Method for searching for raw material compounds for paints

Abstract

To provide a method for searching for raw material compounds for paints that can be searched quickly, accurately, and easily. [Solution] A method for searching for a raw material compound for paint, comprising: a selection step of selecting candidate compounds for paint raw material compounds using a judgment model obtained by machine learning; and a confirmation step of conducting experiments to confirm the properties of the candidate compounds for paint raw material compounds selected in the selection step, wherein the raw material compound for paint is one or more selected from the group consisting of a curing agent, a curing agent component, a resin for forming a coating film, an additive, a pigment, and a solvent.

Description

[Technical Field]

[0001] The present invention relates to a method for searching for raw material compounds for paints. [Background Art]

[0002] In the development of paints, selection of raw material compounds for paints is important to obtain a paint having desired properties. Heretofore, selection of raw material compounds for paints has depended on the knowledge of developers based on composition and property data of paints produced so far. However, since demands for paint properties have become more advanced and complicated, and the types of raw material compounds for paints have increased, many experiments are required to obtain a paint having desired properties. For this reason, time, labor, cost, waste generation amount and the like are increasing in the development of paints.

[0003] In view of such circumstances, the use of Materials Informatics (MI) in the design and development of new materials has been studied. Materials Informatics (MI) is said to enable high-quality and efficient design and development of new materials by integrating experimental science, theoretical science, computational science, and information (data-driven) science in the design and development of new materials, and has been used in various research and development activities.

[0004] The following techniques are conventionally known for the development of new materials. Patent Document 1 discloses a material design apparatus for designing a design target material including a material composed of a plurality of compositions or a material produced by a combination of a plurality of production conditions. This apparatus uses a trained model obtained by acquiring, via machine learning, the correspondence between input information including design conditions of a design target material and output information including material property values. Patent Document 2 discloses a materials informatics-based materials development method for searching for a target material, in which the properties of the material, the composition of the material, and / or the process of the material are predicted based on data including first data showing at least one of composition data, process data, and characteristic data for an individual sample, and second data showing microstructure data including specific features of the sample. Patent Document 3 discloses a method for searching for novolac-type phenolic resins, which is performed by an information processing device. The method includes the steps of: generating a plurality of prediction models corresponding to a plurality of target variables using actual data relating to novolac-type phenolic resins; and searching for a novolac-type phenolic resin having a desired balance of physical properties by performing an inverse analysis using the plurality of prediction models. The actual data includes the polymer composition, structural formula, reaction solvent, and reaction parameters relating to the novolac-type phenolic resin, and the target variables include developability, heat resistance, and molecular weight. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] International Publication No. 2020 / 090848 [Patent Document 2] Japanese Patent Publication No. 2023-24492 [Patent Document 3] Japanese Patent Publication No. 2023-90563 [Overview of the project] [Problems that the invention aims to solve]

[0006] Previously, the search for raw material compounds for paints involved selecting candidate compounds based on past experimental data and researchers' knowledge, then repeatedly preparing paints using these candidate compounds, evaluating them, and deciding whether to accept or reject them. However, since paints contain a wide variety of components, and their compositional ratios also vary widely, finding the optimal raw material compound from among the many available requires a tremendous amount of time, expense, and experimentation, highlighting the need for improvement. Patent documents 1 to 3 do not relate to methods for searching for raw material compounds for paints, and even if they were used for searching for raw material compounds for paints, they would not be satisfactory in terms of search accuracy, etc.

[0007] The problem that this invention aims to solve is to provide a method for searching for raw material compounds for paints that can be searched quickly, accurately, and easily. [Means for solving the problem]

