Learning method for organic compound design model, and organic compound design system

Abstract

Provided are a learning method for an organic compound design model, in which a large language model is used, and an organic compound design system. The present invention comprises: a step for generating, through use of a first large language model, a first organic compound group by using a user sentence; a step for calculating a physical property value of a first organic compound of the first organic compound group by scientific calculation; a step for evaluating a demand achievement degree from a correlation between the user sentence, the molecular structure of the first organic compound, and the physical property value through use of a second large-scale language model; a step for generating training data in which the user sentence, the molecular structure of the first organic compound, and the demand achievement degree are combined; a step for constructing a reward model by using the training data; a step for creating a pseudo user sentence; a step for generating, through use of the first large language model, a second organic compound group by using the pseudo user sentence; a step for generating pseudo reward data by evaluating the molecular structure of a second organic compound of the second organic compound group through use of the reward model; and a step for updating a parameter of the first large language model by using the pseudo reward data.

Description

Learning method for organic compound design models and organic compound design system

[0001] One aspect of the present invention relates to a method and system for designing novel organic compounds that can predict desired physical properties.

[0002] In research or development using organic compounds, organic compounds with corresponding physical properties are selected according to the required characteristics. However, the physical properties of an organic compound can only be known by synthesizing the target substance and directly measuring them. Therefore, when synthesizing a new organic compound, it is necessary to predict from known substances whether the organic compound will have the desired physical properties. Furthermore, the physical properties of an organic compound are determined by the molecular structure of the organic compound, but because there are so many candidate organic compounds, it has not been easy to select the optimal organic compound.

[0003] Furthermore, it is possible to predict the approximate values ​​of the physical properties of an organic compound having a certain molecular structure by using accumulated data or by performing scientific calculations such as first-principles calculations and simulations. However, accurate predictions are difficult even for those skilled in the art, and currently, simulations require enormous costs and time.

[0004] In recent years, methods for classification, estimation, and prediction using machine learning and other methods have made great advances. In particular, the performance of sorting and prediction using deep learning with convolutional neural networks has improved significantly, achieving excellent results in a variety of fields.

[0005] In particular, large-scale language models (LLMs) have attracted attention. Large-scale language models are natural language processing models trained using large amounts of data. The large-scale language models can realize, for example, dialogue models that respond to user instructions. Non-Patent Document 1 discloses GPT-4 (Generative Pre-trained Transformer 4) (registered trademark) as a large-scale language model, and also discloses ChatGPT as a dialogue model.

[0006] However, in the field of organic compounds, there are still few sufficient methods for describing organic compounds that can accurately extract properties-related characteristics while allowing a large-scale language model to accurately understand the molecular structure. Therefore, a design method and system that allows anyone to easily and accurately predict organic compounds with desired properties has not yet been realized.

[0007] Patent Document 1 and Non-Patent Document 2 disclose a method and an apparatus for searching for new substances using machine learning.

[0008] JP 2017-91526 A

[0009] Summary of ChatGPT / GPT-4 Research and Perspective Towards the Future of Large Language Models, Yiheng Liu et al. (Submitted on 4 April 2023, [online], Internet <URL: https: / / arxiv.org / abs / 2304.01852> Takahiro Ohkoshi, Masaki Ohkoshi, Yoshiaki Nagao, Satoshi Yotsuhashi, New Structure Discovery of Optically Functional Organic Materials Using High-Throughput Computation and Machine Learning, Panasonic Technical Journal Vol. 67 No. 2 Nov. 2021 p. 116

[0010] An object of one embodiment of the present invention is to provide a method for designing an organic compound, which allows anyone to easily and accurately design an organic compound having desired physical properties, whether known or unknown. Another object of one embodiment of the present invention is to provide a system for designing an organic compound, which allows anyone to easily and accurately design an organic compound having desired physical properties, whether known or unknown. Another object of one embodiment of the present invention is to provide a method for generating training data that can be used to design an organic compound having desired physical properties or to predict the physical properties of an organic compound. Another object of one embodiment of the present invention is to significantly improve the development speed by using the method or system for designing an organic compound.

[0011] Note that the description of these problems does not preclude the existence of other problems. Note that one embodiment of the present invention does not necessarily solve all of these problems. Note that problems other than these will become apparent from the description of the specification, drawings, claims, etc., and it is possible to extract other problems from the description of the specification, drawings, claims, etc.

[0012] One aspect of the present invention is a method for learning an organic compound design model, comprising: a step of generating a first group of organic compounds using user text with a first large-scale language model; a step of calculating physical property values ​​of the first organic compound contained in the first group of organic compounds by scientific calculation; a step of evaluating a degree of request fulfillment from a correlation between the user text, the molecular structure of the first organic compound, and the physical property values ​​with a second large-scale language model; a step of generating a plurality of training data combining the user text, the molecular structure of the first organic compound, and the degree of request fulfillment; a step of constructing a reward model using the training data; a step of creating pseudo-user text; a step of generating a second group of organic compounds using the pseudo-user text with the first large-scale language model; a step of generating pseudo-reward data by evaluating the molecular structure of the second organic compound contained in the second group of organic compounds with the reward model; and a step of updating parameters of the first large-scale language model using the pseudo-reward data.

[0013] One aspect of the present invention is a method for learning an organic compound design model, comprising: a step of generating a first group of organic compounds using a user text with a first large-scale language model; a step of calculating physical property values ​​of the first organic compounds in the first group of organic compounds through scientific calculations; a step of evaluating a degree of desired achievement based on a correlation between the user text, the molecular structure of the first organic compound, and the physical property values ​​with a second large-scale language model; a step of generating a plurality of learning data combining the user text, the molecular structure, the physical property values, and the desired achievement degree of the first organic compound; and a step of updating parameters of the first large-scale language model using the learning data.

[0014] In the above, the physical property value is a learning method for a design model of an organic compound at the T1 level.

[0015] In the above, the reward model is a learning method for an organic compound design model that predicts the degree of pseudo-request achievement from the pseudo-user sentence and the molecular structure of the second organic compound.

