Device for learning and predicting activation energy in chemical reactions

Abstract

To provide a learning device capable of learning to accurately predict activation energy.SOLUTION: A learning device includes: a first acquisition unit for acquiring a first explanatory variable including a structural parameter of a reaction system and electronic property calculated by quantum chemical calculation; a second acquisition unit for acquiring a second explanatory variable including the structural parameter and electronic property of a production system generated by chemically reacting the reaction system calculated by quantum chemical calculation; and a third acquisition unit for acquiring a target variable including activation energy in a chemical reaction of the reaction system calculated by quantum chemical calculation; and a learning unit for learning a model using learning data in which the first explanatory variable and the second explanatory variable are associated with the target variable.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a learning device, a prediction device, a learning program, a prediction program, a learning method, and a prediction method for activation energy in a chemical reaction. [Background technology]

[0002] Activation energy is the energy required for a chemical reaction to proceed, and is an important factor for estimating the ease of progress of a chemical reaction in various fields such as the chemical industry and pharmaceuticals. Therefore, a method for calculating the magnitude of activation energy in a realistic calculation time and environment is required.

[0003] Here, a model-based prediction technique has been proposed as a technique for predicting the excited state of reactants, new materials, new combinations, and the like.

[0004] For example, a method has been proposed in which machine learning is performed using data on the ground state and excited state of a molecule, and the properties of the excited state of the molecule are predicted using a model generated by the machine learning (see, for example, non-patent document 1). [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Kawai et al. "Predicting excited states from ground state wavefunction by supervised quantum machine learning", Mach. Learn.: Sci. Technol. 1 (2020) 045027 Summary of the Invention [Problem to be solved by the invention]

[0006] When attempting to directly calculate activation energy by quantum chemical calculation, there is a problem that it is difficult to search for transition states. In addition, since the calculation method used to calculate activation energy is different from the calculation method used to calculate excited states, there is a problem that the technique of Non-Patent Document 1 cannot be applied.

[0007] An object of one aspect of the present invention is to provide a learning device that can learn to predict activation energy with high accuracy. [Means for solving the problem]

[0008] One aspect of the learning device of the present invention has a first acquisition unit that acquires first explanatory variables calculated by quantum chemical calculation, the first explanatory variables including structural parameters and electronic properties of a reaction system; a second acquisition unit that acquires second explanatory variables calculated by quantum chemical calculation, the second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system; a third acquisition unit that acquires a target variable calculated by quantum chemical calculation, the second explanatory variables including an activation energy in a chemical reaction of the reaction system; and a learning unit that learns a model using learning data in which the first explanatory variables and the second explanatory variables are associated with the target variables.

[0009] One aspect of the prediction device according to the present invention has a first acquisition unit that acquires first explanatory variables calculated by quantum chemical calculations, the first explanatory variables including structural parameters and electronic properties of a reaction system; a second acquisition unit that acquires second explanatory variables calculated by quantum chemical calculations, the second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system; a trained model trained using learning data in which a first explanatory variable and a second explanatory variable calculated in advance for a specific reaction system are associated with an objective variable including an activation energy in a chemical reaction of the specific reaction system; and a prediction unit that inputs the first explanatory variable and the second explanatory variable of the reaction system to be predicted, and outputs an objective variable including the activation energy in a chemical reaction of the reaction system to be predicted, predicted by the trained model.

[0010] One aspect of the learning program of the present invention causes a computer to execute a first acquisition step of acquiring first explanatory variables including structural parameters and electronic properties of a reaction system calculated by quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system calculated by quantum chemical calculation; a third acquisition step of acquiring a target variable including an activation energy in a chemical reaction of the reaction system calculated by quantum chemical calculation; and a learning step of learning a model using learning data in which the first explanatory variables and the second explanatory variables are associated with the target variables.

[0011] One aspect of the prediction program according to the present invention causes a computer to execute a first acquisition step of acquiring first explanatory variables including structural parameters and electronic properties of a reaction system calculated by quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, calculated by quantum chemical calculation; and a prediction step of inputting the first explanatory variables and the second explanatory variables of a reaction system to be predicted into a trained model trained using training data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system are associated with a target variable including an activation energy in a chemical reaction of the predetermined reaction system, thereby outputting a target variable including an activation energy in a chemical reaction of the reaction system to be predicted, which is predicted by the trained model.

[0012] One embodiment of the learning method according to the present invention includes a first acquisition step of acquiring first explanatory variables calculated by quantum chemical calculation, the first explanatory variables including structural parameters and electronic properties of a reaction system; a second acquisition step of acquiring second explanatory variables calculated by quantum chemical calculation, the second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system; a third acquisition step of acquiring a response variable calculated by quantum chemical calculation, the second explanatory variables including an activation energy in a chemical reaction of the reaction system; and a learning step of learning a model using learning data in which the first explanatory variables and the second explanatory variables are associated with the response variables.

[0013] One aspect of the prediction method according to the present invention includes a first acquisition step of acquiring first explanatory variables calculated by quantum chemical calculation, the first explanatory variables including structural parameters and electronic properties of a reaction system; a second acquisition step of acquiring second explanatory variables calculated by quantum chemical calculation, the second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system; and a prediction step of inputting the first explanatory variables and the second explanatory variables of a reaction system to be predicted into a trained model trained using training data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system are associated with a target variable including an activation energy in a chemical reaction of the predetermined reaction system, thereby outputting a target variable including an activation energy in a chemical reaction of the reaction system to be predicted, the target variable being predicted by the trained model. Effect of the Invention

[0014] One aspect of the learning device according to the present invention is capable of learning to predict activation energy with high accuracy. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a diagram showing the change in free energy as the reaction progresses from a reaction system to a product system. [Diagram 2]2A and 2B are block diagrams showing the schematic configuration of a learning device and a prediction device according to an embodiment of the present invention, where FIG. 2(a) is a block diagram showing the schematic configuration of a learning device, and FIG. 2(b) is a block diagram showing the schematic configuration of a prediction device. [Diagram 3] FIG. 2 is a block diagram showing a schematic configuration of a learning unit. [Figure 4] 4A and 4B are diagrams showing an example of learning data and prediction data, where FIG. 4A shows the learning data and FIG. 4B shows the prediction data. [Diagram 5] FIG. 2 is a block diagram showing a schematic configuration of a prediction unit. [Figure 6] FIG. 2 is a block diagram showing the hardware configuration of a learning device and a prediction device. [Figure 7] 1 is a flowchart illustrating a learning method according to an embodiment of the present invention. [Figure 8] 1 is a flowchart illustrating a prediction method according to an embodiment of the present invention. [Figure 9] FIG. 13 is a diagram showing the distribution of values ​​of a response variable stored in a data table. [Figure 10] FIG. 13 is a diagram showing prediction verification results of learning data in a machine learning model. [Figure 11] FIG. 13 is a diagram in which calculated values ​​and predicted values ​​are plotted together with training data. [Figure 12] FIG. 1 is a plot of calculation time versus the number of electrons in the system. [Figure 13] FIG. 13 is a diagram showing the distribution of the number of electrons in the system for 94 patterns of learning data and prediction targets A to E. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, an embodiment of the present invention will be described in detail. In order to facilitate understanding of the description, the same components in each drawing are given the same reference numerals, and duplicated descriptions will be omitted. In addition, in this specification, a tilde "~" indicating a numerical range means that the numerical values ​​before and after it are included as the lower and upper limits, unless otherwise specified.

