Estimation model generation method / generation device for estimating reaction conditions, reaction condition provision method / provision device, and program

Abstract

In the present invention, information relating to a plurality of chemical substances and products to be reacted, which is acquired, is similar to information relating to a plurality of chemical substances and products, which is set for reaction conditions. A plurality of reaction conditions obtained when a plurality of chemical substances to be reacted are reacted are set from reaction items set for the reaction conditions. Furthermore, an estimation model is used which is obtained by performing machine learning using, as training data, chemical structure information and physical property information of a plurality of chemical substances that have been reacted in the past, reaction conditions in the reaction, chemical structure information and physical property information of a product that has been generated, and a yield when the reaction is performed under the reaction conditions. A yield is estimated for each of a plurality of reaction conditions, and a reaction condition for which the yield satisfies a predetermined condition among the estimated yields is displayed (output).

Description

Technical Field

[0001] The present invention relates to an estimation model generation method and estimation model generation device, a reaction condition providing method and reaction condition providing device, and a program for estimating reaction conditions whose yield satisfies predetermined conditions from information related to a plurality of chemical substances and information related to products. Background Art

[0002] In the past, chemists generated desired products by reacting a variety of chemical substances. In addition, in order to efficiently generate desired products, chemists repeatedly perform chemical experiments while changing reaction conditions for each experiment based on their own empirical rules, thereby exploring reaction conditions whose yields meet predetermined conditions (for example, reaction conditions that provide higher yields, etc.).

[0003] However, in order to explore the reaction conditions that yield meets the predetermined conditions, chemists must rely on themselves to determine the reaction conditions by trial and error and repeatedly implement chemical experiments. Therefore, a large amount of operating man-hours are required, and there is a problem of improving operating efficiency. In order to solve this problem, a flow synthesis technology (for example, patent document 1) that can automatically set the reaction conditions suitable for the synthesis reaction by reflecting the analysis results based on artificial intelligence is known.

[0004] Patent document 1 discloses a flow synthesis device that automatically sets reaction conditions that include at least one of reaction temperature, reaction time, reagent type, reagent amount, and solvent as elements. The flow synthesis device is controlled by an information processing device that analyzes synthesis reaction information contained as a database using artificial intelligence having structural fingerprints and reaction conditions as explanatory variables.

[0005] Reference List

[0006] Patent Literature

[0007] Patent Document 1: Japanese Unexamined Patent Application Publication No. 2019-185506 Summary of the invention

[0008] Problems to be solved by the invention

[0009] However, in the technology disclosed in Patent Document 1, only a vector combining structural fingerprint vector information and reaction condition vector information is used as an explanatory variable, and machine learning is performed based thereon (i.e., at least physical property information (information showing the physical and / or chemical characteristics of a chemical substance) is not considered as an explanatory variable). Therefore, the accuracy of the generated prediction model is insufficient, and as a result, the yield may not reach the desired level under the set reaction conditions, and improvement of this situation is desired.

[0010] The present invention has been made in view of the above-mentioned conventional problems, and an object of the present invention is to reduce the number of man-hours required for setting reaction conditions so that the yield satisfies a predetermined condition.

[0011] Means for solving problems

[0012] That is, the inference model generation method of the present invention includes: an acquisition step of acquiring information related to multiple chemical substances in past reactions, reaction conditions in the reactions, information related to products generated by reacting the multiple chemical substances, and the yield when the reaction is carried out under the reaction conditions; a first generation step of generating training data from chemical structure information and physical property information derived from each piece of information related to the multiple chemical substances in the past reactions, the reaction conditions, chemical structure information and physical property information derived from information related to the products, and the yield; and a second generation step of generating an inference model by performing machine learning using the above-mentioned training data, the above-mentioned inference model taking information related to the multiple chemical substances to be reacted and information related to the products to be generated as input values, and taking reaction conditions under which the yield of the above-mentioned products to be generated meets predetermined conditions as output values.

[0013] The reaction condition providing device of the present invention comprises: an acquisition unit, which acquires information related to a plurality of chemical substances as reaction objects and information related to products to be generated; a storage unit, which stores one or more reaction conditions; a setting unit, which sets a plurality of reaction conditions for reacting the plurality of chemical substances as reaction objects from reaction items set for one or more reaction conditions when the similarity between the information related to the plurality of chemical substances as reaction objects and the information related to products to be generated and the information related to the plurality of chemical substances and the information related to products set for the reaction conditions is equal to or greater than a predetermined threshold value; and an estimation unit, which estimates information related to the plurality of chemical substances as reaction objects using an estimation model. The chemical structure information and physical property information derived from each piece of information, the chemical structure information and physical property information derived from the information related to the above-mentioned product to be generated, and multiple set reaction conditions are used to estimate the yield of each of the above-mentioned multiple reaction conditions, wherein the estimation model is obtained by performing machine learning using the chemical structure information and physical property information derived from each piece of information related to multiple chemical substances in past reactions, the reaction conditions in the reaction, the chemical structure information and physical property information derived from the information related to the product generated in the reaction of the above-mentioned multiple chemical substances, and the yield when the reaction is carried out under the above-mentioned reaction conditions as training data; and a display control unit is used to enable the display device to display the reaction conditions whose yields meet predetermined conditions among the yields estimated for each reaction condition by the estimation unit.

[0014] Effects of the Invention

[0015] According to the present invention, it is possible to reduce the number of man-hours required to set reaction conditions so that the yield satisfies a predetermined condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1A It is a diagram for explaining a model generating device.

[0017] Figure 1B It is a diagram for explaining a device for providing reaction conditions.

[0018] Figure 2 This is a block diagram illustrating the functional structure of the model generation device.

[0019] Figure 3 is a diagram illustrating the chemical structure information of benzene.

[0020] Figure 4A This is a block diagram for explaining the functional configuration of the first feature quantity reduction unit of the feature quantity selection unit.

[0021] Figure 4B This is a block diagram for explaining the functional configuration of the feature quantity selection model generation unit of the feature quantity selection unit.

[0022] Figure 5A This is a block diagram for explaining the functional configuration of a multi-estimation-model generation unit of the estimation model generation unit.

[0023] Figure 5B This is a block diagram for explaining the functional configuration of the estimated model mixing unit of the estimated model generating unit.

[0024] Figure 6 This is a flowchart illustrating the procedure of processing for generating an estimation model in the model generating device.

[0025] Figure 7 1 is a flowchart illustrating the steps of a process of deriving a molecular descriptor in a molecular descriptor deriving unit.

[0026] Figure 8 A diagram illustrating a process of generating a data set (training data) is shown.

[0027] Fig.9A This is a diagram for explaining missing value processing in the missing value processing unit.

[0028] Fig. 9B This is a diagram for explaining missing value processing in the missing value processing unit.

[0029] Fig. 9C This is a diagram for explaining missing value processing in the missing value processing unit.

[0030] Fig.9D This is a diagram for explaining missing value processing in the missing value processing unit.

[0031] Fig.10 It is a diagram for explaining the processing in the category value conversion unit.

[0032] Fig.11 This is a flowchart for explaining the procedure of processing for selecting a feature amount (feature amount group) in a feature amount selecting unit.

[0033] Fig.12 This is a flowchart explaining the procedure of the process of reducing (deleting) the feature amount in the first feature amount reducing unit.

[0034] Fig.13A This is a diagram for explaining the correlation coefficient and the process of reducing the feature amount based on the correlation coefficient.

[0035] Fig. 13B This is a diagram for explaining the correlation coefficient and the process of reducing the feature amount based on the correlation coefficient.

[0036] Fig.14 This is a flowchart explaining the procedure of processing for generating a feature quantity selection model in a feature quantity selection model generating unit.

[0037] Fig.15 is a diagram illustrating a decision tree model.

[0038] Figure 16A- Fig. 16C It is a diagram for explaining the process of reducing (deleting) the feature amount in the second feature amount reducing means.

[0039] Figure 16D- Fig.16F It is a diagram for explaining the process of reducing (deleting) the feature amount in the second feature amount reducing means.

[0040] Fig.17 This is a flowchart illustrating the steps of a process for generating an estimation model in an estimation model generating unit.

[0041] Fig.18 The flowchart explains the steps of a process for generating a plurality of estimation models in a multi-estimation model generating unit.

[0042] Fig.19 The flowchart explains the steps of a process for generating one estimation model among a plurality of estimation models in a multi-estimation model generating unit.

[0043] Fig. 20 Detailed description of the process of mixing a plurality of estimation models.

[0044] Fig.21This is a flowchart for describing the processing contents (data flow) in the estimation model generation unit.

[0045] Fig. 22 It is a block diagram for explaining the functional structure of the reaction condition providing device.

[0046] Fig.23 This is a flowchart illustrating the steps of a process of providing reaction conditions in a reaction condition providing device.

[0047] Fig.24A It is a diagram showing information related to multiple chemical substances and information related to products, information related to multiple chemical substances reacted in experiments conducted in the past, information related to products generated as a result of reacting multiple chemical substances, and information related to reaction conditions when multiple chemical substances are reacted.

[0048] Fig. 24B It is a diagram showing information related to multiple chemical substances and information related to products converted into chemical structure information, information related to multiple chemical substances, information related to products generated as a result of reactions of multiple chemical substances converted into chemical structure information, and information related to reaction conditions when multiple chemical substances are reacted.

[0049] Fig.24C This is a diagram for describing the similarity calculation process in the reaction condition providing device.

[0050] Fig.24D is a diagram for describing the reaction condition selection process.

[0051] Fig.24E is a diagram for describing the reaction condition selection process.

[0052] Figure 25A- Fig.25C is a flowchart for describing the steps of processing for providing reaction conditions in the reaction condition providing device ( Fig.23 ) in S3-8 to S3-12. Fig. 25A is a diagram for describing the reaction simulation execution process in the reaction condition providing device. Fig. 25B is a diagram for describing the level allocation process. Fig.25C This is a diagram for describing a process of determining whether a reaction condition that satisfies a predetermined condition exists, and a process of selecting a reaction condition that satisfies the predetermined condition.

[0053] Fig.26 It is a diagram for describing the reaction simulation execution process.

[0054] Figure 27A- Fig.27E27A is a diagram illustrating the function between reaction conditions and yield and the processing of Bayesian optimization. FIG. 27B is a diagram illustrating the function between reaction conditions and yield and the processing of Bayesian optimization. FIG. 27C is a diagram illustrating the function between reaction conditions and yield and the processing of Bayesian optimization. FIG. 27D is a diagram illustrating the function between reaction conditions and yield and the processing of Bayesian optimization. Fig.27E It is a diagram illustrating the function between reaction conditions and yield and the process of Bayesian optimization.

[0055] Fig.28 is a graph illustrating a normal distribution.

[0056] Fig.29 1 is a flowchart illustrating the steps of a process of determining reaction conditions by performing Bayesian optimization while applying the Constant Liar Approach.

[0057] Fig.30 A diagram for explaining a sample screen showing reaction conditions displayed on a display device.

[0058] Fig.31 This is a block diagram illustrating the hardware configuration of the model generation device. DETAILED DESCRIPTION

[0059] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. The following embodiments do not limit the present invention, and all combinations of the features described in the embodiments are not necessarily necessary for the scheme of the present invention. In addition, various configurations within the scope of the gist of the present invention are also included in the present invention, and the following embodiments can be partially combined as needed.

[0060] Figure 1A is a diagram illustrating a model generation device, Figure 1B It is a diagram for explaining the reaction condition providing device according to the present embodiment. Figure 1AThe model generation device 1 shown generates training data from information related to multiple chemical substances, reaction conditions when multiple chemical substances are reacted, information about products generated by reacting multiple chemical substances, and the yield of the products when the reaction is carried out under the reaction conditions, which are obtained as input data. The model generation device 1 uses the generated training data to perform machine learning, thereby generating an estimation model (which is a regression model) that uses the reaction conditions under which the yield of the product to be generated meets predetermined conditions as output values. When multiple chemical substances are reacted, multiple products are usually generated. Therefore, here, among the multiple products, information related to the products planned to be generated or information related to the products that can be useful for generating the estimation model is input into the model generation device 1 as information related to the products generated by reacting multiple chemical substances.

[0061] Although information on a plurality of chemical substances, reaction conditions, information on products, and yields can be input into the model generating device 1 using a predetermined input device, Figure 1A As shown, the analysis device 2 can be connected to the input side of the model generation device 1, and information related to the chemical substances and reaction conditions, as well as the product generated in the reactor 3 and the yield of the product can be directly input from the analysis device 2 (more precisely, the information processing device attached to the analysis device 2) to the model generation device 1. The external storage device 4 can be connected to the output side of the model generation device 1, and the generated inference model can be stored in the external storage device 4.

[0062] Figure 1B The reaction condition providing device 5 shown is a device for estimating the reaction conditions when a product to be generated is generated using a chemical substance, and automatically providing the estimated reaction conditions. The reaction condition providing device 5 uses the model generating device 1 ( Figure 1A ) and estimates the yield of each of the plurality of reaction conditions based on the following information: chemical structure information and physical property information derived from information related to the plurality of chemical substances obtained as input data from the predetermined operation device 6; chemical structure information and physical property information derived from information related to products generated by reacting the plurality of chemical substances; and a plurality of reaction conditions when reacting the plurality of chemical substances, which are set by one or more reaction items stored in the storage device (storage unit) based on the chemical structure information and physical property information derived from the information related to the plurality of chemical substances obtained as input data and the chemical structure information and physical property information derived from information related to the products. The reaction condition providing device 5 outputs (displays) the reaction condition whose yield satisfies the predetermined condition among the yields estimated for each reaction condition to the external display device 7. The following describes the respective embodiments of the model generating device 1 and the reaction condition providing device 5 in order.

[0063] (Model generation (Model generation device))

[0064] Figure 2 1 is a block diagram illustrating the functional configuration of the model generation device 1 according to the embodiment. Figure 2 As shown, the model generation device 1 mainly includes a control unit 11, a data acquisition unit 12, a training data generation unit 13, a pre-processing unit 14, a feature quantity selection unit 15, an estimated model generation unit 16 and a storage unit 17 as functions.

[0065] The control unit 11 is a functional module that controls the processing of each functional module. The data acquisition unit 12 acquires information related to a plurality of chemical substances, reaction conditions, information related to the product, and yield. The various data (information) acquired by the data acquisition unit 12 is temporarily stored in the storage unit 17 through the control unit 11. It should be noted that the information related to the chemical substance and the information related to the product can be described by tags such as SMILES tags, MOL files, and SDF files. In the present embodiment, the information described by the SMILES tag is acquired. The SMILES tag is a tag that converts a chemical structure into a string of letters to express it. For example, benzoic acid can be marked as "OC(C1=CC=CC=C1)=O.". In addition, here, as information related to a plurality of chemical substances, information related to two chemical substances is used in the description, that is, information related to chemical substance A and information related to chemical substance B.

[0066] As reaction conditions, solvent, temperature, concentration, dropping speed, dropping method, stirrer, etc. can be set. Among the reaction conditions, other reaction conditions (e.g., temperature, etc.) other than the reaction conditions related to a part of the compound are defined as control conditions. When a product is to be obtained from a predetermined chemical substance, the yield is expressed as the ratio between the theoretical yield that can be obtained theoretically and the amount actually obtained.