[0008] The inventors, after diligently studying to solve the above problems, discovered that the above problems could be solved by a method for searching for raw material compounds for paints that comprises a specific process, and thus completed the present invention. That is, the present invention relates to the following items 1 to 9. [Item 1] A method for searching for a raw material compound for paint, comprising a selection step of selecting candidate compounds for paint raw material compounds using a judgment model obtained by machine learning, and a confirmation step of conducting experiments to confirm the properties of the candidate compounds for paint raw material compounds selected in the selection step, wherein the raw material compound for paint is one or more selected from the group consisting of a curing agent, a curing agent component, a resin for forming a coating film, an additive, a pigment, and a solvent. [Item 2] The method for searching for raw material compounds for paints according to Item 1, wherein the selection step comprises a virtual library creation step of creating a focused library for searching for candidate raw material compounds for paints as a virtual library. [Item 3] The method for searching for raw material compounds for paint according to item 1 or 2, wherein the selection step comprises a virtual library creation step, a judgment model creation step, a candidate compound determination step for raw material compounds for paint, a judgment model suitability determination step, and a virtual library suitability determination step. [Item 4] The method for searching for a raw material compound for paint according to any one of Items 1 to 3, wherein the selection step comprises the steps of obtaining an objective variable from data relating to the raw material compound for paint, obtaining explanatory variables by analyzing the data relating to the raw material compound for paint, and creating a decision model that predicts the objective variable from the explanatory variables by performing machine learning using the objective variable and the explanatory variables. [Item 5] The method for searching for raw material compounds for paints according to any one of Items 1 to 4, wherein the selection step involves creating a judgment model by machine learning using one or more different algorithms, and selecting and using one or more judgment models with high regression or classification indicators or consensus models from among the created judgment models. [Clause 6] The method for searching for a raw material compound for paint according to any one of Clauses 1 to 5, wherein the selection step includes a step of determining whether a candidate compound for a raw material compound for paint satisfies the criteria, using a regression model or classification model using molecular descriptors as the determination model. [Clause 7] A method for searching for a raw material compound for paint according to any one of Clauses 1 to 6, characterized in that if the raw material compound for paint confirmed in the confirmation step does not meet the criteria, the suitability of the judgment model is determined, the process returns to the judgment model creation stage based on the judgment result, the suitability of the virtual library is then determined, and the process returns to the virtual library creation stage based on the judgment result. [Clause 8] A method for searching for a raw material compound for paint according to any one of Clauses 1 to 7, wherein the raw material compound for paint is a raw material compound for electrodeposition paint. [Item 9] A method for searching for a raw material compound for paint according to any one of items 1 to 8, wherein the raw material compound for paint is one or more of a UV absorber, a curing catalyst, or a blocking agent for an isocyanate compound. [Effects of the Invention]

[0009] The present invention provides a method for searching for raw material compounds for paints, which allows for rapid, accurate, and convenient searching of such compounds. [Brief explanation of the drawing]

[0010] [Figure 1] A flowchart of a method for searching for raw material compounds for paints according to one embodiment of the present invention. [Modes for carrying out the invention]

[0011] Embodiments of the present invention will be described in detail below with reference to the drawings. It should be understood that the present invention is not limited to the following embodiments, but also includes various modifications that are carried out without altering the essence of the invention.

[0012] Figure 1 is a flowchart of a method for searching for raw material compounds for paints according to one embodiment of the present invention. The method for searching for raw material compounds for paints shown in Figure 1 comprises a selection step of selecting candidate compounds for raw material compounds for paints, and a confirmation step of experimentally confirming the selected candidate compounds for raw material compounds for paints.

[0013] The selection process involves selecting candidate compounds for paint raw materials from a virtual library using a created judgment model, which then proceed to the verification process. In the flowchart of the paint raw material compound search method shown in Figure 1, the process proceeds from the start to the virtual library creation stage and the judgment model creation stage, where the virtual library and judgment model are created. After the virtual library is created, the process proceeds to the virtual library suitability assessment stage as needed to determine the suitability of the virtual library. If the virtual library fails the suitability assessment stage, the process returns to the virtual library creation stage; if it passes, the process proceeds to the candidate compound assessment stage for paint raw material compounds. After the determination model is created, the process proceeds to a determination model suitability determination step as necessary to determine the suitability of the determination model. In the determination model suitability determination step, if the result is unacceptable, the process returns to the determination model creation step, and if the result is acceptable, the process proceeds to a candidate compound determination step for coating raw material compounds.

[0014] After the virtual library and the determination model are respectively completed, in the candidate compound determination step, candidate compounds for coating raw material compounds that proceed to a confirmation step are determined from the virtual library using the created determination model. In the candidate compound determination step, when a large number of candidates for coating raw material compounds are determined to be acceptable, or when no acceptable determination is obtained, the result is regarded as unacceptable, and the process proceeds to a virtual library suitability determination step and / or a determination model suitability determination step.

[0015] In the virtual library suitability determination step, the suitability of the virtual library is determined based on information obtained in the candidate compound determination step as necessary, and if the result is unacceptable, the process proceeds to the virtual library creation step to recreate or correct the virtual library. In the determination model suitability determination step, the suitability of the determination model is determined based on information obtained in the candidate compound determination step as necessary, and if the result is unacceptable, the process proceeds to the determination model creation step to recreate or correct the determination model. Note that when an unacceptable result is obtained in the candidate compound determination step, at least one of the virtual library and the determination model is recreated or corrected via the virtual library suitability determination step and / or the determination model suitability determination step based on the cause of unacceptability in the candidate compound determination step.

[0016] The selection step is finally completed when an appropriate number of coating raw material compounds are determined to be acceptable in the candidate compound determination step.