[0017] A learning device and a prediction device according to the present embodiment will be described. FIG. 1 is a diagram showing a change in free energy when a reaction progresses from a reaction system to a product system. As shown in FIG. 1, the learning device according to the present embodiment performs machine learning using a first explanatory variable of a reaction system in a stable state, a second explanatory variable of a product system in a stable state, and a target variable including an activation energy obtained from the difference between the free energy of the reaction system and the free energy of a transition state, to generate a learned model.

[0018] The prediction device of this embodiment predicts a response variable including activation energy by inputting a first explanatory variable of a reaction system in a stable state and a second explanatory variable of a production system in a stable state into a trained model in which the first explanatory variable and the second explanatory variable correspond to the response variable.

[0019] <Learning device> A learning device according to this embodiment will be described. FIG. 2 is a block diagram showing a schematic configuration of a learning device and a prediction device according to this embodiment. As shown in FIG. 2(a), the learning device 1 includes a first acquisition unit 11, a second acquisition unit 12, a third acquisition unit 13, and a learning unit 14. The learning device 1 may further include a data table creation unit 15 and a display unit 16. The learning device 1 learns the value of activation energy that occurs when reactants such as compounds undergo a chemical reaction, pass through a transition state, and generate a product.

[0020] The chemical reactions to be studied are not particularly limited as long as they are chemical reactions in which a substance passes through a transition state to produce a product. For example, the Diels-Alder reaction, an organic chemical reaction in which an alkene is added to a conjugated diene to produce a six-membered ring structure, can be mentioned.

[0021] The first acquisition unit 11 acquires first explanatory variables including structural parameters and electronic properties of a reaction system calculated by quantum chemical calculation.

[0022] The first explanatory variables are various pieces of information such as structural parameters and electronic properties related to the reaction system obtained by low-precision quantum chemical calculations (low-precision quantum chemical calculations). The first explanatory variables will be described in detail later.

[0023] In this embodiment, the term "low accuracy" varies depending on the conditions of the reaction system and the like, and refers to an accuracy lower than that of various information calculated under high accuracy calculation conditions. The meaning of "low accuracy" below is used in the same sense. The accuracy of the calculation conditions can be determined by the method and basis set used in the quantum scientific calculation, as described later.

[0024] The second acquisition unit 12 acquires second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation.

[0025] The second explanatory variables are various pieces of information such as structural parameters and electronic properties of the product obtained by low-precision quantum chemical calculations. The second explanatory variables will be described in detail later.

[0026] The third obtaining unit 13 obtains a target variable including an activation energy in a chemical reaction of the reaction system, the activation energy being calculated by a quantum chemical calculation.

[0027] The objective variables are various pieces of information including the activation energy obtained by highly accurate quantum chemical calculations (high-precision quantum chemical calculations). Details of the objective variables will be described later.

[0028] In this embodiment, the term "high accuracy" varies depending on the reaction system conditions and the like, and refers to the accuracy of various pieces of information calculated under high accuracy calculation conditions. The term "high accuracy" below is used in the same sense.

[0029] The first explanatory variable, the second explanatory variable, and the objective variable are all calculated by quantum chemical calculation as described above. The quantum chemical calculation can be performed using general-purpose software, and for example, Gaussian16 (manufactured by Gaussian) can be suitably used.

[0030] When using Gaussian 16 to find a stable molecular structure, the calculation method, basis set, initial coordinates of the molecular model, charge, and spin multiplicity are input to create an input file, and the calculation is performed to find the stable molecular structure.

[0031] Quantum chemical calculation methods include the Hartree-Fock method and density functional theory (DFT) method, among which the DFT method is preferred from the viewpoint of practically satisfying the balance between costs such as calculation time and calculation accuracy.

[0032] It should be noted that in the DFT method, the calculation results can vary significantly depending on which functional is selected. Therefore, in calculations that require accuracy, it is preferable to select a functional that includes long-range correction (taking into account the effects acting between two particles at long distances) and dispersion force correction (taking into account weak intermolecular interactions such as van der Waals forces). Representative functionals that combine long-range correction and dispersion force correction include, for example, ωB97XD and M06-2X.

[0033] As an example of the hierarchy of some of the functionals that can be used in Gaussian16, for the M06 series the hierarchy is M06-L, M06, M06-2X, and for the ωB97 series the hierarchy is ωB97, ωB97X, ωB97XD, which increase in calculation accuracy but increase in calculation cost.

[0034] As with quantum chemical calculation methods, the basis set is an important factor that affects the calculation accuracy and calculation cost. In general, in order to reproduce molecular orbitals more accurately, it is desirable to use a basis set that can assign as many functions as possible. However, the higher the level of accuracy, the higher the calculation cost. Therefore, it is necessary to select a basis set that is within the range that can be used realistically. For example, the 3-21G basis set can be calculated at a low calculation cost. However, the accuracy is low, and the quantitativeness of the results obtained is not guaranteed. On the other hand, the 6-311++G (3df, 3pd) basis set can obtain highly accurate results. However, the calculation cost is high, and in the case of systems with large molecular sizes, it may not be possible to obtain results within a realistic time frame.

[0035] As an example, the order of basis functions in the Pople system is STO-3G, 3-21G, 3-21G * The computational accuracy increases but the computational cost increases in the following order: 6-31G, 6-31G(d), 6-311G(d), 6-311+G(d), 6-311+G(d,p), 6-311+G(2d,p), 6-311++G(2d,p), 6-311++G(3df,3pd).

[0036] The learning unit 14 performs machine learning of the model using learning data in which the first explanatory variable, the second explanatory variable, and the objective variable are associated with each other.

[0037] 3 is a block diagram showing a schematic configuration of the learning unit 14. As shown in FIG. 3, the learning unit 14 includes a read unit 141, a model M1, and a prediction accuracy verification unit 142.