[0067] The training data generating unit 13 includes a data set generating unit 131, a molecular descriptor deriving unit 132, and a physical property information database (physical property information DB in the figure) 133, and generates training data based on various data acquired by the data acquiring unit 12. The data set generating unit 131 generates training data in the form of a predetermined data set based on the following: (1) molecular descriptors derived by the molecular descriptor deriving unit 132 for information related to chemical substances; (2) molecular descriptors derived by the molecular descriptor deriving unit 132 for information related to products; (3) reaction conditions; and (4) yield.

[0068] The molecular descriptor deriving unit 132 derives the molecular descriptor from the chemical structure information and physical property information represented as an index (numerical value) when determining the molecular structure. In the derivation of the molecular descriptor, although the Morgan method is used for explanation in this embodiment, the Mordred method, the RDkit descriptor method, etc. can be used. When the Morgan method is used, the chemical structure information is calculated, and the physical property information is obtained from a predetermined database to derive the molecular descriptor.

[0069] Regarding chemical structure information, for example, in the case of benzene, by converting part of the structure into "0" and "1", it is possible to Figure 3 Specifically, regarding benzene, among the sequences from sequence [0] to sequence

[2047] , "1" (i.e., including a partial structure) is stored for sequence

[389] , sequence

[1088] , and sequence

[1873] , and "0" is stored for the other sequences. In other words, chemical structure information can be calculated by storing "1" for sequence

[389] , sequence

[1088] , and sequence

[1873] .

[0070] The molecular descriptor deriving unit 132 obtains physical property information from a physical property information database 133 storing information indicating physical and / or chemical properties of chemical substances (physical property information). For example, for a solvent, density, refractive index, dielectric constant, dipole moment, melting point, boiling point, viscosity, etc. can be used as physical property information.

[0071] Physical property information includes information that cannot be inferred from chemical structure information alone. For example, even when the chemical substance is the same (even if the chemical structure is the same), when a crystal polymorph exists, the melting point or solubility has different physical property values, and therefore, in some cases the yield of the chemical reaction is significantly affected. Therefore, in terms of generating an inference model, by also taking physical property information into account (by mixing the input of physical property value information from a known physical property database or the input of actual physical property measurement values), it is possible to expect the accuracy of the inference model to be improved.

[0072] The molecular descriptors (i.e., the calculated chemical structure information and the obtained physical property information) derived by the molecular descriptor deriving unit 132 are added as data (elements) to the dataset by the dataset generating unit 131. Therefore, the dataset generating unit 131 adds the molecular descriptors as elements of the dataset to the information related to the chemical substance, and similarly, adds the molecular descriptors as elements of the dataset to the information related to the product, and further, adds the reaction conditions and yields obtained by the data acquiring unit 12, thereby generating training data in the form of a predetermined dataset. As a supplementary explanation, the physical property information stored in the physical property information database 133 can be as described above. Figure 1AAs shown, the update is performed by an external server device via a wired or wireless communication network.

[0073] The pre-processing unit 14 includes a category value conversion unit 141, a missing value processing unit 142, and a data standardization / normalization unit 143. When the training data generated by the training data generation unit 13 does not satisfy predetermined requirements as a data set (that is, when the expected estimation model cannot be properly generated in the estimation model generation unit 16), the pre-processing unit 14 performs predetermined missing value processing on the training data (data set) generated by the training data generation unit 13.

[0074] The category value conversion unit 141 converts the category (reaction condition) not shown as a numerical value into a numerical value. This is because the numerical processing by the calculator needs to be performed on the basis of generating the estimation model. For example, when "agitator (three-swept wing)" is set as the reaction condition, the category value conversion unit 141 converts "agitator (three-swept wing)" into "0, 1, 0". For details, refer to Fig.10 This will be described later. Methods for converting categories to numerical values ​​include Label encoding, One hot encoding, and the like.

[0075] When the training data (data set) generated by the training data generation unit 13 includes data that is determined to be missing (determined to have missing values), the missing value processing unit 142 performs a process of deleting a portion of the data or filling in the missing values. More specifically, the following processes are performed: (1) deleting rows or columns that include missing values ​​in the data set; (2) filling in rows or columns that include missing values ​​with statistics (average, median, most frequent value, etc.). The processing here refers to 9A to 9D Described later.

[0076] For each feature quantity and target variable, the data standardization / normalization unit 143 standardizes the data, or uses the maximum / minimum value to normalize the data to a real value between 0 and 1. Here, as a supplementary explanation, the processing performed by the pre-processing unit 14 is a processing performed when the training data generated by the training data generating unit 13 as a data set does not satisfy the predetermined requirements as described above. When the reaction condition satisfies the predetermined requirements (that is, for example, when the reaction condition does not need to be converted into a numerical value, when the training data does not include a missing value, etc.), the processing performed by the pre-processing unit 14 does not necessarily need to be performed.

[0077] The feature quantity selection unit 15 includes a first feature quantity reduction unit 151, a feature quantity selection model generation unit 152, an importance calculation unit 153, a reducibility determination unit 154, a second feature quantity reduction unit 155, and a feature quantity group selection unit 156. In the process of generating the estimated model by the estimated model generation unit 16, the feature quantity selection unit 15 selects a reaction condition that satisfies a predetermined condition (for example, a reaction condition in the case where the accuracy as a feature quantity selection model is the highest) as a feature quantity group.

[0078] The first feature reduction unit 151 is a functional module that performs processing to reduce feature quantities using correlation coefficients in order to avoid a decrease in prediction accuracy associated with unknown input (chemical substance). Figure 4A As shown, the first feature quantity reduction unit 151 includes a correlation coefficient calculation unit 1511 , a correlation coefficient determination unit 1512 , and a feature quantity deletion unit 1513 .

[0079] As reference Fig.13A and Fig. 13B As described later, the correlation coefficient calculation unit 1511 calculates the correlation coefficient between the feature quantities (i.e., an index for measuring the strength of the linear relationship between two reaction conditions). The correlation coefficient determination unit 1512 determines whether the correlation coefficient calculated by the correlation coefficient calculation unit 1511 is equal to or greater than a predetermined threshold value (more specifically, determines that the feature quantity determined to have a correlation value equal to or greater than the predetermined threshold value is a reduction target). The feature quantity deletion unit 1513 deletes the feature quantity determined by the correlation coefficient determination unit 1512 as a reduction target.

[0080] The feature quantity selection model generation unit 152 is a functional module that performs processing to generate a model for selecting a feature quantity (feature quantity group). Figure 4B As shown, the feature quantity selection model generating unit 152 includes a first training data segmenting unit 1521, a first hyperparameter setting unit 1522, a first optimized model generating unit 1523, a first verification unit 1524 and a first expected improvement degree determining unit 1525.

[0081] The first training data segmentation unit 1521 segments the training data generated by the training data generation unit 13 (when the current processing unit 14 has performed pre-processing (pre-processing) on ​​the training data generated by the training data generation unit 13, it is the training data that has been pre-processed) at a predetermined ratio into training data for generating a feature quantity selection model with optimized hyperparameters and verification data for verifying the generated feature quantity selection model by the first verification unit 1524. The predetermined ratio here is set, for example, so that the training data for generating the feature quantity selection model is 90 and the verification data for verifying the generated feature quantity selection model is 10. The segmentation here is performed in units of data sets, which also applies to the subsequent segmentation into training data and verification data.

[0082] The first hyperparameter setting unit 1522 sets hyperparameters for generating a feature quantity selection model in the subsequent first optimized model generating unit 1523. The hyperparameters here are parameters specific to the model and are defined as parameters that need to be set in advance by the designer or model builder (or settable parameters).

[0083] The first hyperparameter setting unit 1522 includes a first optimization unit for setting hyperparameters, and for example, the first optimization unit is caused to scan the hyperparameters by using Bayesian optimization, etc., so as to set the hyperparameters. The first optimization unit outputs the expected improvement when setting the hyperparameters (outputs the expected improvement (EI, Expected Improvement) in Bayesian optimization, etc.). As a supplementary explanation, in the setting (scanning) of the hyperparameters, in addition to Bayesian optimization, random search, grid search, etc. can also be applied in this embodiment.

[0084] When the hyperparameters are set by the first hyperparameter setting unit 1522, the first optimized model generation unit 1523 uses machine learning methods such as decision trees, neural networks, gradient boosting, and support vectors to generate a model (feature quantity selection model) for selecting feature quantities (feature quantity groups).

[0085] The first verification unit 1524 verifies the feature quantity selection model generated by the first optimized model generation unit 1523 using the training data segmented by the first training data segmentation unit 1521 (i.e., verification data for verifying the feature quantity selection model), and generates accuracy information. The generated accuracy information is stored in the storage unit 17 together with the feature quantity (feature quantity group) through the control unit 11.

[0086] The first expected improvement determination unit 1525 determines whether the hyperparameter needs to be adjusted by the first hyperparameter setting unit 1522 (more accurately, the first optimization unit 15221) by comparing the expected improvement output by the first optimization unit 15221 with a predetermined threshold. That is, it is determined whether there is room for improving the accuracy of the feature quantity selection model to be generated. Then, when it is determined that there is room for improvement, the control unit 11 controls so that the processing from the hyperparameter setting process in the first hyperparameter setting unit 1522 to the verification process in the first verification unit 1524 is repeatedly performed until the first expected improvement determination unit 1525 determines that there is no room for improvement.

[0087] The importance calculation unit 153 calculates the importance of the feature quantity in the feature quantity selection model having accuracy information satisfying the predetermined condition among the accuracy information related to the feature quantity selection model stored by the control unit 11 (for example, a feature quantity selection model having the highest accuracy, a feature quantity selection model having an accuracy equal to or greater than a predetermined threshold, etc.). More specifically, what kind of feature quantity contributes to the accuracy of the feature quantity selection model is quantified. Here, the feature quantity is a feature in machine learning (i.e., a feature and / or property that quantitatively shows the characteristics and / or properties of an object), and here, chemical structure information and physical property information derived from information related to a plurality of chemical substances, chemical structure information and physical property information derived from a product, and reaction conditions are equivalent to the feature quantity.

[0088] The reducibility determination unit 154 determines whether a feature quantity is deletable based on the importance calculated by the importance calculation unit 153. Basically, the reducibility determination unit 154 determines a feature quantity with low importance among the feature quantities as a deletion object. Exceptionally, when a feature quantity with low importance corresponds to a control condition, as a feature quantity, in the case where a feature quantity with the next low importance does not correspond to the control condition and there are multiple feature quantities with the next low importance, the reducibility determination unit 154 determines the feature quantity with the next low importance as a deletion object instead of deleting the feature quantity with low importance (that is, based on the characteristics of the feature quantity with low importance, the feature quantity with low importance is not deleted). Such determination processing is repeatedly performed until it is determined that the feature quantity as a deletion object is no longer included. The processing here will refer to Figures 16A to 16B. Fig.16F The second feature amount reducing unit 155 reduces the feature amount by deleting the feature amount determined as a deletion target by the reduction determination unit 154 .

[0089] Each time a feature quantity selection model is generated, the feature quantity group selection unit 156 selects, in the form of a feature quantity group, a feature quantity for generating a feature quantity selection model having accuracy information satisfying a predetermined condition (e.g., a feature quantity selection model having the highest accuracy, a feature quantity selection model having an accuracy equal to or greater than a predetermined threshold, etc.). The selected feature quantity group is stored in the storage unit 17 through the control unit 11, and then transmitted to the estimation model generation unit 16.

[0090] The estimation model generation unit 16 includes a third training data division unit 161, a multi-estimation model generation unit 162, and an estimation model mixing unit 163, and generates a plurality of estimation models based on the feature quantity group selected by the feature quantity selection unit 15. The multi-estimation model generation unit 162 generates a plurality of estimation models according to each machine learning method (decision tree, neural network, gradient boosting, support vector).

[0091] The third training data splitting unit 161 randomly splits the training data generated by the training data generating unit 13 (when the pre-processing unit 14 performs pre-processing (pre-processing) on ​​the training data generated by the training data generating unit 13, the pre-processed training data) into first training data and first verification data. For example, the random splitting is performed at a ratio of 90:10 between the first training data and the first verification data.

[0092] like Figure 5A As shown, the multi-inference model generation unit 162 includes a second training data segmentation unit 1621, a second hyperparameter setting unit 1622, a second optimized model generation unit 1623, a second verification unit 1624, and a second expected improvement determination unit 1625, and has a configuration roughly similar to that of the feature quantity selection model generation unit 152. However, the difference between the multi-inference model generation unit 162 is that the multi-inference model generation unit 162 generates an inference model (regression model) according to a plurality of machine learning methods, while the feature quantity selection model generation unit 152 generates a model (feature quantity selection model) according to any one of the machine learning methods. The following is an explanation of each functional module.

[0093] The second training data dividing unit 1621 further divides the first training data divided by the third training data dividing unit 161 into second training data for generating an estimated model with optimized hyperparameters and second verification data for verifying the generated estimated model by the second verification unit 1624, according to the number of machine learnings used for modeling. As a method for dividing the first training data into the second training data and the second verification data, for example, a bootstrap method or the like can be applied.

[0094] The second hyperparameter setting unit 1622 is implemented according to the number of machine learning used for modeling (i.e., implemented with a second hyperparameter setting unit 1622-1, a second hyperparameter setting unit 1622-2, ... a second hyperparameter setting unit 1622-n, etc.), and each second hyperparameter setting unit 1622 sets the hyperparameters for generating an inferred model in the second optimized model generation unit 1623 corresponding to the machine learning in the subsequent stage.

[0095] Similar to the above description, the hyperparameters here are model-specific parameters, which are defined as parameters that need to be set in advance by the designer or model builder (settable parameters). Each second hyperparameter setting unit 1622 includes a second optimization unit for setting the hyperparameters, for example, by using Bayesian optimization, etc. to scan the hyperparameters by the second optimization unit, thereby setting the hyperparameters. The second optimization unit outputs the expected improvement when setting the hyperparameters (outputting the expected improvement (EI) in Bayesian optimization, etc.).

[0096] Similar to the second hyperparameter setting unit 1622, the second optimized model generation unit 1623 is implemented according to the number of machine learning used for modeling (i.e., the same number as the number of the second hyperparameter setting units 1622, and implemented as the second optimized model generation unit 1623-1, the second optimized model generation unit 1623-2, ... the second optimized model generation unit 1623-n, etc.). Each second optimized model generation unit 1623 is associated with the corresponding second hyperparameter setting unit 1622, and when the hyperparameters are set by the associated second hyperparameter setting unit 1622, an inference model is generated according to the machine learning method set for the second optimized model generation unit 1623 (as described above, for example, decision tree, neural network, gradient boosting, support vector, etc.).