[0016] One aspect of the present invention is an organic compound design system having an input unit, a calculation unit, and an output unit, wherein the input unit has a function of outputting input user text to the calculation unit, the calculation unit has a function of generating a first group of organic compounds using the user text using a first large-scale language model, a function of calculating physical property values ​​of the first organic compound in the first group of organic compounds using scientific calculations, a function of evaluating a degree of fulfillment of requirements from a correlation between the user text, the molecular structure of the first organic compound, and the physical property values ​​using a second large-scale language model, and a function of listing the molecular structures of the organic compounds and the physical property values ​​of the molecular structures in descending order of degree of fulfillment of requirements using the second large-scale language model, and the output unit has a function of outputting the listed molecular structures and physical property values.

[0017] The organic compound design system has a function of generating learning data that combines user text, the molecular structure of the first organic compound, physical property values, and the degree of desired achievement.

[0018] The above-mentioned organic compound design system has a function of generating training data that combines user text, the molecular structure, physical property values, and desired achievement level of the first organic compound, and a function of updating parameters of the first large-scale language model using the training data.

[0019] In the above, the physical property value is a design system of an organic compound at the T1 level.

[0020] In the above-described configuration, the physical property values ​​of the organic compound are one or more of an emission spectrum, a half-width, an emission energy, an excitation spectrum, an absorption spectrum, a transmission spectrum, a reflection spectrum, a molar absorption coefficient, an excitation energy, a transient emission lifetime, a transient absorption lifetime, an S1 level, a T1 level, an Sn level, a Tn level, a Stokes shift value, an emission quantum yield, an oscillator strength, an oxidation potential, a reduction potential, a HOMO level, a LUMO level, a glass transition point, a melting point, a crystallization temperature, a decomposition temperature, a boiling point, a sublimation temperature, a carrier mobility, a refractive index, an orientation parameter, a mass-to-charge ratio, a spectrum in an NMR measurement, a chemical shift value and the number of elements or a coupling constant thereof, a spectrum in an ESR measurement, a g factor, a D value, or an E value.

[0021] One embodiment of the present invention provides a method for designing an organic compound, which allows anyone to easily and accurately design an organic compound having desired physical properties, regardless of whether the compound is unknown or known. Another embodiment of the present invention provides a system for designing an organic compound, which allows anyone to design an organic compound having desired physical properties, regardless of whether the compound is unknown or known. Another embodiment of the present invention provides a method for generating training data that can be applied to the organic compound design method and the organic compound design system. Use of these organic compound design methods or systems facilitates the selection of organic compounds having desired physical properties, significantly improving the development speed.

[0022] The effects of one embodiment of the present invention are not limited to the effects listed above. The effects listed above do not preclude the existence of other effects. Note that the other effects are effects not mentioned in this section, which will be described below. Effects not mentioned in this section can be derived by a person skilled in the art from the description in the specification, drawings, etc., and can be extracted as appropriate from these descriptions. Note that one embodiment of the present invention has at least one of the effects listed above and / or other effects. Therefore, one embodiment of the present invention may not have the effects listed above in some cases.

[0023] FIG. 1 is a conceptual diagram illustrating a method for designing an organic compound. FIG. 2 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 3 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 4 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 5 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 6 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 7 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 8 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 9 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 10 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 11 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 12 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 13 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 14 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 15 is a diagram illustrating a flowchart of the method for designing an organic compound. FIG. 16 is a diagram illustrating a flowchart of the method for designing an organic compound. FIGS. 17A and 17B are diagrams illustrating the configuration of a neural network. FIG. 18 is a diagram illustrating a method for converting molecular structures using fingerprinting. Figures 19A to 19D are diagrams explaining types of fingerprinting methods. Figure 20 is a diagram explaining the process of converting from SMILES notation to notation using the fingerprinting method. Figure 21 is a diagram explaining types of fingerprinting methods and overlapping notations. Figures 22A and 22B are diagrams explaining an example in which a molecular structure is notated using multiple fingerprinting methods. Figure 23 is a diagram showing a property prediction system.

[0024] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be readily understood by those skilled in the art that various changes in form and details can be made without departing from the spirit and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the description of the embodiments shown below.

[0025] (Embodiment 1) One aspect of the present invention is a method for designing an organic compound using an organic compound design system having a system utilization phase in which a user utilizes the system and a learning phase in which a large-scale language model for generating molecules is learned by reinforcement learning.

[0026] <Configuration Example of Design System> An organic compound design system according to one embodiment of the present invention will be described below with reference to a data model 10, which is a conceptual diagram shown in Fig. 1. The data model 10 includes a main system 11, a large-scale language model for molecule generation 12, a training data database 13, a property calculation system 14, a large-scale language model for request achievement evaluation 15, a learning management system 16, a reward prediction system 17, and a user 19.

[0027] Note that, excluding the user 19, the main system 11, the large-scale language model for molecule generation 12, the training data database 13, the physical property calculation system 14, the large-scale language model for request achievement evaluation 15, the learning management system 16, and the reward prediction system 17 may be incorporated into a single device as long as they are capable of exchanging data. Alternatively, they may be different devices, or some may be incorporated into the same device. Furthermore, the data server may be a cloud, but these will be collectively referred to as the organic compound design system.

[0028] The main system 11 outputs to the user 19 information such as the molecular structure of an organic compound or the physical property values ​​of the organic compound according to the data input by the user 19. The main system 11 also exchanges data with the large-scale language model for molecule generation 12, the training data database 13, the physical property calculation system 14, or the large-scale language model for request achievement evaluation 15, and performs processing as necessary.

[0029] The large-scale language model for molecule generation 12 generates molecular structures in response to data input by a user 19 via the main system 11. It also generates molecular structures in response to data output from a learning management system 16.

[0030] The learning data database 13 stores data that combines data input by the user 19 via the main system 11, molecular structures associated with the data, and physical property values ​​of the molecular structures calculated by scientific calculations, as well as the degree of fulfillment of the requirements for the combined data. The degree of fulfillment of the requirements refers to the degree to which the molecular structure generated based on the data input by the user 19 and the calculated physical property values ​​of the molecular structure fulfill the requirements of the user 19.

[0031] The physical property calculation system 14 calculates the physical property values ​​of the molecular structure in accordance with the molecular structure generated by the large-scale language model for molecule generation 12 .