[0038] The reading unit 141 reads out the learning data from the learning data table created by the data table creating unit 15 .

[0039] The model M1 is a model that calculates a response variable including an activation energy value from a first explanatory variable and a second explanatory variable.

[0040] Model M1 uses the computing power of a computer to perform machine learning on target data and makes predictions on unknown data. Generally, a programming language such as Python is often used for a program for executing machine learning, and a development environment with functions such as numerical calculation and data analysis must be constructed. In order to easily prepare a machine learning environment, it is preferable to use a method using, for example, Anaconda (registered trademark), which is software distributed by Anaconda, Inc., USA.

[0041] Anaconda (registered trademark) includes libraries that are often used in Python and machine learning, and these can be introduced all at once, making it easy to set up a machine learning environment. The libraries included in Anaconda (registered trademark) are also useful for data analysis before running machine learning, which will be described next.

[0042] The learning unit 14 can check the learning data. The learning unit 14 can check the distribution of values ​​of each variable, the presence or absence of correlation between variables, and the like for the first explanatory variable, the second explanatory variable, and the objective variable of each sample stored in the learning data table created by the data table creation unit 15.

[0043] The distribution of values ​​can be visually confirmed by plotting a histogram. In general, the more diverse the data group to be trained is, the more likely it is that a wide range of unknown data can be predicted, so it is preferable that the variable distribution of the training data is as wide and unbalanced as possible. In particular, when the distribution of the objective variable is extremely biased, it is preferable to add sample data to the training data that takes values ​​in a range that alleviates the bias.

[0044] The presence or absence of correlation between variables can be estimated by calculating the correlation coefficient between two sets of variable data. The correlation coefficient can be calculated, for example, by the following formula (1).

[0045]

number

[0046] The presence or absence of correlation between variables can also be visually confirmed by plotting a scatter plot of two sets of variable data for all combinations of variables. When an explanatory variable that is strongly correlated with the objective variable exists, it is highly likely that the explanatory variable contributes greatly to determining the objective variable. On the other hand, an explanatory variable that shows no correlation with the objective variable at all is unrelated to the objective variable and is highly likely to be unnecessary for predicting unknown data. In addition, it is generally known that when there is a strong correlation between two explanatory variables, the prediction accuracy deteriorates due to multicollinearity. Therefore, in this case, it is effective to take measures such as deleting one of the explanatory variables from the training data.

[0047] The model M1 can perform regression analysis on the prepared learning data to calculate a response variable from the first explanatory variable and the second explanatory variable. There are generally various regression methods used to predict a response variable that takes a continuous value, such as linear regression and random forest methods, but it is basically preferable to adopt a method that suits the target data set while looking at the verification results of the prediction accuracy after machine learning.

[0048] The prediction accuracy verification unit 142 verifies (checks) the prediction accuracy of the model M1 obtained by the regression analysis. For example, cross-validation is a method for verifying the prediction accuracy. Cross-validation divides prepared data into several groups, one of which is used as test data and the rest as learning data. Then, learning and validation are repeated the same number of times as the number of divided groups while changing the learning data and the testing data, and the obtained results are averaged to evaluate the prediction accuracy of the model. When the number of data is limited, leave-one-out cross-validation, in which only one of the entire data is used as test data, is particularly useful.

[0049] The accuracy evaluation index that shows the degree to which the predicted values ​​by model M1 match the actual measured values ​​is R 2These indexes include the coefficient of determination (COD), RMSE (Root Mean Square Error), MAE (Mean Absolute Error), etc. If it is determined that the prediction accuracy of the model M1 is insufficient based on these indexes, the prediction accuracy verification unit 142 goes back to checking the data again and adds data to the model M1 or makes corrections such as adding or reducing explanatory variables in order to alleviate the bias in the data.

[0050] The completed model M1 is saved so that it can be called up when predicting the target (unknown substance).

[0051] Various regression methods and algorithms for verifying prediction accuracy are provided in Scikit-learn, a library included in the above-mentioned Anaconda (registered trademark).

[0052] As shown in Fig. 2(a), the data table creation unit 15 creates learning data using the first explanatory variable, the second explanatory variable, and the objective variable obtained by the first acquisition unit 11, the second acquisition unit 12, and the third acquisition unit 13. An example of the learning data is shown in Fig. 4(a). As shown in Fig. 4(a), the data table creation unit 15 creates learning data consisting of the first explanatory variable, the second explanatory variable, and the objective variable.

[0053] As sample substances to be included in the training data, it is preferable to prepare at least several dozens of substances with different types and positional relationships of substituents, on the premise that the mechanism of elementary reactions is common. Basically, only relatively small molecular sizes that do not require high calculation costs are required. However, in reactions in which bulkiness directly affects the activation energy, it may be necessary to include sample substances with large molecular sizes and bulky substituents in the training data.

[0054] By determining the molecular structures of the reaction system (stable state before the reaction) and the product system (stable state after the reaction) of the sample substance by quantum chemical calculations, a first explanatory variable and a second explanatory variable are obtained as explanatory variables for learning.

[0055] Once structural optimization using quantum chemical calculations is complete, various information is output, including structural parameters such as bond lengths and bond angles, potential energy of the system, orbital energy, dipole moment, charge distribution, etc.

[0056] The data table creation unit 15 extracts information about each sample substance and adds it to the learning data table.

[0057] The data table creation unit 15 obtains learning objective variables including the activation energy value by determining the molecular structure of the reaction system and transition state of the sample substance through quantum chemical calculations, and adds the learning data to the learning data table.

[0058] It is preferable that the calculation conditions be as accurate as possible. For the transition state calculation, first, a suitable transition state structure is searched for using a sample with the simplest structure with few substituents, and the input structure of other samples is created using the obtained structure as a template, thereby minimizing the number of trial and error steps. After the calculation of the structure optimization is completed, the activation energy is calculated from the difference in potential energy between the reaction system and the transition state.

[0059] The first explanatory variable and the second explanatory variable are preferably calculated by a quantum chemical calculation having a lower calculation accuracy than the objective variable. That is, the quantum chemical calculation performed to obtain the first explanatory variable and the second explanatory variable needs to have the same calculation conditions, but does not need to be highly accurate compared to the case where the objective variable of the reaction system is calculated by the quantum chemical calculation, and can be low in accuracy.

[0060] In this case, it is preferable that the learning data table is prepared in CSV format, since CSV format data is versatile and can be handled by many spreadsheet programs.

[0061] As shown in FIG. 2, the display unit 16 can display information on the learning data used in learning the model M1, and information on the model M1.