[0097] Similar to the second hyperparameter setting unit 1622, the second verification unit 1624 is implemented according to the number of machine learning used for modeling (i.e., the same number as the number of the second optimized model generation unit 1623, and implemented as the second verification unit 1624-1, the second verification unit 1624-2, ... the second verification unit 1624-n, etc.). Each second verification unit 1624 is associated with the corresponding second optimized model generation unit 1623. When the optimized model (inferred model) is generated by the associated second optimized model generation unit 1623, the second verification unit 1624 uses the corresponding second verification data segmented by the second training data segmentation unit 1621 for verification, verifies the accuracy of the inferred model generated by the associated second optimized model generation unit 1623, and generates accuracy information. The generated accuracy information is stored in the storage unit 17 together with the inferred model through the control unit 11.

[0098] Similar to the second hyperparameter setting unit 1622, the second expected improvement determination unit 1625 is implemented according to the number of machine learning used for modeling (i.e., the same number as the number of the second verification units 1624, and implemented as the second expected improvement determination unit 1625-1, the second expected improvement determination unit 1625-2, ... the second expected improvement determination unit 1625-n, etc.). Each second expected improvement determination unit 1625 is associated with the corresponding second verification unit 1624 (second hyperparameter setting unit 1622), and by comparing the expected improvement output by the corresponding second optimization unit 16221 with a predetermined threshold, it is determined whether it is necessary to adjust the hyperparameters through the corresponding second hyperparameter setting unit 1622 (more accurately, the second optimization unit 16221). That is, it is determined whether there is room for improving the accuracy of the inferred model to be generated. Then, when it is determined that there is room for improvement, the control unit 11 controls so that the processing from the hyperparameter setting processing in the second hyperparameter setting unit 1622 to the verification processing in the second verification unit 1624 is repeatedly executed until the second expected improvement degree determination unit 1625 determines that there is no room for improvement.

[0099] like Figure 5B As shown, the estimated model mixing unit 163 includes a weight setting unit 1631, a third verification unit 1632, and a third expected improvement determination unit 1633. The estimated model mixing unit 163 performs mixing processing on the multiple estimated models generated by the multiple estimated model generation unit 162, and sets the optimal weight for each of the multiple estimated models.

[0100] The weight setting unit 1631 includes a fourth optimization unit 16311, and the fourth optimization unit 16311 scans the weights set for the multiple inference models generated in the multiple inference model generation unit 162, for example, by using Bayesian optimization, thereby adjusting (setting) the weights. The fourth optimization unit outputs an expected improvement (in Bayesian optimization, outputs expected improvement (EI) or the like) when setting the weights.

[0101] The third verification unit 1632 multiplies the weights set for each of the multiple estimation models generated in the multiple estimation model generation unit 162 by the predicted value corresponding to the estimation model, and further compares the result calculated by adding them with the first verification data. The weights set for each of the multiple estimation models and the comparison result (verification result) are stored in the storage unit 17 in association with the generated estimation model through the control unit 11.

[0102] The third expected improvement determination unit 1633 determines whether the weight setting unit 1631 (more accurately, the fourth optimization unit 16311) needs to adjust the weight by comparing the expected improvement output by the fourth optimization unit 16311 with a predetermined threshold. That is, for each of the multiple inference models, the weight is multiplied by the predicted value, and the accuracy of the result after mixing them is used to determine whether there is room for improvement. Then, when it is determined that there is room for improvement, the control unit 11 controls so that the weight setting process in the weight setting unit 1631 and the verification process in the third verification unit 1632 are repeatedly performed until the third expected improvement determination unit 1633 determines that there is no room for improvement.

[0103] As described above, the storage unit 17 stores predetermined programs, intermediate data, and execution results (data) of various processes. For example, the training data (data set) output in the process of reducing the feature amount is equivalent to the intermediate data, and for example, the estimated model and the weight associated with the estimated model are equivalent to the execution results of various processes.

[0104] Next, use Figure 6 The process of generating an estimation model in the model generating device 1 is described with reference to a flowchart. In the following description, the reference symbol "S" in the description of the flowchart refers to a step. That is, here, each processing step S1-1 to step S1-8 in the flowchart is abbreviated as S1-1 to S1-8. This is also the case in the flowchart described later.

[0105] In S1-1, the model generation device 1 obtains information related to multiple chemical substances (information related to chemical substance A described by SMILES tags, etc. and information related to chemical substance B described by SMILES tags, etc.), information related to products (information related to products described by SMILES tags, etc.), reaction conditions and yields through the data acquisition unit 12.

[0106] In S1-2, the model generation device 1 derives molecular descriptors based on information related to chemical substances through the molecular descriptor deriving unit 132. The processing here can refer to Figure 7 To describe in detail. Figure 71 is a flowchart illustrating the steps for deriving a molecular descriptor in the molecular descriptor deriving unit 132. In S1-2-1, the molecular descriptor deriving unit 132 calculates the chemical structure information as described above. In S1-2-2, the molecular descriptor deriving unit 132 obtains the physical property information from the physical property information database 133 storing information (physical property information) showing the physical and / or chemical properties of the chemical substance. The processing in S1-2-1 and the processing in S1-2-2 can be performed in parallel without regard to the order. Similar to the molecular descriptor deriving unit 132 deriving the molecular descriptor based on the information related to the chemical substance, the model generating device 1 derives the molecular descriptor based on the information related to the product.

[0107] In S1-3, the data set generation unit 131 adds the molecular descriptors derived from the information related to the chemical substance and the molecular descriptors derived from the information related to the product in S1-2, and the reaction conditions and yields obtained by the data acquisition unit 12 as elements constituting the data, thereby generating training data. Figure 8 The data set shown.

[0108] Figure 8 A. Figure 8 B and Figure 8 C is a diagram illustrating an example of a data set generation process, and more specifically, illustrates an example of a generation process procedure. Figure 8 A illustrates the first step of the data set generation stage, and illustrates a data set including yield, information related to chemical substances (chemical substance A and chemical substance B), information related to products generated when chemical substance A reacts with chemical substance B, and information related to reaction conditions (solvent, temperature, concentration, drop rate). The information related to chemical substances and the information related to products are described using SMILES notation, and a part of the reaction conditions are represented as control conditions.

[0109] Figure 8 B illustrates the second step of the data set generation stage, and illustrates the state in which, in the data set generated in the first step, information related to chemical substances (chemical substance A, chemical substance B, and a product generated when chemical substance A and chemical substance B react, and a solvent) is expressed as a molecular descriptor (converted into a molecular descriptor), and the molecular descriptor is appended to the data set as data. In addition, Figure 8 C describes the third step of the data set generation stage, and describes the following state, namely, in the data set generated in the second step, data obtained from the physical property information DB133 for information related to chemical substances (chemical substance A, chemical substance B, and products generated when chemical substance A reacts with chemical substance B, and solvent) is appended to the data set.

[0110] When there are missing values ​​in the training data (data set) generated by the training data generation unit 13 in S1-3, in S1-4, the model generation device 1 performs missing value processing (i.e., deletion processing of part of the data, or filling processing of missing values) through the missing value processing unit 142.

[0111] 9A to 9D is a diagram for explaining the missing value processing in the missing value processing unit 142. In particular, Fig.9A This describes how to delete rows containing missing values ​​in a data set. Fig. 9B This section describes how to delete columns containing missing values ​​in a data set. Fig. 9C This describes the process of filling in missing values ​​in a data set with the statistic (mean) of the column that contains the missing values. Fig.9D This section describes the process of filling in missing values ​​in a data set using statistics (modes) of columns that contain missing values.

[0112] In S1-5, when there are categories (reaction conditions) not shown as numerical values ​​in the training data (data set) generated in S1-3 or in the training data (data set) after missing value processing has been performed on the training data generated by the training data generating unit 13 in S1-4, the model generating device 1 converts the categories (reaction conditions) not shown as numerical values ​​into numerical values ​​through the category value conversion unit 141.

[0113] Fig.10 This is a diagram for explaining the processing in the category value conversion unit 141 as an example, and specifically, explains the processing of converting the category value into a numerical value using the Onehot encoding method. Fig.10 In , the categorical variables of temperature adjustment method and stirrer are each converted into a dummy variable to facilitate the learning of the machine learning algorithm. More specifically, Fig.10 As shown, for "temperature adjustment method", "temperature adjustment method_stage" is converted to "1,0,0", "temperature adjustment method_continuous" is converted to "0,1,0", and "temperature adjustment method_no adjustment" is converted to "0,0,1". For "mixer", "mixer_mixing rod" is converted to "1,0,0", "mixer_three-swept wing" is converted to "0,1,0", and "mixer_anchor" is converted to "0,0,1".

[0114] In S1 - 6 , the model generation device 1 normalizes the data for each category constituting the data set by the data normalization / normalization unit 143 , or normalizes the data to a real value between 0 and 1 using the maximum value / minimum value of the data constituting the category.

[0115] In S1-7, the model generation device 1 selects the reaction conditions of the feature quantity selection model having accuracy information satisfying the predetermined conditions (for example, a feature quantity selection model having the highest accuracy, a feature quantity selection model having an accuracy equal to or greater than a predetermined threshold, etc.) through the feature quantity selection unit 15 as a feature quantity group (feature quantity group) for generating an estimated model in S1-8. The processing here can refer to Fig.11 Let me explain in detail.

[0116] Fig.11 7-1, the model generation device 1 reduces (deletes) the feature quantity in the training data (data set) by the first feature quantity reduction unit 151. As a supplementary explanation, the processing here can be performed by Fig.12 The flowchart is described in more detail.

[0117] In S1-7-1-1, the model generation device 1 calculates the correlation coefficient by the correlation coefficient calculation unit 1511 of the first feature quantity reduction unit 151. Here, Fig.13A and Fig. 13B is a diagram illustrating the correlation coefficient and the process of reducing the feature quantity based on the correlation coefficient. Fig.13A and Fig. 13B In the matrix, the correlation coefficient between one feature quantity and another feature quantity is calculated, and each of the calculated correlation coefficients is shown at a corresponding position in the matrix.

[0118] In S1-7-1-2, the model generation device 1 determines one of the feature quantities represented by the predetermined correlation coefficient as a deletion object through the correlation coefficient determination unit 1512 of the first feature quantity reduction unit 151. Here, when the predetermined correlation coefficient is assumed to be 0.9, Fig.13A In the above example, since the correlation coefficient between chemical substance 1_FP1 and chemical substance 1_property 5 is calculated to be equal to or greater than the predetermined correlation coefficient (correlation value) (i.e., calculated to be 0.95, which is greater than 0.9), the correlation coefficient determination unit 1512 first determines chemical substance 1_FP1 and chemical substance 1_property 5 as candidates for deletion. It should be noted that, here, the FP (fingerprint) of chemical substance 1_FP1 refers to chemical structure information, which indicates whether the compound has a specific partial structure by "0" or "1".

[0119] Next, the correlation coefficient determination unit 1512 determines whether one of the chemical substance 1_FP1 and the chemical substance 1_property 5 is a deletion target based on each correlation coefficient with another characteristic quantity (in this case, the chemical substance 1_property 100) of the chemical substance 1_FP1 and the chemical substance 1_property 5. Specifically, since the correlation coefficient between the chemical substance 1_FP1 and the chemical substance 1_property 100 is 0.4, and the correlation coefficient between the chemical substance 1_property 5 and the chemical substance 1_property 100 is 0.6 (that is, the correlation coefficient (correlation value) between the chemical substance 1_property 5 and the chemical substance 1_property 100 is higher than the correlation coefficient (correlation value) between the chemical substance 1_FP1 and the chemical substance 1_property 100), the correlation coefficient determination unit 1512 determines the chemical substance 1_property 5 as a deletion target.

[0120] exist Fig. 13B In the case of , the judgment process is similarly performed, and the correlation coefficient judgment unit 1512 first judges chemical substance 2_FP1 and chemical substance 2_property 100 as candidates for deletion objects, and then judges chemical substance 2_property 100 as a deletion object based on the correlation coefficient with another feature value (in this case, chemical substance 2_property 5).

[0121] In S1-7-1-3, the model generation device 1 deletes the feature determined to be deleted in S1-7-1-2 by the feature deleting unit 1513 of the first feature reducing unit 151. By the above-described processing, feature quantities can be deleted from the training data (data set).

[0122] Refer again Fig.11 , when the feature quantity reduction process in S1-7-1 is executed, in S1-7-2, the model generation device 1 generates a feature quantity selection model through the feature quantity selection model generation unit 152 of the feature quantity selection unit 15. As a supplementary explanation, the process here can be performed by Fig.14 The flowchart is described in more detail.

[0123] In S1-7-2-1, the model generation device 1 divides the training data into training data for generating a feature quantity selection model with optimized hyperparameters and verification data for verifying the generated feature quantity selection model by the first verification unit 1524 at a predetermined ratio through the first training data division unit 1521 of the feature quantity selection model generation unit 152. Each of the divided training data is stored in the storage unit 17 through the control unit 11.

[0124] In S1-7-2-2, since the process of generating the feature quantity selection model has not been executed, the model generation device 1 determines n=1 (S1-7-2-2 No) through the control unit 11, and transfers the process to S1-7-2-3. In S1-7-2-3, under the premise that the process of generating the feature quantity selection model has not been executed, the model generation device 1 sets the initial value as the hyperparameter through the first hyperparameter setting unit 1522 of the feature quantity selection model generation unit 152.

[0125] In S1-7-2-4, the model generation device 1 generates a feature quantity selection model according to the set hyperparameter (initial value) through the first optimized model generation unit 1523 of the feature quantity selection model generation unit 152. In S1-7-2-5, the model generation device 1 verifies the accuracy of the feature quantity selection model generated in S1-7-2-4 through the first verification unit 1524 of the feature quantity selection model generation unit 152. The model generation device 1 stores the verification result (verification result) in the storage unit 17 through the control unit 11.

[0126] In S1-7-2-6, the model generation device 1 compares the expected improvement degree set as the initial value with a predetermined threshold value through the first expected improvement degree determination unit 1525 of the feature quantity selection model generation unit 152, thereby determining whether to return the process to S1-7-2-2. Here, the initial value is set in such a way that a necessary and sufficient number of models for selecting the feature quantity is generated (i.e., it is set in such a way that the process returns to S1-7-2-2).

[0127] When the process returns to S1-7-2-2, since the process of generating the feature quantity selection model has been executed, the model generation device 1 determines n≠1 (S1-7-2-2 is) through the control unit 11, and transfers the process to S1-7-2-7. In S1-7-2-7, the model generation device 1 performs the optimization of the hyperparameters through the first hyperparameter setting unit 1522 (first optimization unit 15221) of the feature quantity selection model generation unit 152. The first hyperparameter setting unit 1522 outputs the expected improvement in the process of performing the optimization.

[0128] In S1-7-2-8, the model generation device 1 stores the expected improvement output when executing the optimization process of the hyperparameters in S1-7-2-7 in the storage unit 17 through the control unit 11. Then, the process is transferred to S1-7-2-3, and the model generation device 1 sets the optimized hyperparameters in S1-7-2-7 as hyperparameters for generating the optimized model through the first hyperparameter setting unit 1522 of the feature quantity selection model generation unit 152.

[0129] In S1-7-2-4, the model generation device 1 generates an optimized model (i.e., a feature quantity selection model whose hyperparameters have been optimized) using the hyperparameters optimized in S1-7-2-7 and set in S1-7-2-3 through the first optimized model generation unit 1523 of the feature quantity selection model generation unit 152. In S1-7-2-5, the model generation device 1 verifies the accuracy of the optimized model generated in S1-7-2-4 through the first verification unit 1524 of the feature quantity selection model generation unit 152. The verification result is stored in the storage unit 17 through the control unit 11.