[0032] The large-scale language model for request fulfillment evaluation 15 compares the request input by the user 19 with the calculation results calculated by the physical property calculation system, and evaluates whether the calculation results satisfy the request.

[0033] The learning management system 16 creates pseudo-user sentences. Using the pseudo-user sentences, the large-scale language model for molecule generation 12 generates molecular structures according to the pseudo-user sentences, and inputs the molecular structures to the reward prediction system 17.

[0034] Reward prediction system 17 calculates a predicted value of the degree of wish fulfillment for the molecular structure generated according to the pseudo-user sentence. The predicted value of the degree of wish fulfillment is reflected in large-scale language model for molecule generation 12, and the parameters of large-scale language model for molecule generation 12 are updated. Furthermore, reward prediction system 17 constructs a reward model based on data combining data input by user 19 and molecular structures associated with the data, and the degree of wish fulfillment for the combined data.

[0035] By using one embodiment of the present invention, a novel organic compound can be designed in accordance with a sentence input by a user.

[0036] <Method for Designing Organic Compounds> A method for designing organic compounds according to one embodiment of the present invention will be described below using a flowchart. Note that the alphabets attached to connectors in the flowcharts are connected to connectors formed by adding an apostrophe (') to the alphabet.

[0037] <<User System Use Phase (Flowchart F100)>> In the organic compound design system of one embodiment of the present invention, the user system use phase can be represented, for example, by the flowchart F100 shown in FIGS. 2 and 3, focusing on the main system.

[0038] First, a user creates a sentence (user sentence) according to a desired organic compound, and inputs the created user sentence into the organic compound design system (step S101). Specifically, as an example of the user sentence, a sentence including the physical properties of the organic compound may be created, such as "Make a large amount of molecules whose T1 level is 2.6 eV or more and whose HOMO level is -5.1 eV or more."

[0039] Note that, examples of the physical property values ​​of an organic compound include energy levels such as the lowest triplet excitation energy level (T1 level), the lowest singlet excitation energy level (S1 level), the HOMO (Highest Occupied Molecular Orbital) level, and the LUMO (Lowest Unoccupied Molecular Orbital) level, as well as differences between the energy levels such as the energy level difference in the excited state, and rearrangement energy (λ).

[0040] Next, a large-scale language model for molecule generation is used to generate a group of organic compounds in a state where the chemical structures of the molecules are expressed as character strings, according to the user's text. Specifically, examples of notation in which the chemical structures of the molecules are expressed as character strings include SMILES (Simplified molecular input line entry specification syntax) notation, SMARTS notation, SMIRKS notation, InChI notation, WLN (Wiswesser Line Notation) notation, ROSDAL notation, and SLN notation (Tripos).

[0041] In the flowchart F100, for the sake of simplicity, a case where the SMILES notation is used for character string generation will be specifically described. That is, a molecular SMILES group is generated (step S102). A molecular SMILES is a molecular structure written in the SMILES notation, and a molecular SMILES group is a set of a plurality of molecular SMILES.

[0042] Next, an empty reward data list is created (step S103). Note that the creation of an empty information data list may be created in advance regardless of the timing. Note that the reward data is data that combines the user's sentence, a molecular SMILES generated according to the user's sentence, and the degree of fulfillment of the request of the molecular SMILES.

[0043] Here, in the molecular SMILES group obtained in step S102, physical property values ​​are calculated from each molecular SMILES, and it is determined whether a process (hereinafter also referred to as evaluation process) has been performed to evaluate the physical property values ​​by assigning a degree to which the user's needs are satisfied (degree of fulfillment of needs) (step D101).

[0044] In the main system, if there is any molecular SMILES that has not been evaluated in the molecular SMILES group obtained in step S102, the unprocessed molecular SMILES is extracted from the molecular SMILES group and converted into a format that can be used to calculate physical properties (step S106). Specifically, it is preferable to convert the molecular SMILES into a molecular structural formula, MOL file, GJF file (Gaussian), PDB file, or the like. In the flowchart F100, for the sake of simplicity, the case of conversion into a molecular structural formula will be described.

[0045] Next, the physical property values ​​of the molecular structural formula converted from the molecular SMILES are calculated using a physical property calculation system (step S107). Note that the method of calculating physical property values ​​by scientific calculation using the physical property calculation system will be described later.

[0046] Next, in the main system, the physical property values ​​calculated by the physical property calculation system are documented to create a physical property document (step S108). Here, the documented physical property values ​​are referred to as a physical property document. Alternatively, the documentation of the physical property values ​​may be processed by the physical property calculation system and output as documented information from the physical property calculation system.

[0047] The physical property document may be in any format as long as the names of physical property values ​​are linked to their numerical values. Specifically, an example of the physical property document may be a list format such as "HOMO level: -5.382 eV, LUMO level: -1.951 eV, Bandgap: 3.431 eV, T1: 2.3329999999987194 eV, λ(S0): 0.227 eV, λ(T1): 0.227 eV, ...", or may be written using the syntax of an appropriate language.

[0048] Next, information combining the physical property document and the user's text and an instruction to the effect of "evaluate the degree of requirement fulfillment of the physical property document relative to the user's text" are input to a large-scale language model for requirement fulfillment evaluation. The large-scale language model for requirement fulfillment evaluation evaluates the degree of requirement fulfillment for the input information combining the physical property document and the user's text (step S109).

[0049] Specifically, for a molecular structure generated based on the user's request, "Make a large amount of molecules with a T1 of 2.6 eV or more and a HOMO of -5.1 eV or more," the physical property values ​​are calculated as "HOMO level: -5.382 eV, LUMO level: -1.951 eV, Bandgap: 3.431 eV, T1: 2.3329999999987194 eV, λ(S0): 0.227 eV, λ(T1): 0.227 eV, ...," the user may instruct, as an example of input to the large-scale language model for request fulfillment evaluation, "Compare the calculation results with the following request, and rate on a scale of 100 whether the calculation results meet the request."

[0050] Next, information combining the numerator SMILES and the evaluated degree of request achievement is added to the reward data list (step S110).

[0051] The evaluation process shown in steps S106 to S110 is performed on unprocessed molecular SMILES in the molecular SMILES group obtained in step S102.