[0062] In this embodiment, the learning device 1 does not necessarily have to include the prediction accuracy verification unit 142.

[0063] <Prediction device> A prediction device according to this embodiment will be described. Fig. 2(b) is a system configuration diagram showing the configuration of the prediction device according to this embodiment. As shown in Fig. 2(b), the prediction device 2 includes a first acquisition unit 21, a second acquisition unit 22, a trained model M2, and a prediction unit 23. The prediction device 2 may further include a prediction data table creation unit 24 and a display unit 25.

[0064] The first acquisition unit 21 acquires first explanatory variables including structural parameters and electronic properties of a reaction system calculated by quantum chemical calculations, similar to the first acquisition unit 11. The function of the first acquisition unit 21 is similar to that of the first acquisition unit 11, and therefore details will be omitted.

[0065] The second acquisition unit 22 acquires second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting a reaction system, which are calculated by quantum chemical calculations, similarly to the second acquisition unit 12. The function of the second acquisition unit 22 is similar to that of the second acquisition unit 12, and therefore details will be omitted.

[0066] As shown in Fig. 1, the trained model M2 is trained using training data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system are associated with a target variable including an activation energy in a chemical reaction of the predetermined reaction system. As the trained model M2, a model obtained by machine learning the model M1 included in the learning device 1 can be used. Specifically, as the trained model M2, a machine learning model saved using Anaconda or the like can be used.

[0067] The prediction unit 23 inputs the first explanatory variable and the second explanatory variable of the reaction system to be predicted, and outputs a target variable including the activation energy in the chemical reaction of the reaction system to be predicted, predicted by the trained model M2.

[0068] 5 is a block diagram showing a schematic configuration of the prediction unit 23. As shown in FIG.

[0069] The reading unit 231 reads out the first explanatory variable and the second explanatory variable stored in the prediction data table.

[0070] The prediction unit 23 calls the trained model M2, reads the first explanatory variable and the second explanatory variable stored in the prediction data table by the reading unit 231 into the trained model M2, and outputs a predicted value of the objective variable including the activation energy of the chemical reaction.

[0071] As shown in Fig. 2(b), the prediction data table creation unit 24 creates a prediction data table using the first explanatory variable and the second explanatory variable acquired by the first acquisition unit 21 and the second acquisition unit 22. An example of the prediction data table is shown in Fig. 4(b). As shown in Fig. 4(b), the prediction data table stores the first explanatory variable and the second explanatory variable. The first explanatory variable and the second explanatory variable are similar to the first explanatory variable and the second explanatory variable used in the learning device 1 described above, and therefore details will be omitted.

[0072] The prediction data table, like the learning data table, is preferably prepared in CSV format. CSV format data is versatile and can be handled by many spreadsheet programs.

[0073] The display unit 25 can display the objective variables including the activation energy in the chemical reaction of the reaction system to be predicted, which is predicted by the trained model M2.

[0074] The learning device 1 and the prediction device 2 shown in FIG. 2(a) and FIG. 2(b) may be configured as a learning system and a prediction system. That is, the learning device 1 is a single device such as a PC (Personal Computer) having each component within the device, but one or more of the components may be arranged outside the device and connected via a network. For example, the learning data table may be provided on a cloud. In this case, the learning device 1 is configured as a learning system by the learning data table connected via a network. Similarly, the prediction device 2 shown in FIG. 2(b) may also be configured as a learning system by the learning data table connected via a network, with one or more of the components being arranged outside the device. For example, the prediction data table may be provided on a cloud. In this case, the prediction device 2 is configured as a prediction system by the prediction data table connected via a network.

[0075] <Hardware configuration of learning device 1 and prediction device 2> Next, an example of the hardware configuration of the learning device 1 and the prediction device 2 will be described. FIG. 6 is a block diagram showing the hardware configuration of the learning device 1 and the prediction device 2. As shown in FIG. 6, the learning device 1 and the prediction device 2 are configured as an information processing device (computer), and can be physically configured as a computer system including a CPU (Central Processing Unit: processor) 101 which is an arithmetic processing unit, a RAM (Random Access Memory) 102 and a ROM (Read Only Memory) 103 which are main storage devices, an input device 104 which is an input device, an output device 105, a communication module 106, and an auxiliary storage device 107 such as a hard disk. These are connected to each other by a bus 108. The output device 105 and the auxiliary storage device 107 may be provided externally.

[0076] The CPU 101 controls the overall operation of the learning device 1 and the prediction device 2, and performs various information processing. The CPU 101 executes a learning program and a prediction program stored in the ROM 103 or the auxiliary storage device 107, and controls the display operation of the measurement recording screen and the analysis screen.

[0077] The RAM 102 is used as a work area for the CPU 101 and may include a non-volatile RAM for storing main control parameters and information.

[0078] The ROM 103 stores a basic input / output program, etc. The prediction program may be stored in the ROM 103.

[0079] The input device 104 is a keyboard, a mouse, an operation button, a touch panel, or the like.

[0080] The output device 105 is a monitor display, etc. The output device 105 displays prediction results, etc., and updates the screen in response to input / output operations via the input device 104 and the communication module 106.

[0081] The communication module 106 is a data transmission / reception device such as a network card, and functions as a communication interface that receives information from an external data recording server or the like, and outputs analysis information to other electronic devices.

[0082] The auxiliary storage device 107 is a storage device such as an SSD (Solid State Drive) or an HDD (Hard Disk Drive), and stores, for example, various types of data, files, and the like required for the operation of the learning device 1 and the prediction device 2.

[0083] 2, predetermined computer software (including a learning program and a prediction program) is loaded into a main storage device such as a CPU 101 or a RAM 102 or an auxiliary storage device 107, and the learning program and the prediction program stored in the RAM 102, the ROM 103 or the auxiliary storage device 107 are executed by the CPU 101. Each function of the learning device 1 and the prediction device 2 is realized by operating the input device 104, the output device 105, and the communication module 106 and reading and writing data from and to the RAM 102, the ROM 103, the auxiliary storage device 107, etc. That is, by executing the learning program and the prediction program according to this embodiment on a computer, the learning device 1 and the prediction device 2 function as each configuration of FIG. 2.

[0084] The learning program and the prediction program are stored, for example, in a storage device provided in the computer. Note that the learning program and the prediction program may be configured so that a part or all of them are transmitted via a transmission medium such as a communication line, and are received and recorded (including installed) by a communication module 106 provided in the computer. Also, the learning program and the prediction program may be configured so that a part or all of them are recorded (including installed) in the computer from a state in which they are stored in a portable storage medium such as a CD-ROM, DVD-ROM, or flash memory.