[0130] In S1-7-2-6, the model generation device 1 compares the expected improvement output and stored in S1-7-2-8 with a predetermined threshold value through the first expected improvement determination unit 1525 of the feature quantity selection model generation unit 152, thereby determining whether to return the process to S1-7-2-2. More specifically, through the first expected improvement determination unit 1525 of the feature quantity selection model generation unit 152, when the expected improvement output and stored in S1-7-2-8 is determined to be less than the predetermined threshold value, the model generation device 1 returns the process to S1-7-2-2, and when the expected improvement output and stored in S1-7-2-8 is determined to be equal to or greater than the predetermined threshold value, the model generation device 1 ends the process of generating the feature quantity selection model.

[0131] Therefore, in the process of S1-7-2, until the expected improvement output and stored in S1-7-2-8 becomes equal to or greater than a predetermined threshold, the processes of S1-7-2-2, S1-7-2-7, S1-7-2-8, S1-7-2-3, S1-7-2-4, S1-7-2-5, and S1-7-2-6 are repeatedly executed, and the optimized hyperparameters and verification results are stored for each execution. Then, when the expected improvement output and stored in S1-7-2-8 becomes equal to or greater than a predetermined threshold, Fig.14 The shown processing (ie, the processing of S1-7-2) ends.

[0132] As a supplementary explanation, the feature quantity selection model generated in S1-7-2-4 is described using an example in which a decision tree is used as a machine learning method. Fig.15 This is a diagram illustrating a model (decision tree model) when a decision tree is used as machine learning. Here, a decision tree is a model for estimating (classifying) data by branching according to predetermined conditions. Fig.15 Nodes to which feature values ​​are assigned by reference numerals N10 to N16 (i.e., conditions for setting feature values ​​for each node (i.e., branching conditions)) are shown, and branching is shown so that the node is classified as any of the lower nodes according to the branching conditions.

[0133] For example, the node shown in the reference numeral N10 is branched based on whether temperature_°C≤-0.386 is satisfied (i.e., whether the temperature based on the reaction conditions is -0.386 or lower). More specifically, in the case where temperature_°C≤-0.386 is not satisfied (i.e., when the temperature is greater than -0.386), the node of the reference numeral N10 is branched to the node of the reference numeral N11. In the case where temperature_°C≤-0.386 is satisfied (i.e., when the temperature is -0.386 or lower), the node of the reference numeral N10 is branched to the node of the reference numeral N12. The temperature here (i.e., -0.386) is represented by a normalized / standardized value.

[0134] In the branching from the node marked N10 to the nodes marked N11 and N12, the condition of the characteristic quantity is set in such a way that the sum of the mean square errors becomes the minimum, thereby performing branching (division). That is, branching is performed in such a way that the sum of the mean square errors in the node marked N11 and the node marked N12 becomes the minimum (here, the sum of the mean square errors recorded in the node marked N11 and the node marked N12 (1.587=0.832+0.755) is determined as the minimum sum of the mean square errors, and branching is performed with temperature_℃≤-1.856 and chemical substance B amount_mol%≤1.226 as characteristic quantity conditions).

[0135] In the attached figure, the sample (Sample) recorded in the node indicates the number of samples after segmentation, and the value (Value) indicates the average value, which is used to calculate the mean square error. In addition, as a supplementary explanation, in Fig.15 In the decision tree shown in FIG, the feature quantities of the nodes of the decision tree are represented by temperature, molecular descriptors, etc. However, since the mean square error is used for branching, other reaction conditions (e.g., concentration, dripping speed, etc.) can be selected as feature quantities as described above. Although the depth of the decision tree is Fig.15 2, but as shown in the figure “Continue”, the depth of the decision tree is actually deeper.

[0136] Refer again Fig.11 In the flowchart, when the generation processing of the feature quantity selection model in S1-7-2 is executed, in S1-7-3, the control unit 11 determines (selects) a feature quantity selection model associated with the accuracy information that satisfies a predetermined condition (for example, information related to the highest accuracy, information related to one of the accuracies equal to or greater than a predetermined threshold, etc.) among the accuracy information of the feature quantity selection model whose hyperparameters have been adjusted and saved.

[0137] In S1-7-4, the model generation device 1 calculates the importance in the feature quantity selection model stored as the feature quantity selection model whose accuracy information satisfies the predetermined condition through the importance calculation unit 153. As for the importance, as described above, it uses a numerical value to indicate which feature quantity contributes to the accuracy of the feature quantity selection model, and is calculated by adding the importance of each feature quantity in the constructed feature quantity selection model.

[0138] In S1-7-5, based on the number of feature quantities, whether the feature quantity corresponds to the control condition, the importance of each feature quantity calculated by the importance calculation unit 153, etc., the model generation device 1 determines whether the feature quantity can be reduced (deleted) through the reduction determination unit 154. When it is determined that the feature quantity can be reduced (deleted) (S1-7-5 is), the model generation device 1 transfers the processing to S1-7-6.

[0139] In S1-7-6, the model generation device 1 reduces (deletes) the feature quantity determined to be reducible in S1-7-5 by the second feature quantity reduction unit 155. After that, the process returns to S1-7-2, and the processes of S1-7-6, S1-7-2, S1-7-3, S1-7-4 and S1-7-5 are repeatedly performed until it is determined in S1-7-5 that the feature quantity cannot be deleted (S1-7-5 No).

[0140] Here, as a supplementary explanation, refer to FIGS. 16A to 16B. Fig.16F 16A shows a feature quantity selection model associated with the accuracy information satisfying the predetermined condition (e.g., information related to the highest accuracy, information related to one of the accuracies equal to or greater than a predetermined threshold, etc.) determined (selected) by the control unit 11 as a data set. In FIG. 16A, when the importance calculation unit 153 calculates that the importance of solvent 1_property 100 is the lowest importance, since solvent 1_property 100 does not belong to the control condition and solvent 1_property 1 exists as a feature quantity (in addition, since solvent 1_property 1 to solvent 1_property 99 exist as feature quantities), the reducibility determination unit 154 determines that the feature quantity is reducible.

[0141] Next, the solvent 1_property 100 calculated as having the lowest importance by the importance calculation unit 153 and determined as a reduction target by the reduction determination unit 154 is reduced (deleted) by the second feature reduction unit 155 ( FIG. 16B ). Thereafter, the process of generating a feature selection model, the process of calculating the importance of the feature selection model determined (selected) based on a predetermined condition, and the process of determining whether the feature can be deleted are performed again.

[0142] In the determination process performed by the reduction determination unit 154, it is determined that Fig. 16C Whether the feature quantities shown include deletable feature quantities. Here, Fig. 16C Among the feature quantities shown, although the feature quantity with the lowest importance is temperature (°C), temperature (°C) is not determined as a reduction target because temperature (°C) belongs to the control condition. A feature quantity (chemical substance B_FP1024) with the next lowest importance compared to temperature (°C) is determined, and further, under the condition that the feature quantity (chemical substance B_FP1024) does not belong to the control condition and there are multiple feature quantities (chemical substance B_FP1024) as feature quantities, chemical substance B_FP1024 is determined as a reduction target.

[0143] When the reducibility determination unit 154 determines that the chemical substance B_FP1024 can be reduced (determined as a reduction target), the chemical substance B_FP1024 determined as a reduction target by the reducibility determination unit 154 is reduced (deleted) by the second feature quantity reduction unit 155. Thereafter, the process of generating a feature quantity selection model, the process of calculating the importance of the feature quantity selection model determined (selected) based on a predetermined condition, and the process of determining whether the feature quantity can be deleted are repeatedly performed until the feature quantity determined as a deletion target is no longer included ( FIG. 16C to FIG. 16F )until.

[0144] Refer again Fig.11 When it is determined in S1-7-5 that the feature quantity cannot be deleted (S1-7-5 No), in S1-7-7, the model generation device 1 selects the group of feature quantities that are not deleted in S1-7-6 but retained through the feature quantity group selection unit 156. When the group of retained feature quantities is selected in S1-7-7, Fig.11 A series of processing shown in the flowchart is completed (that is, the processing shown in S1-7 is completed).

[0145] Refer again Figure 6 In S1-8, the model generation device 1 generates a plurality of estimated models through the estimated model generation unit 16. The processing here can refer to Fig.17 Let me explain in detail. Fig.17 1 is a flowchart showing the steps of a process for generating a plurality of estimation models by the estimation model generating unit 16 .

[0146] like Fig.17As shown in the flowchart, in S1-8-1, the model generating device 1 uses the third training data segmentation unit 161 to segment the training data generated by the training data generating unit 13 (when the current processing unit 14 has performed pre-processing (pre-processing) on ​​the training data generated by the training data generating unit 13, it is the training data on which the pre-processing has been performed) into the first training data used in S1-8-2 and the first verification data used in S1-8-3.

[0147] In S1-8-2, a plurality of estimation models are generated by the multi-estimation model generation unit 162 of the estimation model generation unit 16, and then, in S1-8-3, the plurality of estimation models are mixed by the estimation model mixing unit 163 of the estimation model generation unit 16 to calculate the weights set for each of the plurality of estimation models. Fig.17 The processing shown in the flowchart (the processing shown in S1-8-2 and S1-8-3) refers to Fig.18 and Fig. 20 Let's explain in order.

[0148] Fig.18 It is a flowchart illustrating the steps for the processing shown in S1-8-2 (i.e., the processing for generating multiple inference models). In S1-8-2-1, the model generation device 1 divides the first training data divided in S1-8-1 into second training data for generating an inference model with optimized hyperparameters and second verification data for verifying the generated inference model in the second verification unit 1624 by a division method such as a bootstrap method through the second training data division unit 1621 of the multiple inference model generation unit 162 (more specifically, here, the first training data is divided into three sets, wherein one set has second training data for generating an inference model with optimized hyperparameters and second verification data for verifying the generated inference model).

[0149] In S1-8-2-2, the model generation device 1 stores the second training data for generating an inferred model with optimized hyperparameters and the second verification data for verifying the generated inferred model, which have been divided into three sets, in the storage unit 17 in units of feature quantities through the control unit 11. In S1-8-2-3, the model generation device 1 assigns the second training data and the second verification data divided into one set to a pre-set machine learning method through the control unit 11. As described above, since the first training data is divided into three sets by the bootstrap method, the control unit 11 assigns the three sets to machine learning methods 1 to 3 (i.e., machine learning method 1, machine learning method 2, and machine learning method 3), respectively.

[0150] Each of the machine learning methods 1 to 3 is set to any one of the machine learning methods such as decision tree, neural network, gradient boosting and support vector. In S1-8-2-4, the second training data allocated for generating the estimated model is used as input data, and the estimated model is generated using the set machine learning method. Still in S1-8-2-5 and S1-8-2-6, similar to S1-8-2-4, the second training data allocated for generating the estimated model is used as input data, and the estimated model is generated using the set machine learning method.

[0151] The same machine learning method can be used for the processing in S1-8-2-4, the processing in S1-8-2-5, and the processing in S1-8-2-6. The order of the processing does not matter, and the processing can be performed in parallel. In addition, these processing can refer to Fig.19 To explain in more detail, here, the processing in S1-8-2-4 is used as an example.

[0152] In S1-8-2-4-1, since the process of generating an inference model corresponding to the set machine learning method has not been executed, the model generation device 1 determines that n=1 (S1-8-2-4-1 No), and transfers the process to S1-8-2-4-2 through the control unit 11. In S1-8-2-4-2, on the premise that the process of generating an inference model corresponding to the set machine learning method has not been executed, the model generation device 1 sets an initial value as a hyperparameter through the second hyperparameter setting unit 1622 of the multi-inference model generation unit 162.

[0153] In S1-8-2-4-3, the model generation device 1 generates an estimated model according to the set hyperparameter (initial value) through the second optimized model generation unit 1623 of the multi-estimation model generation unit 162. In S1-8-2-4-4, the model generation device 1 verifies the accuracy of the estimated model generated in S1-8-2-4-3 through the second verification unit 1624 of the multi-estimation model generation unit 162. The verified result (verification result) is stored in the storage unit 17 through the control unit 11.

[0154] In S1-8-2-4-5, the model generation device 1 compares the expected improvement degree set as the initial value with a predetermined threshold value through the second expected improvement degree determination unit 1625 of the multi-estimation model generation unit 162, thereby determining whether to return the process to S1-8-2-4-1. Here, the initial value is set so that a necessary and sufficient number of models are generated so that the accuracy of the estimated model is equal to or greater than the predetermined accuracy (that is, it is set so that the process returns to S1-8-2-4-1).

[0155] When the processing returns to S1-8-2-4-1, since the processing of generating the inference model corresponding to the set machine learning method has been executed, the model generation device 1 determines that n≠1 (S1-8-2-4-1 is), and transfers the processing to S1-8-2-4-6 through the control unit 11. In S1-8-2-4-6, the model generation device 1 performs hyperparameter optimization through the second hyperparameter setting unit 1622 (second optimization unit 16221) of the multi-inference model generation unit 162.

[0156] In S1-8-2-4-7, the model generation device 1 stores the expected improvement output in the process of performing the optimization process of the hyperparameters in S1-8-2-4-6 in the storage unit 17 through the control unit 11. Thereafter, the process is transferred to S1-8-2-4-2, and the model generation device 1 sets the hyperparameters optimized in S1-8-2-4-6 as the hyperparameters for generating the inference model through the second hyperparameter setting unit 1622 of the multi-inference model generation unit 162.

[0157] In S1-8-2-4-3, the model generation device 1 generates an optimized model (i.e., an inferred model whose hyperparameters have been optimized) using the hyperparameters optimized in S1-8-2-4-6 and set in S1-8-2-4-2 through the second optimized model generation unit 1623 of the multi-inferred model generation unit 162. In S1-8-2-4-4, the model generation device 1 verifies the accuracy of the optimized model generated in S1-8-2-4-3 through the second verification unit 1624 of the multi-inferred model generation unit 162. The verification result is stored in the storage unit 17 through the control unit 11.

[0158] In S1-8-2-4-5, the model generation device 1 compares the expected improvement output and stored in S1-8-2-4-7 with a predetermined threshold value through the second expected improvement determination unit 1625 of the multi-estimation model generation unit 162, thereby determining whether to return the process to S1-8-2-4-1. More specifically, through the second expected improvement determination unit 1625 of the multi-estimation model generation unit 162, when the expected improvement output and stored in S1-8-2-4-7 is determined to be less than the predetermined threshold value, the model generation device 1 returns the process to S1-8-2-4-1, and when the expected improvement output and stored in S1-8-2-4-7 is determined to be equal to or greater than the predetermined threshold value, the model generation device 1 ends the process of generating the inference model.