[0017] The verification process consists of, in order, a candidate compound acquisition stage in which candidate compounds selected in the selection stage are obtained, a candidate compound experiment stage, and a candidate compound final evaluation stage. If a candidate compound is deemed acceptable in the final evaluation stage, the search for raw material compounds for paints is completed. If it is deemed unacceptable, the process proceeds to the virtual library compatibility evaluation stage and / or the evaluation model compatibility evaluation stage in the selection stage.

[0018] In the virtual library suitability assessment stage, the suitability of the virtual library is assessed, taking into account the information obtained in the candidate compound assessment stage as needed. If it fails the assessment, the process proceeds to the virtual library creation stage, where the virtual library is recreated and modified. In the judgment model suitability assessment stage, the suitability of the judgment model is assessed, taking into account the information obtained in the candidate compound assessment stage as needed. If the model fails to meet the criteria, the process proceeds to the judgment model creation stage, where the judgment model is recreated and modified. If a candidate compound fails the final evaluation stage, at least one of the virtual library and the evaluation model will be recreated and modified based on the reasons for failure in the final evaluation stage, through the virtual library compatibility evaluation stage and / or the evaluation model compatibility evaluation stage.

[0019] The present invention's method for searching for raw material compounds for paints improves the accuracy of the search by conducting experiments in the confirmation step, and if a compound fails the confirmation step, it can improve the prediction accuracy by using the experimental data from that failure to repeat the search for the raw material compound for paints. The details are explained below.

[0020] [Raw material compounds for paints] In the method for searching for raw material compounds for paints according to the present invention, the raw material compound for paints is a raw material compound used in a wide variety of paints. Examples of paints include vehicle paints, architectural paints, household paints, industrial paints, structural paints, antifouling paints for ships, conductive paints, etc. Examples include topcoat paints, intermediate coat paints, base coats, electrodeposition paints, primers, rust-inhibiting paints, etc. In the method for searching for raw material compounds for paints according to the present invention, the raw material compound for paints is preferably a raw material compound for electrodeposition paints.

[0021] In the method for searching for raw material compounds for paints according to the present invention, the raw material compound for paints is any compound used to constitute a paint. For example, it is one or more selected from the group consisting of curing agents, curing agent components, film-forming resins, additives, pigments, and solvents. Examples of curing agent components include curing agent reactivity blocking agents. Examples of additives include one or more ultraviolet absorbers, light stabilizers, thickeners, defoamers, surface modifiers, settling inhibitors, rust inhibitors, chelating agents, dehydrating agents, neutralizing agents, plasticizers, fillers, pigments, and curing catalysts. In the method for searching for raw material compounds for paints according to the present invention, the raw material compound for paints is preferably one or more of the following: an ultraviolet absorber, a curing catalyst, or a blocking agent for isocyanate compounds.

[0022] [Selection process] In the method for searching for raw material compounds for paints according to the present invention, the selection step is a step of selecting candidate compounds for raw material compounds for paints from a created virtual library using a judgment model obtained by machine learning. The selection process comprises at least a virtual library creation stage, a judgment model creation stage, a virtual library compatibility determination stage, a judgment model compatibility determination stage, and a candidate compound determination stage for paint raw material compounds. The selection process may include a step of selecting candidate compounds for paint raw material compounds using a regression model or classification model with molecular descriptors. For example, a step of selecting candidate compounds for paint raw material compounds using a regression model or classification model with molecular descriptors may include a candidate compound determination step in which the determination model is a regression model or classification model with molecular descriptors, and candidate compounds for paint raw material compounds are selected using such a determination model.

[0023] <Virtual library creation stage> The virtual library creation stage involves creating a focused library for searching for candidate raw material compounds for paints, as a virtual library. A focused library is a library of compounds that have substructures or molecular structures that can be expected to have specific effects or functions as paints. If there are substructures that are considered desirable as raw material compounds for paints, or substructures that should be avoided, it is preferable to create a focused library that reflects this information during the virtual library creation stage. When creating a virtual library, candidate compounds and new compounds can be selected based on empirical rules. When creating a virtual library, candidate compounds can be selected by searching databases or other sources. In addition to internal databases, public databases such as PubChem, ZINC, ChemSpider, Dortmund Data Bank, and SPRESI database are preferred as databases.

[0024] In the selection process, the virtual library creation stage may be set before the candidate compound determination stage. The virtual library creation stage can be set before the determination model creation stage, after the determination model creation stage, or in parallel with the determination model creation stage. One embodiment shown in Figure 1 is an example in which the virtual library creation stage is set in parallel with the determination model creation stage.