[0085] <Study Program> The learning program according to this embodiment can use a program having the following configuration. That is, the learning program according to this embodiment includes a first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; A second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; A third acquisition step of acquiring a target variable including an activation energy in a chemical reaction of the reaction system, the activation energy being calculated by a quantum chemical calculation; a learning process of learning a model using learning data in which the first explanatory variable, the second explanatory variable, and the objective variable are associated with each other; A program for causing a computer to execute at least the above can be used.

[0086] <Prediction Program> The prediction program according to this embodiment can use a program having the following configuration. That is, the prediction program according to the present embodiment includes a first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; A second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; a prediction step of inputting the first explanatory variable and the second explanatory variable of a reaction system to be predicted into a trained model trained using learning data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system correspond to a response variable including an activation energy in a chemical reaction of the predetermined reaction system, thereby outputting a response variable including an activation energy in a chemical reaction of the reaction system to be predicted, which is predicted by the trained model; A program for causing a computer to execute at least the above can be used.

[0087] <Learning Method> Next, a learning method according to the present embodiment will be described. The learning method according to the present embodiment is a method for learning a model M1 using learning data in which a first explanatory variable, a second explanatory variable, and a target variable are associated with each other in a learning device 1 having a configuration as shown in Fig. 2(a).

[0088] Fig. 7 is a flowchart for explaining the learning method according to the present embodiment. As shown in Fig. 7, the first acquiring unit 11 acquires the first explanatory variables including the structural parameters and electronic properties of the reaction system calculated by the quantum chemical calculation (first acquiring step: step S11).

[0089] Next, the second acquisition unit 12 acquires second explanatory variables calculated by quantum chemical calculations, the second explanatory variables including structural parameters and electronic properties of the product system generated by chemically reacting the reaction system (second acquisition process: step S12).

[0090] Next, the third acquiring unit 13 acquires a target variable including an activation energy in a chemical reaction of the reaction system, which is calculated by a quantum chemical calculation (third acquiring step: step S13).

[0091] Next, the data table creating unit 15 creates a learning data table including the first explanatory variable, the second explanatory variable, and the objective variable (learning data table creating process: step S14). The data table creating unit 15 links the input first explanatory variable, the second explanatory variable, and the objective variable to create a learning data table, for example, as shown in FIG. 4(a).

[0092] Next, the learning unit 14 learns the model M1 using the learning data in which the first and second explanatory variables are associated with the objective variable (learning step: step S15). The learning unit 14 learns the model M1 in response to the input of the first and second explanatory variables included in the learning data, so as to obtain an output that matches the objective variable linked to these explanatory variables.

[0093] Next, the display unit 16 displays information about the learning data used in learning the model M1 and information about the model M1 (display step: step S16).

[0094] The step of acquiring the first explanatory variable (step S11) may be performed simultaneously with the step of acquiring the second explanatory variable (step S12), or may be performed after the step of acquiring the second explanatory variable (step S12).

[0095] <Prediction method> Next, a prediction method according to the present embodiment will be described. The prediction method according to the present embodiment is a method for outputting a response variable including an activation energy in a chemical reaction of a reaction system to be predicted, which is predicted by a trained model M2 in a prediction device 2 having a configuration as shown in FIG. 2(b).

[0096] Fig. 8 is a flowchart for explaining the prediction method according to the present embodiment. As shown in Fig. 8, a first obtaining unit 21 obtains a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation (first obtaining step: step S21).

[0097] Next, the second acquisition unit 22 acquires second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, calculated by quantum chemical calculations (second acquisition process: step S22).

[0098] Next, a prediction data table including the first explanatory variable and the second explanatory variable is created by the prediction data table creating unit 24 (prediction data table creating process: step S23). The prediction data table creating unit 24 creates a prediction data table including the input first explanatory variable and second explanatory variable as shown in FIG.

[0099] Next, the prediction unit 23 inputs the first explanatory variable and the second explanatory variable of the reaction system to be predicted to the trained model M2, and outputs a response variable including the activation energy in the chemical reaction of the reaction system to be predicted, which is predicted by the trained model M2 (prediction step: step S24). The prediction unit 23 inputs the first explanatory variable and the second explanatory variable input to the prediction data table in the prediction data table creation step (step S23) to the trained model M2, and outputs the response variable including the activation energy output from the trained model M2.

[0100] Next, the display unit 25 displays information on the objective variables including the activation energy output by the prediction unit 23 (display step: step S25).

[0101] The step of acquiring the first explanatory variable (step S21) may be performed simultaneously with the step of acquiring the second explanatory variable (step S22), or may be performed after the step of acquiring the second explanatory variable (step S22).

[0102] The effect of this embodiment will be described. The learning device 1 according to this embodiment includes the learning unit 14, and can perform machine learning of the model M1 to calculate a value that is close to the objective variable calculated when performing a high-precision quantum chemical calculation from the first explanatory variable and the second explanatory variable. When determining the objective variable from the first explanatory variable and the second explanatory variable, generally, if a high-precision calculation condition is selected to perform a high-precision quantum chemical calculation, the objective variable can be calculated with high precision, but the burden on the calculation is large and a high calculation cost is required to obtain the calculation result. The learning device 1 can learn to predict the activation energy of a chemical reaction with high precision from the input first explanatory variable and second explanatory variable easily and at low calculation cost.

[0103] Furthermore, the learning device 1 can reduce the burden and time required for calculations. Therefore, the learning device 1 can learn to predict the activation energy of a chemical reaction simply and with low calculation cost from the input first explanatory variable and second explanatory variable.

[0104] Furthermore, in the learning device 1, the first explanatory variable and the second explanatory variable can be calculated by quantum chemical calculations with lower calculation accuracy than the dependent variable, which allows the prediction device 2 to reduce the calculation load required to predict the dependent variable.

[0105] Furthermore, the learning device 1 can include a display unit 16. This allows the prediction device 2 to visually present to the user information for generating learning data and information about the model M1, allowing the user to easily grasp the learning data and the information about the model M1.

[0106] The prediction device 2 according to this embodiment includes a trained model M2 and a prediction unit 23. As a result, when predicting a response variable from a first explanatory variable and a second explanatory variable, the prediction device 2 can predict an approximate or equivalent response variable when calculated from the first explanatory variable and the second explanatory variable by a high-precision quantum chemical calculation. Therefore, the prediction device 2 can predict the activation energy of a chemical reaction with high precision.