[0052] On the other hand, if all molecular SMILES in the molecular SMILES group obtained in step S102 have been evaluated, a combination (also called a pair) of information integrating the user sentence in the reward data list, the molecular structure generated according to the user sentence, and the physical property values ​​of the molecular structure, and the degree of request achievement is stored in the learning data database (step S104).

[0053] Next, each molecular SMILES of the molecular SMILES group is listed in descending order of the degree of wish fulfillment, and the molecular SMILES and the degree of wish fulfillment of the molecular SMILES are output and presented to the user (step S105). Note that it is not necessary to output the entire list of the molecular SMILES group, and it is advisable to extract and output an appropriate amount of information, such as n items in descending order of the degree of wish fulfillment, according to the user's requirements.

[0054] In this way, the user can obtain molecular SMILES of organic compounds that are highly likely to have the desired physical property values.

[0055] <Learning Phase 1 by Reinforcement Learning (Flowchart F120)> In the organic compound design system of one embodiment of the present invention, the learning phase of the large-scale language model for molecule generation by reinforcement learning can be represented, for example, by the flowchart F120 shown in FIGS. 4 and 5 , with the learning management system at the center.

[0056] The learning phase described in this section uses a reward model to optimize a large-scale language model for molecule generation using a large amount of pseudo-reward data.

[0057] First, the learning management system acquires learning data from the learning data database, which is a combination of a user sentence, a molecular SMILES generated according to the user sentence, and physical property values ​​calculated by scientific calculation of the molecular SMILES, as well as the degree of fulfillment of the request for the combined data (step S121).

[0058] Next, a reward model is constructed using the training data and a reward prediction system (step S122). The reward model pseudo-evaluates the molecular generation ability for unknown requests and can obtain a pseudo degree of request fulfillment (pseudo-degree of request fulfillment). In other words, molecular SMILES (or information on molecular structure represented by molecular structural formulas, etc.) predicted to have a high degree of request fulfillment can be output for a new input sentence (unknown request), so the process of calculating physical property values ​​using a physical property calculation system and the evaluation process using a large-scale language model for request fulfillment evaluation can be omitted.

[0059] Specifically, as an example of the reward model, it is preferable to set "y = f(x) is learned, x = user sentence + numerator SMILES, y = pseudo-request achievement degree."

[0060] Next, a plurality of pseudo-user sentences are created using the request template, and a pseudo-user sentence list is created by listing them (step S123).

[0061] Specifically, an example of a request template would be "Make five molecules with a T1 level of {T1} {below or above} and a HOMO level of {HOMO} {below or above}," where {T1}, {HOMO}, and {below or above} are variables.

[0062] Here, it is determined whether or not a process for generating a molecular SMILES group has been performed for each pseudo-user sentence in the pseudo-user sentence list obtained in step S123 (step D121).

[0063] If there is a pseudo-user sentence for which a molecular SMILES group has not been generated according to the pseudo-user sentence, that is, if there is an unprocessed pseudo-user sentence, a large-scale language model for molecule generation is used to generate a molecular generation SMILES group according to the pseudo-user sentence (step S124).

[0064] Next, it is determined whether or not a pseudo-desire achievement level is assigned to each combination of the molecular SMILES and the pseudo-user sentence in the molecular SMILES group generated in step S124 (step D122).

[0065] If a pseudo-desired achievement level has not been assigned to each combination of molecular SMILES and pseudo-user sentence, that is, if there is a molecular SMILES that has not been processed, the reward prediction system uses a reward model to predict a pseudo-desired achievement level for the combination of the unprocessed molecular SMILES and pseudo-user sentence (step S125).

[0066] Next, the parameters of the large-scale language model for generating molecules are updated according to the predicted value of the pseudo-desire achievement degree predicted in step S125 (step S126). The processes of steps S125 and S126 are performed for all the molecule SMILES in the molecule SMILES group obtained in step S124.

[0067] On the other hand, if a pseudo-desire achievement level has been assigned to all combinations of molecular SMILES and pseudo-user sentences, that is, if all molecular SMILES have been processed, it is determined again whether the process of generating molecular SMILES groups according to each pseudo-user sentence has been performed in the pseudo-user sentence list obtained in step S123 (step D121).

[0068] Here, when molecular SMILES groups have been generated according to all pseudo-user sentences, that is, when all pseudo-user sentences have been processed, the learning phase ends.

[0069] A large amount of pseudo reward data (pseudo reward data) can be obtained by going through the processes of steps S123 to S126, step D121, and step D122. The pseudo reward data is data that combines pseudo user sentences, molecular structures generated according to the pseudo user sentences, and pseudo request achievement levels evaluated for the molecular structures.

[0070] That is, a large amount of pseudo-user sentences are created, a group of molecular SMILES is generated according to the pseudo-user sentences, and the degree of pseudo-desire achievement is predicted for each molecular SMILES in the group of molecular SMILES, thereby obtaining a large amount of pseudo-reward data.The large amount of pseudo-reward data can be used to accelerate the optimization of a large-scale language model for molecule generation, thereby enabling efficient design of organic compounds having desired physical property values.

[0071] <Learning Phase 2 by Reinforcement Learning (Flowchart F140)> In the organic compound design system of one embodiment of the present invention, the learning phase of the large-scale language model for molecule generation by reinforcement learning can be represented, for example, by a flowchart F140 shown in Fig. 6. The flowchart F140 is a learning phase performed without using a reward model.

[0072] The learning phase shown in this section does not use a reward model, so although it is small-scale, it is faithful to the simulation results and can obtain highly accurate learning data.

[0073] First, data combining a user sentence and molecular SMILES associated with the user sentence, and a request achievement level for the combined data are acquired as learning data from the learning data database (step S141).

[0074] Here, it is determined whether the training data acquired in step S141 has been input to the large-scale language model for generating molecules (step D141).

[0075] If there is any training data that has not been input to the large-scale language model for generating molecules, i.e., unprocessed training data, the parameters of the large-scale language model for generating molecules are updated according to the degree of fulfillment of the request (step S142). The processing of step S142 is performed on all training data obtained in step S141.

[0076] On the other hand, when all the training data has been input to the large-scale language model for generating molecules, that is, when all the training data has been processed, the training phase ends.