[0159] Therefore, in the process of S1-8-2-4, until the expected improvement output and stored in S1-8-2-4-7 becomes equal to or greater than a predetermined threshold, the processes of S1-8-2-4-1, S1-8-2-4-6, S1-8-2-4-7, S1-8-2-4-2, S1-8-2-4-3, S1-8-2-4-4 and S1-8-2-4-5 are repeatedly executed. Then, the optimized hyperparameters and verification results are stored for each execution. Refer again Fig.18 , when the processing in S1-8-2-5 and the processing in S1-8-2-6 are performed similarly to the processing in S1-8-2-4, Fig.18 The process of generating a plurality of estimation models shown is completed (ie, the process shown in S1-8-2 is completed).

[0160] Next, refer to Fig. 20 The flowchart shown describes the processing shown in S1-8-3 (i.e., the processing of mixing multiple inference models). In S1-8-3-1, since the weight setting processing has not been performed, the model generation device 1 determines that n=1 (S1-8-3-1 No), and transfers the processing to S1-8-3-2 through the control unit 11. In S1-8-3-2, under the premise that the weight setting processing has not been performed, the model generation device 1 sets the initial value through the weight setting unit 1631 of the inference model mixing unit 163 as the weight added to each of the multiple inference models generated by the multi-inference model generation unit 162.

[0161] In S1-8-3-3, the model generating device 1 multiplies the weights (initial values) set for each of the multiple estimation models generated by the multi-estimation model generating unit 162 by the predicted value of the corresponding estimation model through the third verification unit 1632 of the estimation model mixing unit 163, and further adds them together.

[0162] In S1-8-3-4, the model generation device 1 compares the execution result (calculation result) of S1-8-3-3 with the first verification data through the third verification unit 1632 of the estimated model mixing unit 163. Then, the model generation device 1 stores the comparison result (verification result) and the weights set for each of the plurality of estimated models in association with the generated estimated models in the storage unit 17 through the control unit 11.

[0163] In S1-8-3-5, the model generation device 1 compares the expected improvement degree set as the initial value with a predetermined threshold value through the third expected improvement degree determination unit 1633 of the estimated model mixing unit 163, thereby determining whether to return the process to S1-8-3-1. Here, the initial value is set in such a way that the weights set for each of the plurality of estimated models are added as the optimal value (that is, in order to improve the prediction accuracy, the process is returned to S1-8-3-1).

[0164] When the processing returns to S1-8-3-1, since the weight setting processing has been performed, the model generation device 1 determines that n≠1 (S1-8-3-1 is), and transfers the processing to S1-8-3-6 through the control unit 11. In S1-8-3-6, the model generation device 1 performs optimization of the weights added to each of the multiple inference models through the weight setting unit 1631 (fourth optimization unit 16311) of the inference model mixing unit 163.

[0165] In S1-8-3-7, the model generation device 1 stores the expected improvement output during the weight optimization process in S1-8-3-6 in the storage unit 17 through the control unit 11. Thereafter, the process is transferred to S1-8-3-2, and the model generation device 1 sets the weight optimized in S1-8-3-6 as the additional weight for each of the multiple estimation models generated in the multi-estimation model generation unit 162 through the weight setting unit 1631 of the estimation model mixing unit 163.

[0166] In S1-8-3-3, the model generation device 1 multiplies the weight optimized in S1-8-3-6 and set in S1-8-3-2 by the prediction value of the corresponding estimation model through the third verification unit 1632 of the estimation model mixing unit 163, and further adds them.

[0167] In S1-8-3-4, the model generation device 1 compares the execution result (calculation result) of S1-8-3-3 with the first verification data through the third verification unit 1632 of the estimated model mixing unit 163. Then, the model generation device 1 stores the comparison result (verification result) and the weights set for each of the plurality of estimated models in association with the generated estimated models in the storage unit 17 through the control unit 11.

[0168] In S1-8-3-5, the model generation device 1 compares the expected improvement output and stored in S1-8-3-7 with a predetermined threshold value through the third expected improvement determination unit 1633 of the estimated model mixing unit 163, thereby determining whether to return the process to S1-8-3-1. More specifically, when the expected improvement output and stored in S1-8-3-7 is determined to be less than the predetermined threshold value through the third expected improvement determination unit 1633 of the estimated model mixing unit 163, the model generation device 1 returns the process to S1-8-3-1.

[0169] Therefore, in the process of S1-8-3-5, until the expected improvement degree output and stored in S1-8-3-7 becomes equal to or greater than the predetermined threshold value, the processes of S1-8-3-1, S1-8-3-6, S1-8-3-7, S1-8-3-2, S1-8-3-3, S1-8-3-4 and S1-8-3-5 are repeatedly executed. Then, for each execution, the optimized weight and verification result are associated with the generated estimation model and stored in the storage unit 17. In the process of S1-8-3-5, when the expected improvement degree output and stored in S1-8-3-7 is determined to be equal to or greater than the predetermined threshold value, the process of mixing multiple estimation models ends (that is, the process shown in S1-8-3 ends).

[0170] As a supplementary explanation, see Fig.21 , while paying attention to the circulation of data (data flow), the content of the processing in the inferred model generation unit 16 is explained. In S2-1, the training data generated by the training data generation unit 13 (when the current processing unit 14 has performed pre-processing (pre-processing processing) on ​​the training data generated by the training data generation unit 13, it is the pre-processed training data) is obtained by the inferred model generation unit 16.

[0171] In S2-2, the obtained training data is divided into first training data and first verification data. In S2-3, the divided first training data is further divided into three sets by the bootstrapping method, wherein one set includes second training data and second verification data. In the division here, data duplication is allowed. That is, for the second training data divided into the first set and the second training data divided into the second set, data can be duplicated.

[0172] In S2-4, based on the first set (the second training data and the second verification data as the first set), the estimated model 1 is generated by the machine learning method 1. Then, when the estimated model 1 is generated, in S2-7, the predicted value y1 of the estimated model 1 is calculated. Similarly, in S2-5, the estimated model 2 is generated by the machine learning method 2, and the predicted value y2 of the estimated model 2 is calculated in S2-8. In S2-6, the estimated model 3 is generated by the machine learning method 3. In S2-9, the predicted value y3 of the estimated model 3 is calculated.

[0173] In S2-10, for each of the plurality of estimation models, a weight is set for the estimation model (specifically, a weight w1 is set for estimation model 1, a weight w2 is set for estimation model 2, and a weight w3 is set for estimation model 3). In S2-11, the prediction values ​​of the estimation models are multiplied by the weights corresponding to the estimation models, and their total is calculated (specifically, the total of the product of the prediction value 1 (y1) of estimation model 1 and the weight 1 (w1) corresponding to estimation model 1, the product of the prediction value 2 (y2) of estimation model 2 and the weight 2 (w2) corresponding to estimation model 2, and the product of the prediction value 3 (y3) of estimation model 3 and the weight 3 (w3) corresponding to estimation model 3 is calculated).

[0174] In S2-12, the total (result) calculated in S2-11 is compared with the first verification data segmented in S2-2 to verify the accuracy. In S2-13, the comparison result (verification result) is stored in the storage unit 17 in association with the generated estimation model together with the weights set for each of the multiple estimation models. As described in the flowchart showing the steps for the above-mentioned processing in the estimation model generation unit 16, the data processing from S2-10 to S2-13 here is repeatedly performed until the expected improvement output during the execution of the Bayesian optimization becomes equal to or greater than a predetermined threshold. This is Fig.21 Shown by dotted lines.

[0175] When an estimated model is provided, provided are: estimated model 1 generated in S2-4, estimated model 2 generated in S2-5, and estimated model 3 generated in S2-6; and weight 1, weight 2, and weight 3 corresponding to the above-mentioned estimated models, respectively, which are stored in S2-13 and verified as those with the highest accuracy in the verification results corresponding to the estimated models.

[0176] thus, Figure 6 The processing shown in S1-8 in the flowchart is completed. Figure 6The flowchart of describes the steps of the processing for generating an estimated model in the model generating device 1, but the steps (processing) for generating an estimated model in the model generating device 1 can be further described as a whole. Specifically, the processing of deriving molecular descriptors in S1-2 and the processing of generating a data set in S1-3 can be performed together as a processing for generating training data, and the processing for missing values ​​in S1-4, the processing for category values ​​in S1-5, and the processing for standardization / normalization of data in S1-6 can be performed together as pre-processing.

[0177] In addition, the processing described as pre-processing (i.e., the processing shown by S1-4, S1-5 and S1-6) is processing performed corresponding to the generated training data (for example, when the training data includes missing values, missing value processing (as described above, for example, deletion processing of a part of the data or filling processing of missing values)) is not processing that is necessarily performed.

[0178] exist Figure 6 In the flowchart (this embodiment), in order to improve the accuracy of model generation, the following are performed: (1) a process for deleting feature quantities (S1-7-4, S1-7-6, etc.); (2) a process for adjusting hyperparameters (S1-7-2-7, etc.); and (3) a process for generating multiple estimation models and setting weights (S1-8). Specifically, the process of adjusting hyperparameters and the process of deleting feature quantities from the estimation model generated with the adjusted hyperparameters are repeatedly performed until a predetermined condition is satisfied, and then, the feature quantities (feature quantity group) for generating an estimation model with high accuracy obtained by performing the process for deleting feature quantities are used to generate multiple estimation models and set weights.

[0179] Therefore, in S1-7-2-4, on the premise that appropriate accuracy can be ensured for the feature quantity selection model generated based on the hyperparameters (initial values), the feature quantity selection model generated by the hyperparameters (initial values) can be provided as an estimated model generated by the model generation device 1.

[0180] In addition, it is also possible to separately and independently perform (1) processing for deleting feature quantities and (2) processing for adjusting hyperparameters, and provide a feature quantity selection model with accuracy information satisfying predetermined conditions among the feature quantity selection models generated by the execution (for example, a feature quantity selection model with the highest accuracy, a feature quantity selection model among feature quantity selection models with an accuracy equal to or greater than a predetermined threshold, etc.) as an estimated model generated by the model generation device 1.

[0181] Similarly, (3) the process for generating a plurality of estimation models and setting weights can be performed separately and independently from (1) the process for deleting the feature quantity and (2) the process for adjusting the hyperparameter. That is, it is possible to generate a plurality of estimation models based on the feature quantity of the feature quantity selection model generated with the hyperparameter (initial value) and to set the weights of each of the generated plurality of estimation models.

[0182] In addition, as a supplementary explanation, although the above embodiment uses information related to two chemical substances, that is, information related to chemical substance A and information related to chemical substance B as information related to a plurality of chemical substances, it is also possible to generate a model using information related to three or more chemical substances. When using information related to three or more chemical substances, for example, in addition to the information on chemical substance A and chemical substance B, chemical structure information and physical property information of chemical substances such as chemical substance C, chemical substance D, etc. are also added as elements constituting training data generated in the form of a predetermined data set, and it is only necessary to perform a process of selecting feature quantities (a process of deleting feature quantities) and a process of generating a plurality of estimation models based on the data set (training data) thus constituted.

[0183] Although the decision tree is used as an example to illustrate machine learning in the embodiment, a neural network can also be used as machine learning as described above. The neural network includes an input layer, an output layer, and a plurality of intermediate layers located between the input layer and the output layer. Each of the intermediate layers includes a plurality of nodes (neurons), and biases are set for each of the plurality of nodes. Each of the plurality of nodes is connected to the nodes of the intermediate layer as the front stage and the rear stage through an edge with a predetermined weight.

[0184] Based on this configuration, each of the multiple nodes in each intermediate layer constitutes an intermediate layer of the previous stage, and performs a product-sum operation of the input signal from the node connected to the node by the edge, the weight set to the edge, and the deviation. In addition, the application of the activation function constitutes the intermediate layer of the latter stage, and generates an output signal to be output to the node connected to the node by the edge. The generated output signal is transmitted to the corresponding node of the output layer via the edge. As a supplementary explanation, the weight set to the edge is used for multiplication in the product-sum operation, and the deviation set to the node is used for addition in the product-sum operation.

[0185] In the case of using the above-mentioned neural network, specifically, when training data generated from information related to a plurality of chemical substances, reaction conditions when reacting a plurality of chemical substances, information related to products generated by reacting a plurality of chemical substances, and yields of products when reacting under the reaction conditions, which are obtained as input data in the model generation device 1, are input to the input layer of the neural network, in each layer of the intermediate layer, product-sum operations and application of activation functions are performed as described above, and output signals are generated. The generated output signals are transmitted to corresponding nodes of the output layer via edges.

[0186] The model generation device 1 calculates the loss between the output signal (yield) obtained from the output layer and the target signal (training data), changes the bias of the nodes and the weight of the edges in the neural network based on the calculation result, calculates the gradient value of the loss, further calculates the product of the gradient value of the loss and the learning rate and subtracts the value obtained by the bias of the nodes and the weight of the edges in the neural network. Then, the model generation device 1 repeatedly performs the calculation processing to optimize the bias of the nodes and the weight of the edges in the neural network (that is, the model generation device 1 generates an optimized neural network (estimation model)).

[0187] As described above, according to this embodiment, by generating an estimation model that takes information related to multiple chemical substances that are reaction objects and information related to the products to be generated as input values ​​and takes reaction conditions in which the yield of the products to be generated satisfies predetermined conditions as output values, it is possible to reduce operating hours in setting reaction conditions in which the yield satisfies the predetermined conditions.

[0188] (Providing reaction conditions (reaction condition providing device))

[0189] Fig. 22 1 is a block diagram illustrating the functional configuration of the reaction condition providing device 5 according to the embodiment. Fig. 22 As shown, the reaction condition providing device 5 mainly includes a control unit 51, a data acquisition unit 52, a chemical property information generating unit 53, a reaction condition setting unit 54, a missing value processing unit 55, an inference model 56, an analysis unit 57, a display control unit 58 and a storage unit 59.

[0190] The control unit 51 is a functional module that controls the processing of each functional module. The data acquisition unit 52 acquires information related to a plurality of chemical substances and information related to a product (product to be generated) generated by reacting a plurality of chemical substances. The information related to a plurality of chemical substances and information related to the product to be generated acquired by the data acquisition unit 52 are temporarily stored in the storage unit 59 through the control unit 51. It should be noted that the information related to the chemical substances and the information related to the product can be described using tags such as SMILES tags, MOL files, and SDF files. In this embodiment, similar to the case of the model generation device, information described using SMILES tags is acquired. In addition, here, as information related to a plurality of chemical substances, information related to two chemical substances, namely, information related to chemical substance A and information related to chemical substance B are used for explanation.

[0191] The chemical property information generating unit 53 includes a data set generating unit 531, a molecular descriptor deriving unit 532, and a physical property information database (physical property information DB in the figure) 533, and generates chemical property information from the information related to the chemical substance obtained by the data acquiring unit 52. The data set generating unit 531 generates chemical property information in the form of a predetermined data set based on the chemical structure information derived by the molecular descriptor deriving unit 532 and the physical property information obtained from the physical property information database 533. The processing performed on the information related to the chemical substance here is performed similarly to the processing performed on the information related to the product obtained by the data acquiring unit 52. As described above, Figure 1B As shown, the physical property information stored in the physical property information database 533 can be updated from an external server device via a wired or wireless communication network. This also applies to the reaction condition database 541 described later.