[0107] Furthermore, by including the trained model M2 and the prediction unit 23, the prediction device 2 can predict the objective variable from the first explanatory variable and the second explanatory variable more easily than when a high-precision quantum chemical calculation is performed using a high-precision computer, and can reduce the burden and time required for calculating the objective variable. Therefore, the prediction device 2 can predict the activation energy of a chemical reaction simply and with low calculation cost.

[0108] Here, activation energy is the energy required for a chemical reaction to proceed, and is an important factor for estimating the ease of progress of a chemical reaction in various fields such as the chemical industry and pharmaceuticals. In general, there are experimental methods, computational science methods, and the like for determining the activation energy of a chemical reaction. In the experimental method, for example, an experiment is performed multiple times at different temperatures to determine the rate constant by measuring the concentration change of a substance during the reaction, and the activation energy can be determined from the slope of the straight line in the Arrhenius plot. However, the experiment requires a lot of effort, and depending on the reaction, a straight line cannot be obtained, so that the activation energy may not be determined. Therefore, it is difficult to determine the activation energy correctly and reliably using the experimental method.

[0109] In the computational science method, the activation energy can be directly measured from the difference between the free energy of the starting material in a stable state and the free energy of the transition state, as shown in FIG. 1, by determining the structure of the transition state of the reaction by quantum chemical calculation. However, when the activation energy is directly determined by using quantum chemical calculation, it is difficult to specify the structure of the transition state, and it is not easy to calculate the transition state in quantum chemical calculation, so that it is difficult to predict the activation energy. Therefore, when searching for the structure of the transition state by quantum chemical calculation, the molecular structure of the substance in the transition state is estimated in advance. However, if the molecular structure of the substance cannot be estimated or if the estimated molecular structure is far from the true transition state structure, the activation energy cannot be correctly predicted. When the structure of the transition state is searched for by quantum chemical calculation based on the intuition or experience of the engineer, the value of the activation energy calculated by the engineer is likely to vary. In some cases, it may take a lot of time and effort to trial and error. In addition, when the activation energy is calculated by performing high-precision calculations using a highly reliable and highly accurate computer, the burden and time required for the calculation increases. In particular, the larger the molecular size of the reaction system or product system, the more time and high-spec computers are required, making it difficult to balance time and cost. Therefore, if the accuracy of the calculation is sacrificed in order to perform the calculation in a realistic calculation time and environment, only unreliable calculation results can be obtained, and as a result, the molecules that can be handled are limited to small sizes.

[0110] The prediction device 2 uses various information on the reaction system and the product system obtained by low-precision quantum chemical calculation as the first explanatory variable and the second explanatory variable. The various information on the reaction system and the product system obtained by the low-precision quantum chemical calculation is information on the stable structure of the reaction system and the product system, and can be easily and stably calculated by the low-precision quantum chemical calculation. As shown in FIG. 1, the prediction device 2 can predict an approximate or equivalent activation energy ΔG when a transition state is obtained by a high-precision quantum chemical calculation by applying the first explanatory variable and the second explanatory variable obtained by the low-precision quantum chemical calculation to the trained model M2. In addition, the prediction device 2 can reduce the calculation time and calculation load compared to the conventional high-precision quantum chemical calculation.

[0111] Therefore, the prediction device 2 can predict activation energy with high accuracy with simple and low calculation cost without being restricted by molecular size by combining low-precision quantum chemical calculation and machine learning, and can therefore predict the objective variable with high accuracy even for reaction systems with relatively large molecular sizes. Therefore, the prediction device 2 can be suitably used to speed up research and development, since it can quickly and accurately predict the activation energy of reaction systems with relatively large molecular sizes, such as materials and catalysts in the chemical industry.

[0112] The learning program according to the present embodiment includes a learning step, and thus can cause a computer to execute machine learning of the model M1. Therefore, by using the learning program according to the present embodiment, it is possible to learn so as to predict the activation energy in a chemical reaction with high accuracy based on the input first explanatory variable and second explanatory variable.

[0113] In addition, the learning program according to the present embodiment uses a first explanatory variable including the structural parameters and electronic properties of the reaction system and a second explanatory variable including the structural parameters and electronic properties of the product system, thereby making it possible to predict the objective variable more easily than in the case of using high-precision quantum chemical calculations, and to reduce the burden and time required for calculation. Therefore, the learning program according to the present embodiment can learn to predict the activation energy of a chemical reaction from the input first explanatory variable and second explanatory variable in a simple manner and at a low calculation cost.

[0114] The prediction program according to the present embodiment includes a prediction step, and can be executed by a computer to output a response variable to be predicted. Therefore, when the prediction program according to the present embodiment is used to predict a response variable from the first explanatory variable and the second explanatory variable, it is possible to predict an approximate or equivalent response variable when calculated from the first explanatory variable and the second explanatory variable by a high-precision quantum chemical calculation. Therefore, the prediction device 2 can predict the activation energy of a chemical reaction with high precision.

[0115] Furthermore, since the prediction program according to this embodiment includes a prediction step, it is possible to easily output a response variable, including the activation energy of a chemical reaction, in a short time without selecting highly accurate calculation conditions for the response variable from the first explanatory variables and the second explanatory variables and without performing highly accurate quantum chemical calculations, and therefore it is possible to cause a computer to easily predict a response variable including the activation energy of a chemical reaction with low calculation cost.

[0116] The learning method according to the present embodiment includes a learning step (step S13), and thus can perform machine learning of the model M1. Therefore, the learning method according to the present embodiment can learn to predict the activation energy in a chemical reaction with high accuracy based on the input first explanatory variable and second explanatory variable.

[0117] Furthermore, the learning method according to the present embodiment includes a learning step (step S13), and thus can predict the objective variable from the first explanatory variable and the second explanatory variable more easily than when using high-precision quantum chemical calculations, and can reduce the burden and time required for calculations. Therefore, the learning method according to the present embodiment can learn to predict the activation energy of a chemical reaction from the input first explanatory variable and second explanatory variable simply and at low calculation cost.

[0118] The prediction method according to the present embodiment includes a prediction step (step S23), and can output a predicted value of the response variable to be predicted. Therefore, when predicting a response variable from the first explanatory variable and the second explanatory variable, the prediction method according to the present embodiment can predict an approximate or equivalent response variable when calculated from the first explanatory variable and the second explanatory variable by a high-precision quantum chemical calculation. Therefore, the prediction method according to the present embodiment can predict the activation energy of a chemical reaction with high precision.