[0025] <Decision model creation stage> The judgment model creation stage is the stage in which a judgment model is created to determine whether or not the selected candidate compounds should proceed to the verification process when selecting candidate compounds for paint raw materials. The decision model is preferably one that uses machine learning with a target variable and explanatory variables, and that predicts the target variable from the explanatory variables. The decision model creation stage may include one or more phases: obtaining the objective variable from data related to paint raw material compounds, and obtaining explanatory variables by analyzing data related to paint raw material compounds.

[0026] In the method for searching for raw material compounds for paints according to the present invention, the objective variable is a variable used to evaluate whether or not a raw material compound for paints constitutes a paint that fulfills the intended purpose (satisfies the characteristics). Examples of objective variables include one or more of the following: properties of the coating film (gel fraction, elongation, gloss, haze, light transmittance, storage modulus, loss tangent, solar reflectance, ultraviolet transmittance, heat resistance, water resistance, color tone, acid resistance, chemical resistance, weather resistance, surface roughness, gloss, shine, smoothness, adhesion, contact angle, scratch resistance, electrical resistance, stain resistance, light reflectance, flip-flop properties, micro-glossiness, coating film appearance, coating film standards, etc.), properties of the paint (storage stability, low-temperature curing properties, blocker dissociation temperature, viscosity, dispersibility, defoaming properties, foaming properties, color, viscoelasticity, stringiness, sagging properties, surface tension, solid content concentration, pH, repellency, paint standards, etc.), and properties of the paint components (constituent atoms, environmental properties, heat resistance, reactivity, toxicity, cost, availability, legal regulations, safety, etc.).

[0027] In the method for searching for raw material compounds for paints according to the present invention, the explanatory variables are variables that affect one or more of the following: the properties of the raw material compound for paints, the properties of the paint using it, or the objective variable. Examples of explanatory variables include: the type of paint component, the content ratio of the paint component, the properties of the paint component (constituent atoms, environmental properties, heat resistance, reactivity, toxicity, safety, cost, availability, legal regulations, molecular weight, surface tension, contact angle, solid content concentration, pH, viscosity, SP value, solubility, vapor pressure, evaporation rate, particle size, etc.), the physical properties of the paint (storage stability, low-temperature curing properties, blocker dissociation temperature, viscosity, dispersibility, defoaming properties, foaming properties, color, viscoelasticity, stringiness, sagging properties, surface tension, solid content concentration, pH, repellency, paint-related standards, etc.), and the manufacturing conditions of the paint (stirring conditions, manufacturing temperature, etc.). One or more of the following can be mentioned: painting conditions (temperature, humidity, film thickness, drying temperature, etc.), paint application (electrodeposition paint, intermediate coat paint, topcoat paint, basecoat paint, glossy paint, primer, etc.), and paint film characteristics (film thickness, layer structure, elongation, gloss, haze, light transmittance, storage modulus, loss tangent, solar reflectance, ultraviolet transmittance, heat resistance, water resistance, color tone, water resistance, acid resistance, chemical resistance, weather resistance, surface roughness, gloss, shine, smoothness, adhesion, contact angle, scratch resistance, electrical resistance, stain resistance, light reflectance, flip-flop properties, micro-glossiness, paint film appearance, standards related to the paint film, etc.).

[0028] Molecular descriptors can be used as explanatory variables. This makes it possible to quantify the explanatory variables, which is advantageous when generating a decision model using machine learning.

[0029] Molecular descriptors are numerical representations of molecular structure, characterized by their features. Molecular descriptors consist of experimental values, 0-dimensional descriptors, 1-dimensional descriptors, 2-dimensional descriptors, 3-dimensional descriptors, and 4-dimensional descriptors.

[0030] Experimental measurements are descriptors that are considered universal properties, and examples include log P, refractive index, dipole moment, polarity, molecular weight, and NMR shift.

[0031] A 0-dimensional descriptor is a descriptor that can be positioned as a constructor descriptor or count descriptor, and examples include the number of atom types, bonding modes, the number of atoms such as C, H, O, N, halogens, the number of heavy atoms, the number of bonds, the number of rotatable bonds, the number of multiple bonds, and so on.

[0032] One-dimensional descriptors are descriptors that can be described as fragment lists or fingerprints, and examples include the number of hydrogen bond donors and acceptors, the number of rings, the number of functional groups, and fingerprints (structure keys, hashed fingerprints).

[0033] Two-dimensional descriptors are descriptors that can be positioned as topological indices, and examples include topological indices (indexes that quantify the connection patterns based on the adjacency matrix) and Wiener indices (the sum of the shortest distances between atoms within a molecule).

[0034] Three-dimensional descriptors are descriptors that can be categorized as geometric descriptors or quantum chemical descriptors. Examples include geometric descriptors, polar surface area, 3D-MoRSE descriptors (3D molecular representation based on electron diffraction), WHIM descriptors (calculated from atomic coordinates of molecular structure), GETAWAY descriptors (calculated by calculating the molecular influence matrix and combining it with geometric atomic distances), HOMO / LUMO, charge, dipole moment, polarizability, binding energy, reaction heat, activation barrier, IR frequency and intensity, and NMR chemical shift.