[0077] By following this learning phase, the large-scale language model for molecular generation is optimized using reward data consisting only of physical property values ​​calculated using the physical property calculation system, making it possible to design new organic compounds with high accuracy in response to requirements.

[0078] <Method for calculating physical property values ​​by scientific calculation (flowchart F200)> Below, a method for calculating physical property values ​​by scientific calculation that can be used in a physical property calculation system will be described. Specifically, when the molecular structure converted from a molecular SMILES generated using a large-scale language model for molecule generation is an organometallic complex A1, a method for calculating the physical property values ​​of the organometallic complex A1 will be described with reference to Figures 7 to 16. In this section, quantum chemical calculation (calculation software Gaussian) is used for the scientific calculation.

[0079] The physical property values ​​of an organic compound include, for example, each energy level such as the T1 level, the S1 level, the HOMO level, and the LUMO level, as well as each energy level such as the energy level difference in an excited state, and the rearrangement energy (λ).

[0080] First, a description will be given of a flowchart F200 shown in Figures 7 and 8. In the flowchart F200, an example of calculating the T1 level as a physical property value will be described below.

[0081] 7, the ground state of an arbitrary organometallic complex A1 is optimized (step S201). Next, the excitation energy of the organometallic complex A1 is calculated (step S202). Subsequently, the excited site of the organometallic complex A1 is analyzed (step S203).

[0082] Next, the molecular structure is edited according to the excited site of the organometallic complex A1 (step S204). Note that in an organic compound, it is necessary to confirm multiple excited states for the T1 level, so it is necessary to repeatedly edit the molecular structure and confirm the energies of multiple excited states. Step S204 will be described later using a flowchart F210.

[0083] Next, the excited state of the organometallic complex A1 edited in step S204 is optimized (step S205), and the energy level (T1 level) in the excited state of the organometallic complex A1 edited in step S204 is calculated (step S206).

[0084] The processes of steps S205 and S206 are performed for all excited sites in the organometallic complex A1 group (step D201). Then, as shown in Figure 8, the structure with the lowest energy level is selected from the molecular structures of the edited organometallic complex A1 group (step S207). Next, the physical properties of the molecular structure selected in step S207 are obtained (step S208).

[0085] As described above, by following the flowchart F200, quantum chemical calculations (calculation software Gaussian) can be performed on the organometallic complex A1 group.

[0086] <<Editing Molecular Structure According to Excitation Site (Flowchart F210)>> Below, a method for editing the molecular structure according to the excitation site of the organometallic complex A1 group (step S204) will be described using the flowchart F210 shown in Figures 9, 10, and 11. Note that the hole distribution and electron distribution in the excited state can be obtained using, for example, open source software such as Multiwfn. Furthermore, the molecular structure can be edited using RDKit, an open source cheminformatics toolkit.

[0087] 9, three-dimensional grid data of excited-state hole distribution (hereinafter referred to as hole distribution) is read for an arbitrary organometallic complex A1 (step S211). Next, three-dimensional grid data of excited-state electron distribution (hereinafter referred to as electron distribution) is read for the organometallic complex A1 (step S212). Subsequently, information on all bonds (pairs of carbon atoms) in the organometallic complex A1 is listed (step S213).

[0088] Next, a threshold value is set to determine whether the values ​​in the grid of the hole distribution of organometallic complex A1 or the grid of the electron distribution of organometallic complex A1 are valid (step S214). Specifically, in order to efficiently search for locations with high density of hole distribution or electron distribution, it is preferable to set the distribution so that it is invalid when it is below a certain density.

[0089] Next, the coordinates of the initiating and terminating atoms of the bond in the hole distribution of the organometallic complex A1, or the coordinates of the initiating and terminating atoms of the bond in the electron distribution of the organometallic complex A1 are read (step S215).

[0090] 10 , the presence or absence of coordinates in the grid of hole distribution or the grid of electron distribution of organometallic complex A1 that are near the starting atom and whose numerical value exceeds a threshold value is determined (step D213). If there are coordinates in the grid of hole distribution or the grid of electron distribution of organometallic complex A1 that are near the starting atom and whose numerical value exceeds a threshold value is determined (step D214).

[0091] In addition, if there are no coordinates near the initial atom whose numerical value exceeds the threshold value in both the hole distribution and the electron distribution of organometallic complex A1, or no coordinates near the terminal atom whose numerical value exceeds the threshold value, the molecular structure of organometallic complex A1 is not edited (step D212). On the other hand, if there are coordinates near the initial atom whose numerical value exceeds the threshold value and coordinates near the terminal atom whose numerical value exceeds the threshold value, the molecular structure of organometallic complex A1 is edited.

[0092] 10 , when editing the molecular structure of organometallic complex A1, first, a coordinate in the hole distribution of organometallic complex A1 that is near the initiating atom and whose numerical value exceeds a threshold is set as the start coordinate, and a coordinate in the vicinity of the terminal atom and whose numerical value exceeds the threshold is set as the target coordinate. Alternatively, a coordinate in the electron distribution of organometallic complex A1 that is near the initiating atom and whose numerical value exceeds a threshold is set as the start coordinate, and a coordinate in the vicinity of the terminal atom and whose numerical value exceeds the threshold is set as the target coordinate (step S216).

[0093] Here, it is determined whether the start coordinates and the target coordinates are connected in the hole distribution or electron distribution of the organometallic complex A1 (step D215). The determination method in step S215 will be described later with reference to a flowchart F220.

[0094] If the starting coordinates and the target coordinates are not connected in both the hole distribution and the electron distribution of the organometallic complex A1, the molecular structure of the organometallic complex A1 is not edited. On the other hand, if the starting coordinates and the target coordinates are connected in either the hole distribution or the electron distribution of the organometallic complex A1, the molecular structure of the organometallic complex A1 is edited.

[0095] Specifically, if the starting coordinates and the target coordinates are connected in the hole distribution of the organometallic complex A1, the bond length between the starting atom and the terminal atom is shortened (step S217), or if the starting coordinates and the target coordinates are connected in the electron distribution of the organometallic complex A1, the bond length between the starting atom and the terminal atom is extended (step S218).

[0096] The processes of steps S215 to S218 and steps D212 to D216 are performed for all molecular bonds of the organometallic complex A1 (step D211).