[0119] Furthermore, since the prediction method according to this embodiment includes a prediction step (step S23), it is possible to easily and quickly output the objective variables from the first explanatory variables and the second explanatory variables without selecting highly accurate calculation conditions and performing highly accurate quantum chemical calculations, and without increasing the burden and time required for calculating the objective variables. Therefore, it is possible to easily predict the objective variables including the activation energy of a chemical reaction with high accuracy at low calculation cost. EXAMPLES

[0120] The embodiments will be described in more detail below with reference to examples, but the embodiments are not limited to these examples and comparative examples.

[0121] <Example 1> [Machine learning environment] In this embodiment, Anaconda (registered trademark) was used for machine learning, numerical calculation, and data analysis.

[0122] [Target reaction] In this example, we will deal with the Diels-Alder reaction, an organic chemical reaction in which an alkene is added to a conjugated diene to produce a six-membered ring structure. The general formula for the Diels-Alder reaction is shown in the following formula (I). For simplicity, in this example, we will only consider the case where acyclic compounds react with each other.

[0123] [ka] (In the formula, TS represents a transition state, and X1 to X4 and Y1 to Y4 each independently represent any one of substituents such as -H, -NO2, -CN, -CF3, -SO2OH, -SO2CH3, -Ph, -CHO, -COOH, -COOCH3, -F, -Cl, -Br, -C(CH3)3, -CH3, -OCH3, and -OH.)

[0124] [Create Model] (Acquiring training data and creating a training data table) For sample compounds with different combinations of substituents X1 to X4 and Y1 to Y4, information obtained from quantum chemical calculations was used as learning data in this embodiment, and 94 patterns No. 1 to No. 94 shown in Tables 1 to 3 were used. In order to distribute the activation energy value, which is the objective variable, compounds with a combination of 17 types of electron-donating and electron-withdrawing substituents (-H, -NO2, -CN, -CF3, -SO2OH, -SO2CH3, -Ph, -CHO, -COOH, -COOCH3, -F, -Cl, -Br, -C(CH3)3, -CH3, -OCH3, -OH) were targeted. In order to reduce calculation costs and human labor, substituents with large molecular size and large degree of orientation freedom were avoided as much as possible.

[0125] [Table 1]

[0126] [Table 2]

[0127] [Table 3]

[0128] For the 94 patterns shown in Tables 1 to 3, No. 1 to No. 94, structural optimization calculations were performed using the Gaussian16 program by the DFT method. The calculation conditions and calculation targets for the first explanatory variable, the second explanatory variable, and the objective variable are shown in Table 4. The first explanatory variable and the second explanatory variable were obtained using the low-precision basis 3-21G. * ,To obtain the objective variable data, we used the high ,precision basis 6-311+G(2d,p).

[0129] [Table 4]

[0130] From the output calculation results, data for the first explanatory variable, the second explanatory variable, and the objective variable were extracted. The objective variable, activation energy, was calculated according to the following formula (2). Activation energy = free energy of the transition state – (free energy of reactant 1 + free energy of reactant 2) (2)

[0131] As the first and second explanatory variables, data on a total of 54 types shown in Table 5 (27 variables derived from reactants 1 and 2, 26 variables derived from products, and one variable for the stabilization energy of the product system) were extracted, and a learning data table was created together with the objective variables.

[0132] [Table 5]

[0133] The carbon atom labels C1 to C6 are defined in accordance with the following reaction formula (II).

[0134] [ka]

[0135] [Check the training data] The distribution of the objective variable values ​​stored in the created training data table is shown in Figure 9. As shown in Figure 9, the range of activation energy is from 27 kcal / mol to 43 kcal / mol, and since there is continuous data within this range, it can be said that there is generally no problem.

[0136] Moreover, the correlation coefficients between the first explanatory variable and the objective variable and the second explanatory variable are shown in Tables 6 and 7 in descending order of the absolute value of the correlation coefficient.

[0137] [Table 6]

[0138] [Table 7]

[0139] From Tables 6 and 7, it was confirmed that the stabilization energy derived from reactant 2 (alkene), the product, and the stabilization energy of the product system are explanatory variables having a relatively strong correlation with the objective variable.

[0140] (Creating a learning model and verifying prediction accuracy) We carried out regression analysis and verified the prediction accuracy of the prepared learning data. We used Bayesian ridge regression as the regression method and leave-one-out cross-validation as the accuracy verification method. In addition, we found that the best prediction accuracy (R 2 We adopted a model that used the top 52 correlation variables (i.e., excluding the bottom two correlation variables) that had the largest correlation coefficient (evaluated by correlation coefficient value).

[0141] The results of predicting and verifying the learning data using the created learning model are shown in Figure 10. As shown in Figure 10, although there is some variation depending on the sample, the accuracy of the learning model R 2The value was 0.834, and the MAE (mean absolute error) was 1.1 kcal / mol. It was confirmed that Example 1 was able to provide activation energy with an accuracy comparable to that of directly determining activation energy using high-precision quantum chemical calculations for the training data.

[0142] [Prediction of activation energy of the predicted target (unknown substance)] Using the completed trained model, we predicted the activation energy when an unknown substance that is the prediction target and not included in the training data reacts. The prediction results are shown in Table 8. Table 8 also shows five patterns of prediction targets A to E that were used as prediction targets. In Table 8, 1-Ad is a 1-adamantyl group (-C 10 H 15 ), and TBS stands for t-butyldimethylsilyl group (-Si(CH3)2C(CH3)3).

[0143] [Table 8]

[0144] Low-precision quantum chemical calculations were performed for five prediction targets A to E. The calculation conditions and calculation targets are shown in Table 9. From the output calculation results, the first explanatory variable and the second explanatory variable were extracted, and a prediction data table was created.

[0145] [Table 9]

[0146] The prediction data table we created was used as input, and the predicted value of the activation energy was output as the objective variable using the trained model. The output results are shown in Table 10.

[0147] [Table 10]

[0148] <Verification of prediction results for the prediction target> The above prediction results were compared with the results of the activation energy actually calculated by high-precision quantum chemical calculations to verify the prediction accuracy of the activation energy predicted using the trained model. The calculation conditions and calculation targets for the verification are shown in Table 11.

[0149] [Table 11]

[0150] The predicted values ​​of the activation energies and the calculated values ​​of the activation energies obtained are shown in Table 12. Moreover, a diagram in which the predicted values ​​and the calculated values ​​are plotted together with the learning data is shown in FIG.

[0151] [Table 12]

[0152] As shown in Table 12 and Figure 11, it was confirmed that the predicted value of the activation energy predicted using the trained model created in this embodiment was able to predict roughly the same result as the calculated value obtained by high-precision quantum chemical calculation, even for untrained data. Therefore, it can be said that even for an unknown substance to be predicted, if the first explanatory variable and the second explanatory variable obtained by low-precision quantum chemical calculation of untrained data are applied to the trained model, it is possible to predict an activation energy equivalent to or close to the activation energy obtained by high-precision quantum chemical calculation.