[0035] A four-dimensional descriptor is a descriptor that can be positioned as an interaction descriptor, and examples include the GRID method (extensive conformational search and narrowing down of active conformation candidates), the CoMFA method (a regression analysis method for predicting interactions between proteins and small molecules), and Volsurf (calculated from the interaction energy between a molecule and atoms placed at surrounding grid points). Of these molecular descriptors, those obtained through calculation can be acquired from the SMILES notation compound information obtained from the aforementioned public database using publicly available software such as RDKit, Mordred, or commercially available software such as alvaDesc and Codessa. Furthermore, when using quantum chemical calculation results as molecular descriptors, programs such as Gaussian, GAMESS, Jaguar, Q-Chem, and Spartan can also be used.

[0036] It is not easy to directly link the structure and composition of a material with its physical properties and characteristics. Therefore, molecular descriptors, which are surrogate indices governing physical properties, are often used to clarify the relationship between the material's structure and composition and these surrogate indices before predicting physical properties. The use, examination, and exploration of these surrogate indices are crucial for creating effective judgment models. Machine learning involves extracting one or more surrogate indices (molecular descriptors) that correlate with the material's physical properties, which are the target variables. Creating a more accurate machine learning model leads to the selection of better candidate compounds.

[0037] In the stage of creating a decision model, the algorithm used for machine learning when creating the decision model is not particularly limited, regardless of whether it is regression or classification, as long as it is a machine learning algorithm capable of creating a decision model. Regression algorithms include CAT-boosted regression, gradient-boosted regression, extra-tree regression, random forest regression, XG-boosted regression, LIGHTGBM regression, ADA-boosted regression, linear regression, ridge regression, minimum-angle regression, Bayesian linear regression, decision tree regression, elastic network regression, Lasso regression, Hoover regression, orthogonal matching tracking, K-nearest neighbor regression, Lasso minimum-angle regression, PA, relevance determination, RANSAC, Teilsen regression, kernel ridge regression, support vector machine regression, multilayer perceptron, partial least-squares regression, Gaussian process regression, etc., or a consensus model using one or more algorithms can be used. A consensus model for regression is a model that calculates the average value of models using one or more of the regression algorithms with high regression indices. As classification algorithms, random forest classification, Lite GBM classification, XG-Boost classification, Cat-Boost classification, extra-tree classification, gradient boosting classification, decision tree classification, Ada-Boost classification, K-nearest neighbor classification, logistic regression, linear discriminant analysis, ridge classification, support vector machine (linear kernel), naive Bayes, quadratic discriminant analysis, etc., or a consensus model using one or more algorithms can be used. A consensus model for classification is a model that calculates the majority vote of models using one or more of the classification algorithms with high classification indices. Regardless of whether it is regression or classification, it is preferable to optimize the hyperparameters in each algorithm. Regression metrics that can be used include the coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the square root of mean squared logarithmic error (RMSLE). Classification metrics that can be used include accuracy, AUC (Area Under the Curve), recall, precision, F1 (F1-score), Kappa coefficient, and Matthews correlation coefficient (MCC).

[0038] When creating a decision model, an appropriate machine learning framework or library can be used depending on the data size, execution environment, etc. The machine learning framework or library may be one that you have developed yourself, or one or more of the available ones may be used. Examples of machine learning frameworks and libraries include one or more of TensorFlow, PyTorch, Keras, Caffe, Microsoft Cognitive Toolkit, MXNet, Chainer, Scikit-learn, OpenCV, spaCy, Caret, and PyCaret.

[0039] In machine learning, the input of the target variable and explanatory variables can be performed by transmitting data via wired, wireless, or a combination thereof communication methods, or via recording media. Examples of inputs using communication methods include one or more combinations of various communication networks such as LAN (Local Area Network), WAN (Wide Area Network), the Internet, and telephone networks. Inputs via recording media can be performed by reading data from recording media such as magnetic recording media, optical recording media, and paper recording media using appropriate reading means.

[0040] In machine learning, the reliability of the dependent and independent variables can be enhanced by improving consistency (consistency of labeling criteria), reproducibility (highly reproducible labeling for the same data), and transparency (clarity of the data creation process and the basis for labeling).

[0041] In machine learning, it is preferable to preprocess the data related to the target variable and explanatory variables. Examples of preprocessing include data processing such as imputation of missing data, filtering such as removal of outliers, encoding to convert qualitative variables (text data) to quantitative variables (numerical data), data normalization / standardization, and arithmetic processing (logarithmic scaling, power scaling, etc.). This makes the data applicable to various machine learning algorithms, and also allows for data specific to particular machine learning algorithms.