[0097] As described above, by following the flowchart F210, the molecular structure of the organometallic complex A1 group can be edited according to the excited site.

[0098] <<Determining Connection Between Start Coordinates and Target Coordinates Within Distribution (Flowchart F220)>> Below, a method for determining the connection between the start coordinates and target coordinates of the bond possessed by the organometallic complex A1 within the hole distribution or the electron distribution (Step D215) will be described using the flowchart F220 shown in Figures 12, 13, 14, 15, and 16.

[0099] First, as shown in FIG. 12 , grid data for the three-dimensional hole distribution or three-dimensional electron distribution of the target organometallic complex A1 is read (step S221). Next, the start coordinates and target coordinates of the organometallic complex A1 set in step S216 of flowchart F210 are read (step S222). Next, an empty path stack is created (step S223). Next, the start coordinates of the bond in the organometallic complex A1 are set to the current coordinates (step S224).

[0100] In steps S221 to S224, the processes may be executed in any order and may be interchanged. A stack is a basic data structure that stores data in a last-in, first-out (LIFO: Last In First Out; FILO: First In Last Out) structure. The search history records searched coordinates regardless of whether the surrounding coordinates of the current coordinates are searchable, while the path stack stores coordinates that are searchable as surrounding coordinates of the current coordinates.

[0101] Next, as shown in FIG. 13, if the starting coordinates (current coordinates) of the bond of organometallic complex A1 are set to coordinates (Xs, Ys, Zs), the coordinates (Xs-1, Ys, Zs) are set as the destination relative to the current coordinates (Xs, Ys, Zs) (step S225).

[0102] Here, it is determined whether the destination coordinates (Xs-1, Ys, Zs) are unsearched and are equal to or greater than the threshold (step D221).

[0103] If the destination coordinates (Xs-1, Ys, Zs) have already been searched or are smaller than the threshold, the coordinates (Xs+1, Ys, Zs) are set as the destination for the current coordinates (Xs, Ys, Zs) (step S226), and it is determined whether the destination coordinates have not yet been searched and are equal to or greater than the threshold (step D222). As shown in Figures 13 and 14, similar processes are then performed for the Y coordinate and the Z coordinate (steps S227, D223, S228, D224, S229, D225, D226, and S230).

[0104] As shown in FIG. 15, in the processing of steps S225 to S230 and steps D223 to D226, if all coordinates of the destinations are unsearched or are not above the threshold, it is determined whether the path stack is empty (step D227).

[0105] In step D227, if the path stack is empty, it is determined that the distributions of the current coordinates (start coordinates) and the target coordinates are not connected (step S231).

[0106] On the other hand, if coordinate information is held in the path stack in step D227, the current coordinate is added to the search history (step S232). Next, the latest coordinate is retrieved from the path stack and set as the current coordinate (step S233), and then the process returns to step S225. Since the coordinates held in the path stack cannot be searched for in the vicinity of the coordinate, by returning to step S225, it is possible to search for coordinates in the vicinity of the updated current coordinate.

[0107] 16, in the processing of steps S225 to S230 and steps D223 to D226, if the destination coordinate is unsearched or equal to or greater than the threshold, the current coordinate of the bond in organometallic complex A1 is added to the path stack and the search history (steps S237 and S238). Note that steps S237 and S238 can be performed in any order. Therefore, the path stack holds coordinates that can be searched for as neighboring coordinates of the current coordinate.

[0108] For example, if the destination coordinates (Xs-1, Ys, Zs) are unsearched and are equal to or greater than the threshold, the current coordinates (Xs, Ys, Zs) are added to the path stack and the search history. That is, the latest history of the path stack and the search history becomes the coordinates (Xs, Ys, Zs).

[0109] Next, the current coordinates (Xs, Ys, Zs) are set to the destination coordinates (Xs-1, Ys, Zs) and updated to (Xs-1, Ys, Zs) (step S239).

[0110] Next, it is determined whether the current coordinates are the target coordinates (step D228). If any of the coordinates where the bond in the organometallic complex A1 has moved is equal to the target coordinates, it is determined that the start coordinates and the target coordinates are connected (step S240). For example, if the current coordinates are (Xs-1, Ys, Zs) and the target coordinates are (Xs-1, Ys, Zs), it is determined that the start coordinates (Xs, Ys, Zs) and the target coordinates are connected in distribution.

[0111] On the other hand, if any of the coordinates to which the bond of organometallic complex A1 has moved is not the target coordinate, the process returns to step S225. For example, if the current coordinates set in step S239 have been updated to the coordinates (Xs-1, Ys, Zs), the process returns to step S225, whereby peripheral coordinates of the coordinates (Xs-1, Ys, Zs) can be searched for.

[0112] From the above, by following the flowchart F220, it is possible to determine the connection between the start coordinate and the target coordinate of the bond of the organometallic complex A1 in the hole distribution or the electron distribution.

[0113] The above is an explanation of a method for designing the molecular structure of an organic compound having desired properties.

[0114] <Neural Network> Values ​​predicted by a neural network can be used for the molecular structure of an organic compound generated by the large-scale language model for molecule generation 12, the degree of desire fulfillment evaluated by the large-scale language model for desire fulfillment evaluation 15, or the predicted value of the degree of desire fulfillment predicted by the reward prediction system 17. In addition, predicted values ​​by an artificial neural network (hereinafter simply referred to as a neural network) can be used as a method for quickly calculating physical property values ​​by the physical property calculation system 14.

[0115] Below, as an example of a neural network, a neural network that can be used for supervised learning will be described.

[0116] As shown in FIG. 17A, the neural network NN can be composed of an input layer IL, an output layer OL, and a hidden layer HL. Each of the input layer IL, output layer OL, and hidden layer HL has one or more neurons (units). The hidden layer HL may have one layer or two or more layers. A neural network having two or more hidden layers HL can also be called a deep neural network (DNN). Learning using a deep neural network can also be called deep learning.

[0117] Input data is input to each neuron in the input layer IL. An output signal from a neuron in the previous or next layer is input to each neuron in the hidden layer HL. An output signal from a neuron in the previous layer is input to each neuron in the output layer OL. Note that each neuron may be connected to all neurons in the previous or next layer (fully connected), or may be connected to only a portion of the neurons.