[0153] In addition, the calculation cost, which is a problem when directly obtaining activation energy by quantum chemical calculation, was compared with the prediction device method of this embodiment. In general, the calculation time depends on the number of atoms and electrons constituting the target molecule, more precisely, the number of basis functions expressing each electronic state. For the five prediction targets A to E for which data was obtained in this embodiment, the number of atoms, the number of electrons, the number of basis functions, and the time required for calculation (actual results) are shown in Table 13. In addition, a diagram plotting the calculation time against the number of electrons in the system is shown in FIG. 12. The required time is the sum of the time required for optimizing the structures of reactant 1, reactant 2, and the product, and the time required for vibration calculation. The number of computer cores was 16 in both cases.

[0154] In addition, as Comparative Example 1, a conventional method was performed in which the activation energy was directly calculated by calculating the free energy of the reaction system and the free energy of the transition state by high-precision quantum chemical calculation. In Comparative Example 1, the required time was the sum of the time required for optimizing the structures of reactant 1, reactant 2, and the transition state, and the time required for vibration calculation.

[0155] [Table 13]

[0156] As shown in Table 13 and FIG. 12, the time required for calculation in Example 1 was shorter than that in Comparative Example 1. In particular, the larger the molecular size, the greater the effect of shortening the time required for calculation. Therefore, it was confirmed that the conventional method of directly calculating activation energy by high-precision quantum chemical calculation requires more calculation time as the molecular size of the system being handled increases. In contrast, it was confirmed that the prediction device according to this embodiment can significantly shorten the calculation time because it is sufficient to apply the first explanatory variable and the second explanatory variable calculated by low-precision calculation to the trained model. In addition, it was confirmed that the calculation time is prevented from increasing even when the molecular size increases, and a significant effect of shortening the calculation time is obtained.

[0157] FIG. 13 shows the distribution of the number of electrons in the reaction system for 94 patterns of learning data and five prediction targets A to E. As shown in FIG. 13, the number of molecules in the reaction system was small in the learning data, whereas the number of molecules in the reaction system was mostly large in the prediction targets A to E. That is, it can be seen that the data group acquired when creating the trained model is contained in a relatively small molecular size region, whereas the prediction targets A to E are distributed in a region with a larger molecular size than the trained data. Therefore, it can be said that the trained model created in this embodiment can accurately predict the activation energy of the prediction target that exists in the untrained region with a larger molecular size than the training data.

[0158] Therefore, by using the trained model created in this embodiment, it is possible to accurately predict the activation energy of a reaction system that exists in a region of large molecular size, where high-precision quantum chemical calculations are difficult, by using small-sized samples that impose relatively little computational cost and time when obtaining training data.

[0159] Although the embodiment has been described above, the above embodiment is presented as an example, and the present invention is not limited to the above embodiment. The above embodiment can be implemented in various other forms, and various combinations, omissions, substitutions, modifications, etc. can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0160] 1 Learning device 2. Prediction Device 11, 21 First Acquisition Section 12, 22 Second Acquisition Section 13 Third Acquisition Section 14 Learning Department 142 Prediction Accuracy Verification Department 15 Data table creation section 16, 25 Display section 23 Prediction Department 24. Prediction data table creation section M1 Model M2 pre-trained model

Claims

1. A first acquisition unit that acquires a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; A second acquisition unit that acquires second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; A third acquisition unit that acquires a target variable including an activation energy in a chemical reaction of the reaction system, the activation energy being calculated by a quantum chemical calculation; a learning unit that performs model learning using learning data in which the first explanatory variable, the second explanatory variable, and the objective variable are associated with each other; A learning device having the above configuration.

2. The learning device according to claim 1 , wherein the first explanatory variable and the second explanatory variable are calculated by a quantum chemical calculation having a lower calculation accuracy than the target variable.

3. The learning device according to claim 1 , further comprising a display unit that displays information for generating learning data used in learning the model and information about the model to be learned.

4. A first acquisition unit that acquires a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; A second acquisition unit that acquires second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; a trained model trained using training data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system are associated with a target variable including an activation energy in a chemical reaction of the predetermined reaction system; a prediction unit that receives a first explanatory variable and a second explanatory variable of a reaction system to be predicted, and outputs a target variable including an activation energy in a chemical reaction of the reaction system to be predicted, predicted by the trained model; A prediction device having the above configuration.

5. A reception unit that receives a selection of information regarding the reaction system of the prediction target input to the trained model and information regarding a product system generated by chemically reacting the reaction system of the prediction target; a display unit that displays a response variable including an activation energy in a chemical reaction of the reaction system to be predicted; The prediction device according to claim 4 , further comprising:

6. A first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; A third acquisition step of acquiring a target variable including an activation energy in a chemical reaction of the reaction system, the activation energy being calculated by a quantum chemical calculation; a learning process of learning a model using learning data in which the first explanatory variable, the second explanatory variable, and the objective variable are associated with each other; A learning program that causes a computer to execute the following:

7. A first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; a prediction step of inputting the first explanatory variable and the second explanatory variable of a reaction system to be predicted into a trained model trained using learning data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system correspond to a response variable including an activation energy in a chemical reaction of the predetermined reaction system, thereby outputting a response variable including an activation energy in a chemical reaction of the reaction system to be predicted, which is predicted by the trained model; A prediction program that causes a computer to execute the above.

8. A first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; A third acquisition step of acquiring a target variable including an activation energy in a chemical reaction of the reaction system, the activation energy being calculated by a quantum chemical calculation; a learning process of learning a model using learning data in which the first explanatory variable, the second explanatory variable, and the objective variable are associated with each other; The computer executes the learning method.

9. A first acquisition step of acquiring a first explanatory variable including a structural parameter and an electronic property of a reaction system calculated by a quantum chemical calculation; a second acquisition step of acquiring second explanatory variables including structural parameters and electronic properties of a product system generated by chemically reacting the reaction system, the second explanatory variables being calculated by quantum chemical calculation; a prediction step of inputting the first explanatory variable and the second explanatory variable of a reaction system to be predicted into a trained model trained using learning data in which a first explanatory variable and a second explanatory variable calculated in advance for a predetermined reaction system correspond to a response variable including an activation energy in a chemical reaction of the predetermined reaction system, thereby outputting a response variable including an activation energy in a chemical reaction of the reaction system to be predicted, which is predicted by the trained model; A prediction method implemented by a computer.

Images