[0192] The molecular descriptor deriving unit 532 derives the molecular descriptor from the chemical structure information and physical property information shown as an index (numerical value) when determining the molecular structure. In the derivation of the molecular descriptor, similar to the case of the model generation device, although the Morgan method is used for explanation, the Mordred method, the RDkit descriptor method, etc. can also be used. That is, here, similar to the case of the model generation device, the chemical structure information is calculated, and the physical property information is obtained from a predetermined database, thereby deriving the molecular descriptor.

[0193] The molecular descriptors (i.e., the calculated chemical structure information and the acquired physical property information) derived by the molecular descriptor deriving unit 532 are added as data (elements) to the data set by the data set generating unit 531. Therefore, the data set generating unit 531 adds the molecular descriptors as elements of the data set to the information related to the chemical substance, and further adds the molecular descriptors as elements of the data set to the information related to the product, thereby generating chemical property information in the form of a predetermined data set.

[0194] The reaction condition setting unit 54 includes a reaction condition database (reaction condition DB in the figure) 541, a similarity calculation unit 542, a first reaction condition selection unit 543, and a reaction condition range setting unit 544. The reaction condition setting unit 54 sets the reaction conditions to be input to the analysis unit 57 (more specifically, the reaction simulation unit 571 of the analysis unit 57) based on the chemical property information generated by the chemical property information generation unit 53.

[0195] The reaction condition database 541 is a database that stores data related to experiments conducted in the past. Specifically, a plurality of reacted chemical substances, reaction conditions when reacting the plurality of chemical substances, and products generated by the reactions are included in one data set, and the data set is stored in a necessary and sufficient amount for estimating the reaction conditions. The plurality of chemical substances and products stored in the reaction condition database 541 are described using SMILES tags.

[0196] The similarity calculation unit 542 uses a predetermined index to calculate the similarity between the information on the plurality of chemical substances acquired by the data acquisition unit 52 and the information on the plurality of chemical substances in past reactions stored in the reaction condition database 541. Similarly, for the product, the similarity calculation unit 542 uses a predetermined index to calculate the similarity between the information on the product acquired by the data acquisition unit 52 and the information on the product generated by the past reaction.

[0197] Thus, by calculating the similarity and inputting the reaction conditions selected by evaluating the calculated similarity into the estimation model, it is possible to provide (estimate) reaction conditions with a high yield. In the present embodiment, as described above, in the process of determining a chemical substance similar to the input information related to a plurality of chemical substances (or a chemical substance identical thereto), although the similarity is used, the distances between the chemical substances (distance matrix, similarity matrix) can also be used.

[0198] Indicators (methods) for calculating similarity include contingency coefficient, Tanimoto coefficient, Dice coefficient, etc. Here, as an example, the calculation of similarity using Tanimoto coefficient is briefly described. In Tanimoto coefficient, for each compound as the object of similarity calculation, the molecular structure described by SMILES tag is converted into a fingerprint, and the similarity between fingerprints is calculated.

[0199] More specifically, the Tanimoto coefficient (here, the Tanimoto coefficient between chemical substance x and chemical substance y) is shown as a value obtained by dividing the number of partial structures common to chemical substance x and chemical substance y by the total number of partial structures included in chemical substance x and chemical substance y (i.e., as shown in the following formula). In the formula, n(x∩y) represents the number of common partial structures, and n(x∪y) represents the total number of partial structures (natural number), and the closer the divided value is to "1", the higher the similarity between chemical substance x and chemical substance y becomes.

[0200] [Mathematical formula 1]

[0201]

[0202] Here, when chemical substance x and chemical substance y each have a fingerprint (FP) as shown in Table 1 below, n(x∩y) is "4" (FP2, FP3, FP5, FP10), and n(x∪y) is "7" (FP1, FP2, FP3, FP5, FP6, FP8, FP10), and therefore, the Tanimoto coefficient is 4 / 7 (=0.57). For reference, Table 2 below shows the fingerprints of chemical substance x and chemical substance z, in which case the Tanimoto coefficient is 5 / 7 (=0.71). Therefore, when chemical substance x is an unknown chemical substance, chemical substance z is determined to be a chemical substance that is more similar to chemical substance x than chemical substance y.

[0203] [Table 1]

[0204]

[0205] [Table 2]

[0206]

[0207] The first reaction condition selection unit 543 selects a reaction condition by comparing the similarity calculated by the similarity calculation unit 542 with a predetermined threshold value. The reaction condition range setting unit 544 sets a combination of assumed reaction conditions according to the selected reaction condition. The processing contents in the similarity calculation unit 542, the first reaction condition selection unit 543, and the reaction condition range setting unit 544 will be referred to in detail. FIG. 24A to FIG. 24E A specific description will be given later.

[0208] When the reaction conditions set by the reaction condition range setting unit 544 cannot meet the predetermined requirements as a data set (that is, when it is expected that appropriate analysis cannot be performed by the analysis unit 57 and the reaction conditions cannot be output appropriately), the missing value processing unit 55 performs predetermined missing value processing on the reaction conditions (data set).

[0209] In the missing value processing here, similar to the case of the model generation device, a process of deleting a part of the data or filling (patching) the missing values ​​is performed. More specifically, (1) the rows or columns including missing values ​​in the data set are deleted, and (2) the rows or columns including missing values ​​are filled using statistics (mean, median, mode, etc.).

[0210] The inferred model 56 is an inferred model learned by performing machine learning in advance using a machine learning method such as a decision tree, a neural network, a gradient boosting, and a support vector. Here, in the present embodiment, it is envisaged that a device (model generation device 1) different from the reaction condition providing device 5 is used to perform the processing related to learning and relearning in the inferred model 56. In this case, for example, with respect to the inferred model 56, it can be provided to the reaction condition providing device 5 via a predetermined recording medium, or, for example, it can also be transmitted to the reaction condition providing device 5 via a wired or wireless communication network through a predetermined server device. The inferred model 56 can be provided to the reaction condition providing device 5 in any manner. Therefore, although an example in which the reaction condition providing device and the model generating device are configured as different devices is shown here, the present invention is not necessarily limited to this, and a configuration having the functions of the reaction condition providing device and the functions of the model generating device in one device can be adopted.

[0211] The analysis unit 57 includes a reaction simulation unit 571, a classification unit 572, a determination unit 573, a second reaction condition selection unit 574, and a third optimization unit 575. The analysis unit 57 performs simulation according to the reaction conditions set by the reaction condition setting unit 54, selects a reaction condition whose yield meets a predetermined condition based on the execution result (analysis result), and sets the selected reaction condition as an output (display) object. When there is no reaction condition whose yield meets the predetermined condition, the reaction condition optimized by the third optimization unit 575 is set as an output object.

[0212] The reaction simulation unit 571 inputs the reaction conditions set by the reaction condition range setting unit 544 (when the missing value processing is performed, the reaction conditions set by the reaction condition range setting unit 544 are subjected to the missing value processing) into the estimation model 56, thereby performing the reaction simulation. When the execution of the reaction simulation is completed, the control unit 51 stores the execution result (yield) of the reaction simulation in the storage unit 59 in association with the reaction conditions.

[0213] The classification unit 572 assigns a grade (classifies) each of the plurality of reaction conditions associated with the yield and stored in the storage unit 59 according to a predetermined condition. For example, in the case of a yield of 80% or more, it is classified as "A", in the case of a yield of less than 80% and 50% or more, it is classified as "B", and in the case of a yield of less than 50%, it is classified as "C". The determination unit 573 determines whether a reaction condition whose yield satisfies a predetermined condition exists. Specifically, based on the grade assigned by the classification unit 572, it is determined whether a reaction condition of a predetermined grade or higher exists (for example, it is determined whether a reaction condition of grade "B" or higher exists).

[0214] The second reaction condition selection unit 574 selects the reaction condition determined by the determination unit 573 to satisfy the predetermined condition (determined to have a predetermined level or higher level) as the output object, and further assigns a display priority to the reaction condition selected as the output object. When the determination unit 573 determines that there is no reaction condition that satisfies the predetermined condition (when the reaction condition determined to have a predetermined level or higher level does not exist), the third optimization unit 575 performs optimization processing in the process of exploring (resetting) the reaction condition. It should be noted that although the processing here will be described by referring to Figures 27A to 27B described later, Fig.29 Although the description is made with reference to the accompanying drawings, regarding the setting (scanning) of the reaction conditions, in addition to the Bayesian optimization applied in the present embodiment, random search, grid search, etc. may also be applied.

[0215] When it is determined by the determination unit 573 that there is a reaction condition satisfying the predetermined condition (when it is determined that a reaction condition of a predetermined level or higher exists), the display control unit 58 controls the external display device 7 according to the display priority to display the reaction condition selected as an output object by the second reaction condition selection unit 574. When it is determined by the determination unit 573 that there is no reaction condition satisfying the predetermined condition, the display control unit 58 controls to display the result (reaction condition) of executing the optimization process in the third optimization unit 575.

[0216] Next, refer to Fig.23The flowchart of FIG. 3 is used to explain the steps of the process of estimating (providing) the reaction conditions by the reaction condition providing device 5. In S3-1, the reaction condition providing device 5 obtains information related to a plurality of chemical substances (information related to chemical substance A described by SMILES tags and information related to chemical substance B described by SMILES tags) and information related to the product (information related to the product described by SMILES tags) through the data acquisition unit 52.

[0217] In S3-2, the reaction condition providing device 5 uses the molecular descriptor deriving unit 532 and derives the molecular descriptor based on the information related to the chemical substance and the information related to the product. The processing here is similar to that of the model generating device, and the molecular descriptor deriving unit 532 performs the calculation of the chemical structure information and the acquisition of the physical property information. In S3-3, the reaction condition providing device 5 uses the data set generating unit 531 to add the molecular descriptor derived in S3-2 as an element constituting the data to the information related to the chemical substance, and further add the molecular descriptor as an element of the data set to the information related to the product, thereby generating a predetermined data set.

[0218] In S3-4, the reaction condition providing device 5 calculates the similarity between the information related to the plurality of chemical substances acquired by the data acquisition unit 52 and the information related to the plurality of chemical substances in the past reactions stored in the reaction condition database by the similarity calculation unit 542. Similarly, for the product, the reaction condition providing device 5 calculates the similarity between the information related to the product acquired by the data acquisition unit 52 and the information related to the product generated by the past reactions by the similarity calculation unit 542.

[0219] In S3-5, the reaction condition providing device 5 selects a reaction condition by comparing the similarity calculated in S3-4 with a predetermined threshold value using the first reaction condition selecting unit 543. In S3-6, the reaction condition providing device 5 sets the combination of the reaction conditions envisioned from the reaction conditions selected in S3-5 using the reaction condition range setting unit 544.

[0220] In S3 - 7 , when the reaction conditions set by the reaction condition range setting unit 544 do not satisfy predetermined requirements as a data set, the reaction condition providing device 5 uses the missing value processing unit 55 to perform predetermined missing value processing on the reaction conditions (data set).

[0221] In S3-8, the reaction condition providing device 5 uses the reaction simulation unit 571 of the analysis unit 57 to perform reaction simulation by inputting the reaction conditions set in S3-6 (among the reaction conditions, when the reaction conditions include missing values, the reaction conditions are the reaction conditions for which a predetermined missing value processing is performed) to the estimation model 56. When the execution of the reaction simulation is completed, the control unit 51 stores the execution result (yield) of the reaction simulation in the storage unit 59.

[0222] In S3-9, the reaction condition providing device 5 assigns a grade (performs a grade) to each of the plurality of reaction conditions stored in the storage unit 59 based on the yield (execution result of the reaction simulation) associated with the reaction condition using the grader 572 of the analysis unit 57. In S3-10, the reaction condition providing device 5 determines whether there is a reaction condition of a predetermined grade or higher based on the grade assigned in S3-9 using the determination unit 573 of the analysis unit 57. When it is determined by the determination unit 573 of the analysis unit 57 that there is a reaction condition of a predetermined grade or higher among the plurality of reaction conditions (S3-10 Yes), the reaction condition providing device 5 transfers the process to S3-11. When it is determined that there is no reaction condition of a predetermined grade or higher among the plurality of reaction conditions (S3-10 No), the process is transferred to S3-13.

[0223] When it is determined that there is a reaction condition of a predetermined level or higher, in S3-11, the reaction condition providing device 5 selects the reaction condition determined to satisfy the predetermined condition in S3-10 (determined to have a predetermined level or higher) as an output object by using the second reaction condition selecting unit 574 of the analyzing unit 57, and further assigns a display priority to the reaction condition selected as the output object. In S3-12, the reaction condition providing device 5 controls the external display device 7 to display the reaction condition selected as the output object by using the second reaction condition selecting unit 574 in S3-11 according to the display priority.

[0224] When it is determined that there is no reaction condition of the predetermined level or higher, in S3-13, the reaction condition providing device 5 optimizes (resets) the reaction condition using the third optimization unit 575. It should be noted that although the processing here will be described with reference to FIGS. 27A to 27B, Fig.29 The accompanying drawings will be described later, but when the number of yields obtained by the experiment (experimental points) is insufficient in the execution of optimization, the experimental points can be virtually set based on the results of the simulation performed using the inferred model. In S3-14, the reaction condition providing device 5 controls the reaction conditions obtained as a result of the execution of the optimization process in S3-13 to be displayed in descending order of the value of the acquisition function.

[0225] Next, refer to FIG. 24A to FIG. 24E , with respect to a flowchart showing the steps of the process of estimating (providing) the reaction conditions by the reaction condition providing device 5 described above ( Fig.23 ) will be specifically described in detail from S3-4 to S3-6. FIG. 24A to FIG. 24E (in particular FIG. 24C to FIG. 24E ) describes the similarity calculation process (S3-4), the reaction condition selection process (S3-5) and the reaction condition selection process (S3-6).

[0226] Fig.24A The data acquisition unit 52 acquires information related to a plurality of chemical substances (information related to chemical substance A and chemical substance B) and information related to products ( Fig.24A ); and information stored in the reaction condition database 541, which is related to a plurality of chemical substances (chemical substance 1 to chemical substance 8) reacted in experiments conducted in the past, information related to products (product 1 to product 4) generated as a result of reacting the plurality of chemical substances, and information related to reaction conditions when reacting the plurality of chemical substances ( Fig.24A on the right side of the Fig.24A In the , information related to the multiple chemical substances and information related to the products are described using SMILES tags. Information related to the reaction conditions when the multiple chemical substances react is shown as a reaction item.

[0227] Fig. 24B The ( Fig. 24B on the left side), Fig.24A The information on the left side of the diagram related to multiple chemical substances (information related to chemical substance A and chemical substance B) and information related to products ( Fig.24A on the left side of the ). Fig. 24B shows information related to a plurality of chemical substances (chemical substance 1 to chemical substance 8) converted into chemical structure information (fingerprint) by the molecular descriptor deriving unit 532 and information related to products (product 1 to product 4) generated as a result of reacting the plurality of chemical substances, and Fig.24A The information on the reaction conditions when reacting a plurality of chemical substances is shown on the right side of Fig. 24B on the right side of the screen).