[0118] An example of a computation by a neuron is shown in Figure 17B. Here, neuron N and two neurons in the previous layer that output signals to neuron N are shown. Neuron N receives the output x of the previous layer neuron. １ and the output of the previous layer neuron x ２ is input. Then, in neuron N, the output x １ And weight lol １ The multiplication result (x １ w １ ) and output x ２ And weight lol ２ The multiplication result (x ２ w ２ ) the sum x １ w １ +x ２ w ２ After is calculated, a bias b is added if necessary, and the value a = x １ w １ +x ２ w ２+b is obtained. The value a is then transformed by the activation function h, and an output signal y=ah is output from the neuron N. As the activation function h, for example, a sigmoid function, a tanh function, a softmax function, a ReLU function, a threshold function, or the like can be used.

[0119] In this way, the operation by a neuron includes the operation of adding the product of the output of the neuron in the previous layer and the weight, that is, the sum-of-products operation (the above x １ w １ +x ２ w ２ ). When the product-sum operation is performed by hardware, a product-sum operation circuit can be used. This product-sum operation circuit may be a digital circuit or an analog circuit. When an analog circuit is used for the product-sum operation circuit, it is possible to reduce the circuit size of the product-sum operation circuit or the number of times memory is accessed, thereby improving processing speed and reducing power consumption.

[0120] The product-sum operation circuit may be configured using a transistor containing silicon (such as single crystal silicon) in a channel formation region (hereinafter also referred to as a Si transistor) or a transistor containing an oxide semiconductor in a channel formation region (hereinafter also referred to as an OS transistor). OS transistors have extremely low off-state current and are therefore suitable as transistors constituting the analog memory of the product-sum operation circuit. The product-sum operation circuit may be configured using both Si transistors and OS transistors.

[0121] The above is a description of the neural network. In one aspect of the present invention, it is preferable to use deep learning. In other words, it is preferable to use a neural network having two or more hidden layers HL.

[0122] <Representation of Molecular Structure of Organic Compound> Note that in the learning method, design method, or design system for an organic compound design model of one embodiment of the present invention, information about the organic compound can be represented by a structural formula, a molecular structure represented by SMILES (Simplified Molecular Input Line Entry Specification Syntax), a name according to the compound nomenclature established by IUPAC, or the like.

[0123] On the other hand, in one embodiment of the present invention, since arithmetic processing is performed on input information, it is preferable to use numerical notation for information about organic compounds (for example, molecular structure).

[0124] Therefore, in one aspect of the present invention, the molecular structure that identifies an organic compound is converted using a specific method. The specific method may be capable of expressing molecular similarity. Well-known methods for expressing molecular similarity include quantitative structure-activity relationship (QSAR), fingerprinting, and graph structure. For example, it is preferable to digitize the molecular structure using linear or matrix notation.

[0125] When quantifying the structure of an inorganic compound, it is advisable to use descriptors such as the radial distribution function (RDF) and the orbital field matrix (OFM).

[0126] [Example of Numerical Expression of Molecular Structure of Organic Compound] Here, an example of numerical expression (mathematical expression) of the molecular structure of an organic compound will be described.

[0127] When information about organic compounds is input as non-numeric data other than SMILES notation, it is preferable to first convert it to SMILES notation. SMILES notation expresses organic compounds as a continuous string of characters, making it preferable for computer-handled data. SMILES notation and the fingerprinting method described below are both classified as linear notations, making them easy to convert between, making them preferable.

[0128] The open-source cheminformatics toolkit RDKit can be used to quantify molecular structures. RDKit can convert the SMILES notation of the input molecular structure into mathematical data (quantification) using the fingerprinting method.

[0129] In fingerprinting, as shown in FIG. 18, for example, a molecular structure is represented by assigning a partial structure (fragment) of the molecular structure to each bit, and if the corresponding partial structure exists in the molecule, the bit is set to "1", and if not, "0". In other words, by using fingerprinting, it is possible to obtain a mathematical formula that extracts the characteristics of the molecular structure. Furthermore, the molecular structure formula represented by fingerprinting is generally several hundred to tens of thousands of bits long, which is an easy-to-handle size. Furthermore, by using fingerprinting to represent a molecular structure as a mathematical formula of 0s and 1s, it is possible to achieve extremely high-speed calculation processing.

[0130] In addition, there are many types of fingerprinting methods (those that take into account different bit generation algorithms, atom types, bond types, and aromaticity conditions, and those that dynamically generate bit lengths using hash functions, etc.), each with its own characteristics.

[0131] 19A to 19D show examples of types of fingerprinting. Typical types of fingerprinting include the circular type shown in FIG. 19A (surrounding atoms up to a specified radius from a starting atom are considered to be the partial structure), the path-based type shown in FIG. 19B (atoms from the starting atom up to a specified path length are considered to be the partial structure), the substance keys type shown in FIG. 19C (partial structures are defined for each bit), and the atom pair type shown in FIG. 19D (atom pairs generated for all atoms in a molecule are considered to be the partial structure). Fingerprints of each of these types are implemented in the RDKit.

[0132] Figure 20 shows an example of the molecular structure of an organic compound actually expressed as a mathematical formula using the fingerprint method. In this way, the molecular structure can be converted into a SMILES notation and then converted into a fingerprint.

[0133] When expressing the molecular structure of an organic compound using a fingerprinting method, the resulting mathematical formula may be identical between different organic compounds with similar structures. As described above, there are several types of fingerprinting methods depending on the notation method, but the tendency for compounds to be identical varies depending on the notation method, as shown in Figure 21, which includes the circular type (Morgan Fingerprint), path-based type (RDK Fingerprint), substance keys type (Avalon Fingerprint), and atom pair type (Hash atom pair). In Figure 21, the molecules within each double arrow each represent the same mathematical formula (notation). Therefore, it is preferable to use a fingerprinting method in which, when the molecular structure of each organic compound to be learned is expressed using at least one of the fingerprinting methods, the notations for each organic compound are all different. In FIG. 21, it can be seen that compounds with different atom pair types can be expressed without overlap, but depending on the population of organic compounds to be learned, other expression methods may also be possible without overlap.