[0042] In the selection process, a decision model can be created using machine learning with one or more different algorithms. In the selection process, one or more of the created judgment models can be selected and used if they have high regression or classification evaluation metrics, or if they are consensus models. In the selection process, a decision model can be created using machine learning with one or more different algorithms, and one or more of the created decision models that have high regression or classification indicators, or a consensus model, can be selected and used. In the selection process, the judgment model creation stage may be set before the candidate compound determination stage. The judgment model creation stage can be set before the virtual library creation stage, set after the virtual library creation stage, or set in parallel with the virtual library creation stage. One embodiment shown in Figure 1 is an example in which the judgment model creation stage is set in parallel with the virtual library creation stage.

[0043] <Virtual Library Compatibility Assessment Stage> The virtual library suitability assessment stage is the stage in which, if a candidate compound fails the candidate compound assessment stage in the selection process or the final candidate compound assessment stage in the confirmation process, it is determined, based on existing knowledge and newly acquired knowledge, whether the candidate compound is suitable as a focused library (virtual library) to which the assessment model should be applied. As long as a virtual library is used in which no candidate compounds that meet the initial characteristics are found, no candidate compounds will be obtained, so the virtual library creation stage is a very important step.

[0044] In the selection process, the virtual library compatibility determination stage may be set after the virtual library creation stage, before the determination model creation stage, or after the determination model creation stage. It can also be set in parallel with the determination model creation stage.

[0045] <Determination stage for suitability of the judgment model> The judgment model suitability assessment stage is the stage in which, if a candidate compound fails the candidate compound assessment stage in the selection process or the final candidate compound assessment stage in the confirmation process, the accuracy of the judgment model is examined, and it is considered whether it is necessary to recreate the judgment model (improve the judgment model).

[0046] In the candidate compound determination stage of the selection process or the final candidate compound determination stage of the confirmation process, explanatory variables for each candidate compound are obtained. In particular, in the final candidate compound determination stage of the confirmation process, the correspondence between the explanatory variables and the dependent variable is obtained. In the stage of determining the suitability of the decision model, the necessity of recreating the decision model (improving the decision model) is considered based on these explanatory variables or the correspondence between the explanatory variables and the dependent variable.

[0047] In the method for searching for raw material compounds for paints according to the present invention, if a candidate compound fails the final judgment stage in the confirmation process, the prediction accuracy of the judgment model can be further improved by repeatedly recreating the judgment model (improving the judgment model).

[0048] In the judgment model suitability determination stage, if it is determined that the judgment model does not need to be recreated, or if it is determined that the judgment model needs to be recreated and a judgment model is created, the process proceeds to the virtual library suitability determination and creation stage. If it is determined that the virtual library needs to be recreated, a focused library of new candidate compounds is created and the process proceeds to the candidate compound determination stage. If it is determined that the virtual library does not need to be recreated, the process proceeds to the candidate compound determination stage. If the process returns to the virtual library creation stage, it is preferable to create the library using a method other than the previously used library creation method (for example, compounds with different substructures). If it is necessary to recreate the judgment model, experiments may be conducted using the candidate compounds that failed the test, the target variable may be obtained, and the judgment model may be recreated (improved) using the obtained target variable.

[0049] If it is determined during the judgment model suitability assessment stage that the judgment model needs to be recreated, the process returns to the judgment model creation stage, and machine learning is performed again to recreate (improve) the judgment model. Furthermore, if a candidate compound fails the final evaluation stage in the verification process, it has been proven through experiments that it cannot exhibit the desired properties. Therefore, by adding the obtained correspondence between explanatory variables and the objective variable to the data used to create the evaluation model and subjecting it to machine learning, it is possible to improve the trained evaluation model and enhance its accuracy.

[0050] <Candidate compound selection stage> The candidate compound determination stage involves using the determination model created in the determination model creation stage to predict whether or not the candidate compounds in the virtual library meet the desired characteristics. In the candidate compound selection stage, in addition to selection using a selection model, factors such as availability, ease of acquisition, synthetic suitability, price, environmental characteristics, and safety can be used in the selection process. The selection model used in the candidate compound determination stage is preferably a machine learning model that includes explanatory variables as input and outputs an objective variable. In the candidate compound evaluation stage, it is preferable to predict the target variable by inputting the explanatory variables of candidate compounds selected from a virtual library into the evaluation model, passing candidate compounds that meet the threshold for the target variable, and failing candidate compounds that do not meet the threshold for the target variable. In the candidate compound identification stage, electronic devices such as computers and other electronic equipment can be used to easily identify a large number of candidate compounds. If there are many candidate compounds that meet the initial threshold, the threshold may be changed to narrow down the number of candidates.