[0228] Fig.24C The above-mentioned Tanimoto coefficients are used to obtain Fig. 24B The information on the left side of the table related to various chemical substances (chemical structure information of chemical substance A and chemical substance B) is Fig. 24BThe similarity between the information on the right side of the plurality of chemical substances (chemical structure information of chemical substance 1 to chemical substance 8), and Fig. 24B The information on the left side of the product (chemical structure information of the product) and Fig. 24B The calculation result of the similarity between the information related to a plurality of products (chemical structure information of product 1 to product 4) on the right side of . The calculated similarity is associated with the information related to the reaction conditions.

[0229] For example, in Fig.24C In the figure, the top row shows, from the left, (1) the similarity between the chemical structure information of chemical substance A obtained by the data acquisition unit 52 and the chemical structure information of chemical substance 1 reacted in a past experiment stored in the reaction condition database 541, and the calculation formula (Tanimoto coefficient); (2) the similarity between the chemical structure information of chemical substance B obtained by the data acquisition unit 52 and the chemical structure information of chemical substance 2 reacted in a past experiment stored in the reaction condition database 541, and the calculation formula (Tanimoto coefficient); (3) the similarity between the chemical structure information of the product obtained by the data acquisition unit 52 and the chemical structure information of product 1 generated as a result of reacting chemical substance 1 with chemical substance 2, stored in the reaction condition database 541, and the calculation formula (Tanimoto coefficient); and (4) the reaction conditions (temperature, concentration and solvent) when chemical substance 1 reacts with chemical substance 2.

[0230] Fig.24D Shown by Fig.24C , and the reaction conditions selected by comparing the similarity calculated in the data acquisition unit 52 with a predetermined threshold value. Specifically, the reaction conditions selected by calculating the average value of the similarities between the chemical structure information of the chemical substance A acquired by the data acquisition unit 52 and the chemical structure information of the chemical substance reacted in the past experiment stored in the reaction condition database 541, the similarity between the chemical structure information of the chemical substance B acquired by the data acquisition unit 52 and the chemical structure information of the chemical substance reacted in the past experiment stored in the reaction condition database 541, and the similarity between the chemical structure information of the product acquired by the data acquisition unit 52 and the chemical structure information of the product generated as a result of reacting the chemical substances stored in the reaction condition database 541 are shown. Here, the predetermined threshold value is set to 0.8, and the reaction conditions whose average value of the similarity is 0.8 or more are selected and shown.

[0231] exist Fig.24DIn the above, as described above, the reaction conditions are selected using the average of the similarities between the chemical structure information of the chemical substance A acquired by the data acquisition unit 52 and the chemical structure information of the chemical substance reacted in the past experiment stored in the reaction condition database 541, the similarity between the chemical structure information of the chemical substance B acquired by the data acquisition unit 52 and the chemical structure information of the chemical substance reacted in the past experiment stored in the reaction condition database 541, and the similarity between the chemical structure information of the product acquired by the data acquisition unit 52 and the chemical structure information of the product generated as a result of reacting the chemical substances stored in the reaction condition database 541. However, a weighted average may be used for selecting the reaction conditions. For example, a relatively high weight set for the similarity between the chemical structure information of the product acquired by the data acquisition unit 52 and the chemical structure information of the product generated as a result of reacting the chemical substances stored in the reaction condition database 541 may be used together with the average.

[0232] Fig.24E Show according to Fig.24D Here, as the reaction conditions, Fig.24D As shown, temperature, concentration and solvent constitute one set, and, based on the premise that there are two sets (i.e., based on the premise that the reaction conditions of temperature, concentration and solvent each have two values), 8 combinations (=2×2×2) are generated and shown. FIG. 24A to FIG. 24E As shown, by performing data processing in this way, reaction conditions are set.

[0233] Then, when the reaction conditions are set as described above, the reaction simulation is performed by inputting the set reaction conditions (when the set reaction conditions include missing values, the reaction conditions after missing value processing are performed) into the estimation model 56. Fig.25C , with respect to a flowchart showing the steps of the process of estimating (providing) the reaction conditions by the reaction condition providing device 5 described above ( Fig.23 ) will be specifically described in S3-8 to S3-12. That is, the execution process of the reaction simulation (S3-8), the classification process (S3-9), the process of determining whether there are reaction conditions that meet the predetermined conditions (S3-10), and the selection process of the reaction conditions that meet the predetermined conditions (S3-11) will be described.

[0234] FIG25A shows a data set obtained by adding (merging) the above reaction condition set shown in FIG24 to the predetermined data set generated in S3-3 (i.e., the data set includes molecular descriptors and physical property information as elements). By inputting the data set of FIG25A into the estimation model 56, a reaction simulation is performed.

[0235] The execution process of the reaction simulation here can refer to Fig.26 Let me explain in detail. Fig.26 2 is a diagram for explaining the reaction simulation process when a plurality of estimation models (three estimation models) are used. Fig.26 As shown, the data set of Figure 25A was input into each of the three imputation models.

[0236] Each of the estimation models performs calculation processing on the input (the data set in FIG. 25A ) and calculates a predicted value. Fig.26 In, from Fig.26 From the left side of , the estimation model 1 calculates the predicted value as y1, the estimation model 2 calculates the predicted value as y2, and the estimation model 3 calculates the predicted value as y3. Then, when the predicted value is calculated in each estimation model, the calculated predicted value is multiplied by the weight set for each estimation model, and the weights are further added.

[0237] That is, the following processing is performed: the calculated predicted value y1 is multiplied by the weight w1 set for the estimation model 1, the calculated predicted value y2 is multiplied by the weight w2 set for the estimation model 2, and the calculated predicted value y3 is multiplied by the weight w3 set for the estimation model 3, and the multiplication results are added. Then, the summed result (that is, the result of calculating y1×w1+y2×w2+y3×w3) is output as the yield.

[0238] Referring again to FIG. 25A , the calculated yields are shown on the right side of the data set of FIG. 25A . FIG. 25B shows the grades assigned to the calculated yields. Here, the grade is assigned to "A" when the yield is 80% or more, the grade is assigned to "B" when the yield is less than 80% and 50% or more, and the grade is assigned to "C" when the yield is less than 50%.

[0239] In FIG. 25B , reaction condition 2 is classified as “A”, reaction condition 1 is classified as “B”, and the other cases (reaction conditions 3 to 8) are classified as “C”. Based on this premise, the determination unit 573 of the analysis unit 57 determines whether there is a reaction condition of a predetermined level of “B” or higher. Here, since reaction condition 2 is “A” and reaction condition 1 is “B”, it is determined that there is a reaction condition of a level of “B” or higher. It should be noted that when there is no reaction condition of a predetermined level or higher (that is, here, when there is no reaction condition of a level of “B” or higher), as in the above description, Fig.23 As shown in the flowchart as a transition from S3-10 to S3-13, the third optimization unit 575 of the analysis unit 57 performs optimization processing to explore reaction conditions.

[0240] Fig.25CIt is shown that reaction condition 1 and reaction condition 2 are selected as reaction conditions of a predetermined level or higher, and further display priorities are set according to the levels (yields when the levels are the same). Here, since the level of reaction condition 2 is "A" and the level of reaction condition 1 is "B" among the selected reaction conditions, the display priority of reaction condition 2 is set to 1 and the display priority of reaction condition 1 is set to 2 (i.e., it is set in a manner that reaction condition 2 is displayed preferentially compared to reaction condition 1).

[0241] Next, referring to FIGS. 27A to 27B, Fig.29 , the processing of Bayesian optimization performed when the reaction conditions output as the estimation result do not satisfy the predetermined conditions (yield) is described. As described above, with respect to the reaction condition providing device 5, when the reaction conditions output as the estimation result do not satisfy the predetermined conditions (yield) ( Fig.23 In the flowchart of S3-10 (no case), Bayesian optimization is performed, and the reaction conditions are output based on the execution result of Bayesian optimization. When the yield is determined to be a function of the reaction conditions, Bayesian optimization is provided as a method for obtaining the maximum value of the yield (here, a method for obtaining the reaction conditions with the maximum yield).

[0242] In this embodiment, (instead of performing Bayesian optimization once and outputting (providing) one reaction condition), multiple Bayesian optimizations are performed at one time to provide multiple reaction conditions. Therefore, by providing multiple reaction conditions to the operator at one time, the operator can perform experiments (operations) in parallel in some cases, and as a result, the operation time can be reduced. It should be noted that when multiple Bayesian optimizations are performed at one time, the Constant Liar Approach is applicable.

[0243] 27A to 27B. Fig.27E The processing content here is explained as a supplementary explanation. For the convenience of explanation, the case where the number of reaction conditions as a characteristic quantity is 1 (for example, temperature) will be explained. Figure 27A shows a function (prediction line) predicted based on eight experimental points (results), with the horizontal axis representing temperature (reaction conditions) and the vertical axis representing yield, and the range of change is also shown in a veil shape. The eight experimental points are set based on the yields actually obtained through experiments. It should be noted that when the number of yields obtained through experiments is insufficient for performing Bayesian optimization, the experimental points can also be set based on the inferred results using the inferred model ( Fig.23 8 in the flowchart of FIG. 1 ).

[0244] In this case, the third optimization unit 575 determines the next experimental point using a predetermined acquisition function. As an acquisition function, for example, expected improvement (EI), probability of improvement (PI), upper confidence bound (UCB), etc. can be used. Here, PI is used as an example for explanation. In the case of PI, in a normal distribution including the sum of the mean value and the variance value as elements, the value ( Fig.28 The shaded portion of ) shows that the temperature at which the integral value becomes the highest is selected (determined) as the temperature at which the probability of exceeding the champion data yield is high. It should be noted that the normal distribution here is a type of probability distribution and is represented by the following formula.

[0245] [Mathematical formula 2]

[0246]

[0247] The value obtained by integrating from the champion data Ymax to infinity can be calculated by the following formula: In the following formula, ε is used so that a point very close to the champion data Ymax is not selected as the next experimental point according to the degree of variance, and is set to 0.01, for example.

[0248] [Mathematical formula 3]

[0249]

[0250] Then, by calculating the above formula while changing the reaction conditions (temperature), PI (x (n+1) ) becomes the highest value when x (n+1) (i.e., determine Fig.28 The area of ​​the shaded portion becomes the largest at x (n+1) (ie, the next sample)). In FIG. 27B, the temperature indicated by the triangle is determined as a temperature at which the probability of exceeding the yield of the champion data in the acquisition function is high.

[0251] In this embodiment, instead of directly conducting experiments (operations) to determine the yield based on the provided temperature, the Constant Liar Approach is applied as described above, and Bayesian optimization is further performed by assuming that the experimental points are located on the prediction line corresponding to the provided temperature (assuming the position of the square in Figure 27C).

[0252] When Bayesian optimization is further performed, the relationship (function) between temperature and yield is shown in FIG. 27D , and the temperature ( Fig.27E ). In addition, the Bayesian optimization is further performed by assuming that the experimental point is located on the prediction line corresponding to the provided temperature (assuming the position of × in the figure).

[0253] Therefore, by performing multiple Bayesian optimizations at one time, multiple reaction conditions are output (provided). That is, the reaction conditions are output at one time according to the number of experiments (operations) that the operator can perform in parallel. Below, the processing steps here are supplemented using a flowchart.

[0254] Fig.29 A flowchart showing the steps of applying the Constant Liar Approach to perform Bayesian optimization to determine reaction conditions. Fig.29 In the flowchart of , for example, the start of the process is triggered by the fact that the reaction conditions output as the estimation result do not satisfy the predetermined conditions (yield).

[0255] In S4-1, the third optimization unit 575 performs Bayesian optimization based on known experimental points. As described above, when the number of known experimental points is insufficient for performing Bayesian optimization, the experimental points are set based on the results of estimation using the estimation model. In S4-2, the third optimization unit 575 determines a reaction condition with a high probability of exceeding the yield of the champion data based on the acquisition function.

[0256] In S4-3, the third optimization unit 575 determines whether the number of determined reaction conditions reaches a predetermined number. Here, the variable n is set to 1 as an initial value (set to n=1), and when three reaction conditions are determined (set), the process of determining the reaction conditions ends. The predetermined number can be set by the operator according to the operator's experience and the content of the experiment. As a predetermined number, in addition to the number of experiments that the operator can implement in parallel, for example, when you want to know other experimental candidates as a reference, you can examine the number of experimental candidates, or, when you want to keep the feature quantity within a specific range (for example, you want to keep the temperature at 50°C or lower), you can examine the number of parallel outputs.

[0257] In S4-4, as described above with reference to FIG. 27, the third optimization unit 575 temporarily sets the experimental point based on the value of the reaction condition with a high probability of exceeding the yield of the champion data. In S4-5, the third optimization unit 575 performs Bayesian optimization again, including the temporarily set experimental point. In S4-6, the third optimization unit 575 determines the reaction condition with a high probability of exceeding the yield of the champion data based on the acquisition function.

[0258] Then, when executing the process of S4-6, the control unit 51 sets the variable n (in Fig.29) counts (i.e., adds 1 to the variable n and sets n to 2), and returns the process to S4-3. Therefore, the processes of S4-3 to S4-6 are repeatedly performed until the variable n reaches a predetermined number. When the variable n reaches a predetermined number, Fig.29 The processing shown in the flowchart is completed. By executing the processing in this way, the reaction conditions can be output in the number of operations that can be performed in parallel, and as described above, the operation man-hours can be reduced.

[0259] Although the number of reaction conditions (i.e., feature quantity) is 1 in the above description, and in FIG. Fig.27E In the example, the two-dimensional method is used to describe the reaction conditions. However, in reality, multiple reaction conditions are set. Therefore, in the reaction condition providing device, calculations are performed by matrix operations. Fig. 22 Although not shown in the figure, the third optimization unit 575 includes as functions a derivation unit for deriving reaction conditions whose yields satisfy predetermined conditions using a predetermined acquisition function (for example, reaction conditions with the highest yield, reaction conditions with a yield equal to or greater than a predetermined threshold, etc.), and a setting unit for temporarily setting experimental points using a function of the yield versus the reaction conditions according to the reaction conditions derived by the derivation unit.

[0260] Fig.30 It is a diagram for explaining a sample screen of reaction conditions displayed on the external display device 7 . Fig.30 It is shown that three reaction conditions selected as reaction conditions having a predetermined rank or higher by the second reaction condition selection unit 574 of the analysis unit 57 are displayed according to the assigned display priorities.

[0261] like Fig.30 As shown, the displayed reaction conditions (dropping rate, temperature, control conditions) are associated with the display priority, the predicted value (yield) when chemical substance A reacts with chemical substance B, and chemical substance A and chemical substance B as reaction object materials to form a data set, and these data sets are displayed in descending order according to the display priority.

[0262] The operator selects the reaction conditions displayed on the display device 7 based on the availability of materials, cost and law, and performs experiments (operations) according to the reaction conditions. It should be noted that by inputting the experimental results into the model generation device 1 and further performing machine learning, a model that can provide more accurate reaction conditions can be generated.

