectively corresponding to a plurality of objective variables using the performance data related to the sol-gel film, and searches for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models. The performance data includes at least the type and ratio of the metal alkoxide and the film formation conditions of the sol-gel film. The objective variable includes a numerical value representing at least one of optical properties, mechanical properties, surface properties, and chemical properties.

14. Causing a computer to generate a plurality of prediction models respectively corresponding to a plurality of objective variables using performance data related to a sol-gel film containing a metal alkoxide, and searching for a sol-gel film having a desired physical property balance by inverse analysis using the plurality of prediction models. The performance data includes at least the type and ratio of the metal alkoxide and the synthesis information of the sol-gel film. A program for searching a sol-gel film, wherein the objective variable includes a numerical value representing at least one of optical properties, mechanical properties, surface properties, and chemical properties.

Images