[0051] If a candidate compound passes the candidate compound evaluation stage, it proceeds to the verification stage. If a candidate compound fails the candidate compound evaluation stage, the process proceeds to the evaluation model suitability evaluation stage, and then to the virtual library suitability evaluation stage.

[0052] [Confirmation process] In the method for searching for raw material compounds for paints according to the present invention, the confirmation step comprises, in order, a candidate compound acquisition step for obtaining the candidate compounds selected in the selection step, a candidate compound experiment step, and a candidate compound final determination step.

[0053] <Candidate compound acquisition stage> The candidate compound acquisition stage is the stage in which the candidate compounds for paint raw materials selected in the selection process are specifically obtained. Possible means of obtaining candidate compounds include, for example, commercial acquisition based on catalogs, acquisition by synthesis based on literature, etc., or one or more other methods.

[0054] <Candidate compound experimental stage> The candidate compound experimentation stage involves obtaining specific candidate compounds for paint raw materials selected in the selection process, preparing paint using these candidate compounds, and acquiring data on the properties of the candidate compounds, the paint, and the coating film. During the candidate compound experimentation phase, the numerical values ​​of the target variable will be specifically obtained.

[0055] <Final stage of candidate compound evaluation> The final stage of candidate compound evaluation involves obtaining the candidate compounds selected in the selection process, conducting experiments to confirm the objective variables obtained, and determining whether they can be used as raw material compounds for paints. In the final stage of candidate compound evaluation, the suitability of use can be determined based on the objective variables, which are characteristic values ​​obtained in the candidate compound experiment stage.

[0056] If a candidate compound is deemed acceptable at the final evaluation stage, the search for raw material compounds for paints will be terminated. If a candidate compound fails to meet the criteria during the final evaluation stage, it proceeds to the evaluation model suitability evaluation stage in the selection process. The suitability of the evaluation model is assessed, and based on the evaluation result, it proceeds to the evaluation model creation stage. Next, it proceeds to the virtual library suitability evaluation stage. The suitability of the virtual library is assessed, and based on the evaluation result, it proceeds to the virtual library creation stage.

[0057] In the final stage of candidate compound determination, electronic computers, etc., can be used. When creating or recreating a virtual library, candidate compounds can be selected based on empirical rules. When creating or recreating a virtual library, candidate compounds can be selected by searching a database or similar source. The search process may include a step where a machine learning model using molecular descriptors is used to search for candidate compounds for paint raw materials. The virtual library creation step may also involve identifying compounds with novel structures. During the virtual library creation and virtual library recreation stages, explanatory variables for the selected candidate compounds are obtained.

[0058] [Applications of methods for discovering raw material compounds for paints] The present invention's method for searching for raw material compounds for paints makes it possible to predict the state of the paint produced and the coating film produced from that paint when a specific raw material compound for paints is used as a raw material for paints. Therefore, it can be suitably used in applications where it is necessary to produce paints or coating films with desired properties, or to predict changes in the properties of the produced paints or coating films. For example, it can be used to accelerate the development of paints and coatings that meet required characteristics, and to predict the function and properties of coatings, making it useful in the development and proposal of paints.

[0059] [System for searching for raw material compounds for paints] The present invention provides a system for searching for raw material compounds for paints, which includes a determination model that determines whether a raw material compound satisfies predetermined characteristics. The determination model is generated by machine learning using data in which the target variable and explanatory variables are correlated.

[0060] A system for searching for raw material compounds for paints may be configured within a computer. Specifically, it includes a judgment model (machine learning model), a target variable prediction unit, a virtual library database, a reception unit, a display unit, and a recording unit. The database may be an external database accessed via a communication line. If necessary, it may be possible to connect to an external user terminal (one or more devices such as a personal computer, mobile phone (smartphone), or personal digital assistant) via a communication line. The display unit may be able to connect to a virtual space, and may be configured to display the predicted state of paints and coatings obtained using the raw material compounds in the virtual space. For external users, it may be possible to charge at least some of them for using the system or for some of the operations performed while using the system. [Examples]

[0061] [Search for blocking agents for blocked polyisocyanates] This document presents an example of searching for a blocking agent for blocked polyisocyanates used in thermosetting paints using the paint raw material compound search method of the present invention.

[0062] Using past in-house experimental data and publicly available literature, a dataset was created with the curing start and end temperatures of the blocking agent in blocked polyisocyanates as the dependent variable, and the mordred molecular descriptor of the blocking agent as the independent variable. Next, machine learning was performed using all applicable algorithms, and finally, the random forest algorithm, which had the smallest mean absolute error (MAE), was selected as the algorithm. After optimizing the hyperparameters, a decision model (machine learning model) was created with the curing start and end temperatures of the blocking agent in blocked polyisocyanates as the dependent variable.