[0263] In addition, as a supplementary explanation, in the above-mentioned embodiment, although information related to two chemical substances, that is, information related to chemical substance A and information related to chemical substance B are used as information related to multiple chemical substances for explanation, even if information related to three or more chemical substances is input, reaction conditions can be provided. When information related to three or more chemical substances is input, as an element constituting chemical property information generated as a predetermined data set, for example, in addition to chemical substance A and chemical substance B, chemical structure information and physical property information of chemical substances such as chemical substance C and chemical substance D are added, and the data set (chemical property information) constructed in the above manner is taken as a premise, and similarity calculation processing, reaction condition selection and setting processing and reaction simulation are performed. In this case, experimental data for reacting three or more chemical substances can be stored in the reaction condition database. This also applies to the Bayesian optimization of S3-13.

[0264] As described above, according to an embodiment, when the information obtained related to a plurality of chemical substances and products is similar to the information related to a plurality of chemical substances and products set for the reaction conditions (or the conditions are the same), the plurality of reaction conditions obtained when the plurality of chemical substances are reacted are set from the reaction items set for the reaction conditions. In addition, an inference model obtained by performing machine learning using the chemical structure information and physical property information of a plurality of chemical substances reacted in the past, the reaction conditions in the reaction, and the yield when the reaction is carried out under the reaction conditions as training data is used to infer the yield for each of the plurality of reaction conditions, and the reaction conditions whose yields meet the predetermined conditions among the inferred yields are displayed (output). Thus, the operation man-hours can be reduced in the setting of the reaction conditions whose yields meet the predetermined conditions.

[0265] Fig.31 1 is a block diagram illustrating the hardware configuration of the model generation device 1. Fig.31 As shown, the model generation device 1 includes a CPU 901, a RAM 902, a ROM 903, a HDD 904, a GPU 905, an operating device interface 906, a network interface 907, a device interface 908, and an external storage interface 909. In addition to the operating device connected via the operating device interface 906, the model generation device 1 is also connected to an external device (display device 7) and an external storage device 4.

[0266] A CPU (Central Processing Unit) 901 is a processing unit that integrally controls the modules of the model generation device 1. The CPU 901 is connected to a RAM 902, a ROM 903, a HDD 904, a GPU 905, an operation device interface 906, a network interface 907, a device interface 908, and an external storage interface 909 via a system bus 910.

[0267] RAM (Random Access Memory) 902 includes a storage area for temporarily storing the calculation processing results of CPU 901, various set values, parameters, etc., and a loading area for various control programs. ROM (Read Only Memory) 903 stores various programs (for example, startup programs). HDD (Hard Disk Drive) 904 stores physical property information, reaction conditions, generated estimated models, etc. The function of the model generation device is implemented as follows, for example, CPU 901 reads the program stored in ROM 903 into RAM 902 and executes the program.

[0268] The GPU (Graphics Processing Unit) 905 is a processing unit for executing processing operations at high speed, and performs predetermined calculations transferred from the CPU 901 (for example, matrix operations performed in the generation stage of the estimated model, etc.). The operating device interface 906 is an interface for inputting data to the operating device 6 or obtaining data from the operating device 6. The network interface 907 is connected to the local area network (LAN) by wire or wirelessly, and allows information input and output with external devices (for example, allowing physical property information, reaction conditions and various programs to be updated). The device interface 908 is an interface for connecting to, for example, an external analysis device. The external storage interface 909 is an interface for connecting to an external storage device. For example, when the model generation device and the reaction condition providing device are configured as different devices, the external storage interface 909 is used to store the estimated model generated by the model generation device 1 and stored in the HDD in the external storage device via the external storage interface 909.

[0269] It should be noted that the hardware structure of the reaction condition providing device 5 is almost similar to the hardware structure of the model generating device 1, but the various storage elements, objects (data) stored in the storage device, and external devices (for example, the display device 7, etc.) connected via the device interface 908 are different.

[0270] Furthermore, the present invention can be implemented by supplying a program that implements one or more functions in the above-described embodiments to a device via a network or a storage medium, and having one or more processors in a computer of the device read and execute the program.

[0271] A part or all of the above-mentioned embodiment can be described as the following supplementary notes, but the above-mentioned embodiment is not limited thereto.

[0272] (Note)

[0273] A reaction condition providing device, comprising:

[0274] an optimization unit, which optimizes the function of yield versus reaction conditions based on experimental points during the process of exploring reaction conditions;

[0275] A deriving unit, which uses a predetermined acquisition function to derive a reaction condition whose yield satisfies a predetermined condition;

[0276] a setting unit that temporarily sets the experimental point according to the derived reaction condition using the function of the yield to the reaction condition; and

[0277] a display control unit that causes a display device to display the reaction conditions derived by the deriving unit,

[0278] The optimization of the function of the yield to the reaction conditions, the derivation of the reaction conditions in which the yield satisfies the predetermined conditions, and the setting of the experimental points are repeatedly performed a predetermined number of times.

[0279] Description of Reference Numerals

[0280] 1 Model generation device

[0281] 2 Analysis device

[0282] 3 Reactor

[0283] 4 External storage devices

[0284] 5. Provide equipment

[0285] 6 Operating device

[0286] 7 Display device

[0287] 11 Control Unit

[0288] 12 Data acquisition unit

[0289] 13 Training data generation unit

[0290] 14 Pre-processing unit

[0291] 15. Feature selection unit

[0292] 16 Inference model generation unit

[0293] 17 Storage Unit

[0294] 51 Control unit

[0295] 52 Data acquisition unit

[0296] 53 Chemical property information generation unit

[0297] 54 Reaction condition setting unit

[0298] 55 Missing Value Processing Unit

[0299] 56 Estimated Model

[0300] 57 Analysis Units

[0301] 58 Display control unit

[0302] 59 Storage Units

Claims

1. A method for generating an inference model, comprising: An acquisition step of acquiring information related to a plurality of chemical substances reacted in the past, reaction conditions in the reaction, information related to products generated by reacting the plurality of chemical substances, and yields when the reaction is performed under the reaction conditions; A first generation step of generating training data from chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances in the past reaction, the reaction conditions, chemical structure information and physical property information derived from information related to the product, and the yield; and The second generation step generates an inference model by performing machine learning using the training data, wherein the inference model takes information related to the multiple chemical substances to be reacted and information related to the products to be generated as input values, and takes reaction conditions in which the yield of the products to be generated meets predetermined conditions as output values.

2. The method for generating an inference model according to claim 1, wherein: The second generation step generates the estimation model based on a feature quantity as an element constituting the training data, and the second generation step includes: a first storing step of storing the verified accuracy of the generated estimation model in association with the estimation model; a calculation step of calculating the importance of the feature quantity in the generated estimation model; a determination step of determining whether the feature quantities include deletable feature quantities based on the properties of the feature quantities and the importance of the feature quantities; and a deleting step, when it is determined in the determining step that the deletable feature exists, deleting the feature determined to be deletable from the elements constituting the training data, Among them, until it is determined in the determination step that the feature quantities do not include deletable feature quantities, the generation of the estimated model, the first storage step, the calculation step, the determination step and the deletion step are repeatedly performed in sequence, and the estimated model whose verified accuracy meets the predetermined conditions is selected from the estimated models stored in the first storage step.

3. The method for generating an inference model according to claim 1, wherein: The second generating step comprises: An optimization step of optimizing the hyperparameters set for the machine learning algorithm; and a second storage step of generating the estimated model with the optimized hyperparameters and storing the verified accuracy of the generated estimated model in association with the estimated model, Among them, until the expected improvement degree output in the optimization step satisfies the predetermined condition, the optimization step, the generation of the estimated model and the second storage step are repeatedly performed in sequence, and when the expected improvement degree satisfies the predetermined condition, the estimated model whose verified accuracy satisfies the predetermined condition is selected from the estimated models stored in the second storage step.

4. The method for generating an inference model according to claim 1, further comprising: A selection step of selecting a feature quantity for generating the estimated model; a segmentation step of segmenting the training data according to the number of multiple inference models to be further generated; A third generation step of generating a plurality of estimation models by performing machine learning using a predetermined machine learning method for each of the segmented training data based on the selected feature quantities; and The setting step sets a predetermined weight to each of the plurality of generated estimation models.

5. The method for generating an inference model according to claim 4, wherein: In the segmentation step, the training data is segmented based on a bootstrapping method.

6. The method for generating an inference model according to claim 4, wherein: The predetermined machine learning method is any one of a decision tree, a neural network, a gradient boosting, and a support vector, and A part of the plurality of generated inference models is allowed to be generated by performing the machine learning using the same machine learning method.

7. The method for generating an inference model according to claim 4, wherein: The predetermined weights set for the plurality of generated estimation models are estimated by Bayesian optimization.

8. The estimation model generation method according to claim 1, further comprising the step of performing predetermined pre-processing corresponding to the training data when the training data generated in the first generation step does not satisfy predetermined requirements.

9. The method for generating an inference model according to claim 8, wherein: The pre-processing is at least one of missing value processing, category value conversion processing, or data standardization or normalization.

10. A non-transitory computer-readable storage medium storing instructions for causing a computer to generate an inference model, wherein the instructions are configured to cause the computer to execute a function, the function comprising: Acquiring information on a plurality of chemical substances reacted in the past, reaction conditions in the reactions, information on products generated by reacting the plurality of chemical substances, and yields when the reactions are performed under the reaction conditions; generating training data from chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances in the past reactions, the reaction conditions, chemical structure information and physical property information derived from information related to the products, and the yields; and An estimation model is generated by performing machine learning using the training data. The estimation model takes information about a plurality of chemical substances to be reacted and information about a product to be generated as input values ​​and takes a reaction condition in which the yield of the product to be generated satisfies a predetermined condition as an output value.

11. An inference model generating device comprising a processor programmed to: Acquiring information on a plurality of chemical substances reacted in the past, reaction conditions in the reactions, information on products generated by reacting the plurality of chemical substances, and yields when the reactions are performed under the reaction conditions; generating training data from chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances in the past reactions, the reaction conditions, chemical structure information and physical property information derived from information related to the products, and the yields; and An estimation model is generated by performing machine learning using the training data. The estimation model takes information about a plurality of chemical substances to be reacted and information about a product to be generated as input values ​​and takes a reaction condition in which the yield of the product to be generated satisfies a predetermined condition as an output value.

12. A reaction condition providing device, comprising: a memory storing one or more reaction conditions; and A processor programmed to: Acquiring information related to a plurality of chemical substances as reaction objects and information related to products to be generated; When the similarity between the information related to the plurality of chemical substances as reaction objects and the information related to the products to be generated and the information related to the plurality of chemical substances and the information related to the products set for the reaction conditions is equal to or greater than a predetermined threshold value, setting a plurality of reaction conditions for reacting the plurality of chemical substances as reaction objects from among the reaction items set for one or more reaction conditions; Using an estimation model, estimating a yield for each of the plurality of reaction conditions based on chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances as reaction objects, chemical structure information and physical property information derived from information related to the products to be generated, and a plurality of set reaction conditions, wherein the estimation model is obtained by performing machine learning using as training data the chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances reacted in the past, the reaction conditions in the reaction, the chemical structure information and physical property information derived from information related to the products generated in the reaction of the plurality of chemical substances, and the yield when the reaction is performed under the reaction conditions; and The display device displays the reaction condition whose yield satisfies the predetermined condition among the yields estimated for the respective reaction conditions.

13. The reaction condition providing device according to claim 12, wherein: The processor is programmed to: calculating similarities between the information on the plurality of chemical substances as reaction targets and the information on the products to be generated and the information on the plurality of chemical substances and the information on the products set for the one or more reaction conditions; and Calculate the average or weighted average of the similarities between the chemical structure information and physical property information derived from each piece of information related to the multiple chemical substances as reaction objects and the chemical structure information and physical property information derived from each piece of information related to the multiple chemical substances set for the one or more reaction conditions, and the average or weighted average of the similarities between the chemical structure information and physical property information derived from the information related to the product to be generated and the chemical structure information and physical property information derived from the information related to the product set for the one or more reaction conditions.

14. The reaction condition providing device according to claim 12, wherein: The processor is programmed to: Determining whether there is a reaction condition whose yield satisfies a predetermined condition among the yields estimated for each reaction condition; and When it is determined that a plurality of reaction conditions exist whose yields satisfy the predetermined condition, display priorities are assigned to the plurality of reaction conditions.

15. The reaction condition providing device according to claim 14, wherein: The processor is programmed to optimize the function of the yield versus the reaction condition in the process of exploring the reaction condition when it is determined that the reaction condition whose yield satisfies the predetermined condition does not exist.

16. The reaction condition providing device according to claim 15, wherein: The processor is programmed to: The yield was optimized as a function of the reaction conditions based on the experimental points; Using a predetermined acquisition function, deriving a reaction condition whose yield satisfies a predetermined condition; Using the function of yield versus reaction conditions, set the experimental points according to the derived reaction conditions; Repeating the optimization of the function of the yield to the reaction conditions, the derivation of the reaction conditions whose yield satisfies the predetermined conditions, and the setting of the experimental points for a predetermined number of times; and The display device is caused to display the derived reaction conditions.

17. A method for providing reaction conditions, comprising: An acquisition step of acquiring information related to a plurality of chemical substances as reaction objects and information related to products to be generated; a setting step of setting a plurality of reaction conditions for reacting the plurality of chemical substances as reaction objects from reaction items set for one or more reaction condition settings stored in a memory, when a degree of similarity between the information related to the plurality of chemical substances as reaction objects and the information related to the product to be generated and the information related to the plurality of chemical substances and the information related to the product set for the reaction conditions is equal to or greater than a predetermined threshold; an estimation step, using an estimation model to estimate the yield for each of the plurality of reaction conditions based on chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances as reaction objects, chemical structure information and physical property information derived from information related to the products to be generated, and a plurality of set reaction conditions, wherein the estimation model is obtained by performing machine learning using the chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances reacted in the past, the reaction conditions in the reaction, the chemical structure information and physical property information derived from information related to the products generated in the reaction of the plurality of chemical substances, and the yield when the reaction is performed under the reaction conditions as training data; and A display control step of causing a display device to display, among the yields estimated for each reaction condition in the estimating step, a reaction condition whose yield satisfies a predetermined condition.

18. A non-transitory computer-readable storage medium storing instructions for causing a computer to provide a reaction condition, wherein the instructions are configured to cause the computer to perform a function, the function comprising: Acquiring information related to a plurality of chemical substances as reaction objects and information related to products to be generated; When the similarity between the information related to the plurality of chemical substances as reaction objects and the information related to the products to be generated and the information related to the plurality of chemical substances and the information related to the products set for the reaction conditions is equal to or greater than a predetermined threshold value, setting a plurality of reaction conditions for reacting the plurality of chemical substances as reaction objects from reaction items set for one or more reaction conditions stored in a memory; Using an estimation model, estimating a yield for each of the plurality of reaction conditions based on chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances as reaction objects, chemical structure information and physical property information derived from information related to the products to be generated, and a plurality of set reaction conditions, wherein the estimation model is obtained by performing machine learning using as training data the chemical structure information and physical property information derived from each piece of information related to the plurality of chemical substances reacted in the past, the reaction conditions in the reaction, the chemical structure information and physical property information derived from information related to the products generated in the reaction of the plurality of chemical substances, and the yield when the reaction is performed under the reaction conditions; and The display device displays the reaction condition whose yield satisfies the predetermined condition among the yields estimated for the respective reaction conditions.

Images