[0063] From the database (PubChem: approximately 120 million compounds), a search was performed on substructures extracted from previously used blocking agents. Filtering by molecular weight and heteroatom count, 1428 compounds were selected and formed a virtual library. Next, undesirable structures were removed to select 882 compounds. Using the aforementioned random forest model, the curing start and end temperatures of the blocked polyisocyanates were judged to be acceptable if they were 140°C or lower, and 9 compounds were selected.

[0064] Experiments were conducted to verify the properties of the nine selected compounds, and all of them were found to be unsuitable. The evaluation model determined them to be suitable, but the virtual library determined them to be unsuitable, so the virtual library was recreated.

[0065] A search was performed using a different substructure as the search key from the first search in the database (PubChem). Filtering by molecular weight and structural complexity, 8128 compounds were selected, and a virtual library was created. Next, undesirable structures were removed, and 7238 compounds were selected. From the 7238 selected compounds, a random forest model was used to determine which compounds passed the test, with curing start temperature below 90°C and curing end temperature below 130°C. This resulted in the selection of 668 compounds. From these 668 compounds, 38 compounds were selected based on their availability, and from these 38 compounds, four compounds were selected based on price and synthetic suitability.

[0066] For the four selected compounds, we obtained them, conducted experiments, and measured the curing start temperature and curing end temperature of the blocking agent when it was used to construct blocked polyisocyanates. Ultimately, two compounds were identified: N-methylaniline (curing start temperature 122°C, curing end temperature 134°C) and tetrahydroquinoline (curing start temperature 109°C, curing end temperature 125°C). Conventional blocking agents have a curing start temperature of 151°C and a curing end temperature of 168°C.

[0067] Blocked polyisocyanates were synthesized using N-methylaniline or tetrahydroquinoline as blocking agents, and paints were prepared using these. Each of the obtained paints was applied to a cold-rolled steel sheet that had undergone chemical conversion treatment, and cured by baking at 130°C for 20 minutes. The storage stability and finish quality were then checked and found to be equivalent to that of conventional paints.

Claims

1. A selection process for selecting candidate compounds for paint raw material compounds using a judgment model obtained by machine learning, and A confirmation step is to conduct experiments using candidate compounds of paint raw material compounds selected in the selection step to confirm their properties. A method for searching for raw material compounds for paints, comprising: The aforementioned raw material compound for paint is one or more selected from the group consisting of curing agents, curing agent components, film-forming resins, additives, pigments, and solvents. A method for searching for the aforementioned raw material compounds for paints.

2. The method for searching for raw material compounds for paints according to claim 1, wherein the selection step comprises a virtual library creation step of creating a focused library for searching for candidate raw material compounds for paints as a virtual library.

3. The method for searching for raw material compounds for paints according to claim 1 or 2, wherein the selection step comprises a virtual library creation step, a judgment model creation step, a candidate compound determination step for raw material compounds for paints, a judgment model suitability determination step, and a virtual library suitability determination step.

4. The method for searching for a raw material compound for paint according to claim 1 or 2, wherein the selection step comprises the steps of obtaining an objective variable from data relating to a raw material compound for paint, obtaining explanatory variables by analyzing data relating to a raw material compound for paint, and creating a decision model that predicts the objective variable from the explanatory variables by performing machine learning using the objective variable and explanatory variables.

5. The method for searching for raw material compounds for paints according to claim 1 or 2, wherein the selection step involves creating a judgment model by machine learning using one or more different algorithms, and then selecting and using one or more of the created judgment models that have high regression or classification indicators or a consensus model.

6. The method for searching for a raw material compound for paint according to claim 1 or 2, wherein the selection step includes a step of determining whether a candidate compound for a raw material compound for paint satisfies the criteria, using a regression model or classification model using molecular descriptors as the determination model.

7. A method for searching for a raw material compound for paint according to claim 1 or 2, characterized in that, if the raw material compound for paint confirmed in the confirmation step does not meet the criteria, the suitability of the judgment model is judged, and based on the judgment result, the process returns to the judgment model creation stage, then the suitability of the virtual library is judged, and based on the judgment result, the process returns to the virtual library creation stage.

8. The method for searching for a raw material compound for paint according to claim 1 or 2, wherein the raw material compound for paint is a raw material compound for electrodeposition paint.

9. The method for searching for a raw material compound for paint according to claim 1 or 2, wherein the raw material compound for paint is one or more of a UV absorber, a curing catalyst, or a blocking agent for isocyanate compounds.

Images