[0134] Therefore, when describing the molecular structure of an organic compound using a fingerprinting method, it is preferable to use multiple different types of fingerprinting methods. Any number of types may be used, but two or three types are preferable because they are easier to handle in terms of data volume. When using multiple types of fingerprinting methods, the molecular structure of an organic compound may be described by connecting a formula represented by one type of fingerprinting method to a formula represented by another type of fingerprinting method, or by describing the molecular structure of an organic compound as having multiple different formulas. Figures 22A and 22B show an example of a method for describing the molecular structure of an organic compound using multiple fingerprints of different types.

[0135] Fingerprints are a method of describing the presence or absence of substructures, and information about the overall molecular structure is lost. However, if a molecular structure is expressed mathematically using multiple fingerprints of different types, different substructures are generated for each fingerprint type, and information about the presence or absence of these substructures can be used to complement information about the overall molecular structure. If a feature that cannot be fully expressed by a fingerprint has a significant impact on the characteristics of a light-emitting device, it can be complemented by other fingerprints, making the method of describing a molecular structure using multiple fingerprints of different types effective.

[0136] As shown in FIG. 22A, when representation is performed using two types of fingerprint methods, it is preferable to use the Atom Pair type and the Circular type, as this allows for accurate prediction of physical properties.

[0137] Furthermore, as shown in FIG. 22B , when notating using three types of fingerprint methods, it is preferable to use the Atom Pair type, the Circular type, and the Substructure Keys type, as these enable accurate prediction of physical properties.

[0138] Furthermore, when a circular fingerprinting method is used, the radius r is preferably 3 or more, and more preferably 5 or more. Note that the radius r is the number of atoms counted in connection from a certain atom that serves as the starting point, which is set as 0.

[0139] When selecting a fingerprinting method to be used, as mentioned above, it is preferable to select at least one method in which the molecular structures of the organic compounds are all represented differently.

[0140] By increasing the bit length (number of bits) used to represent a fingerprint, the possibility of generating descriptions that completely match between organic compounds can be reduced; however, increasing the bit length too much results in a trade-off in increased computational costs or database management costs. On the other hand, by simultaneously using multiple fingerprints for representation, even if there are multiple molecular structures whose representations are completely identical in a certain fingerprint type, combining different fingerprint types may result in an overall incomplete match. As a result, a state in which multiple organic compounds whose representations are completely identical in fingerprints can be generated with as small a bit length as possible. There is no particular limit to the bit length of the fingerprint to be generated; however, considering computational costs or database management costs, for molecules with molecular weights of up to about 2000, fingerprints that do not completely match between organic compounds can be generated with a bit length of 4096 or less, preferably 2048 or less, and in some cases 1024 or less for each fingerprint type.

[0141] Furthermore, the bit length of the fingerprints generated for each fingerprint type can be adjusted appropriately taking into account the characteristics of the type and the overall molecular structure, and does not need to be unified. For example, the bit length may be expressed as 1024 bits for the Atom Pair type and 2048 bits for the Circular type, and then these may be concatenated.

[0142] The above is an explanation of how to quantify the molecular structure of an organic compound.

[0143] As described above, one embodiment of the present invention can provide a learning method, a design method, and a design system for a design model of a molecular structure of an organic compound having desired properties.

[0144] One aspect of the present invention makes it possible to predict the properties of the molecular structure of a novel organic compound. Furthermore, by using past experimental data, it is possible to speed up the optimization of the molecular structure of an organic compound with desired properties through virtual screening. Even if a human observes the data and finds no interpolation, the nonlinear or higher-order representation of the machine learning model may make it so. Furthermore, by extracting and examining fragments of the representation obtained by the machine learning model, it is possible to discover patterns that would not have been previously noticed.

[0145] This embodiment can be carried out by combining parts thereof as appropriate.

[0146] Embodiment 2 In Embodiment 2, a system for designing an organic compound, which is one embodiment of the present invention, will be described.

[0147] A design system 20 for an organic compound according to one embodiment of the present invention includes at least an input unit, a learning unit, a prediction unit, an output unit, and a data server. These units may be incorporated into a single device, or may be separate devices, or some may be incorporated into the same device, or the data server may be a cloud, as long as they are capable of exchanging data. However, these units are collectively referred to as a property prediction system.

[0148] 23, one embodiment of the present invention will be described by taking as an example an organic compound design system comprising an information terminal having input means, learning means, prediction means, and output means, and a data server 28. The information terminal 21 has at least an input section, learning means, prediction means, and output section, and is capable of exchanging data with the separately installed data server 28. Note that the data server 28 may be installed within the information terminal 21.

[0149] The information terminal 21 mainly comprises an input unit 22, a calculation unit 24, and an output unit 26. The calculation unit 24 functions as both a learning means and a prediction means. Preferably, the calculation unit 24 also includes a neural network circuit. Data provided from the data server 28 is used as data for learning or prediction by the neural network circuit. By using a portion of this data as verification data and training data for the learned learning means, it is possible to update the weight coefficients in the neural network circuit and generate learned weight coefficients. This can further improve the accuracy of prediction.

[0150] 23, arrows indicate the flow of signals in the order of input unit 22, calculation unit 24, data server 28, and output unit 26. In this specification, signals can be appropriately interpreted as data or information.

[0151] The data server 28 provides the structures and physical property values ​​of the organic compounds to be learned to the learning means of the calculation unit 24. The structures of the organic compounds provided are represented using two or more types of fingerprints. The learning means of the calculation unit 24 preferably has a neural network circuit.

[0152] The input unit 22 has a function for the user to input information. Specific examples of the input unit 22 include any input means such as a keyboard, a mouse, a touch panel, a pen tablet, a microphone, or a camera.

[0153] Input information D ｉｎ is the data output from the input unit 22 to the calculation unit 24. ｉｎ is information input by the user. For example, if the input unit 22 is a touch panel, it is information obtained by inputting characters by operating the touch panel. Alternatively, if the input unit 22 is a microphone, it is information obtained by voice input by the user. Alternatively, if the input unit 22 is a camera, it is information obtained by image processing of captured data.

[0154] Input information D ｉｎis information about the structure of the organic compound whose physical properties are to be predicted. If a structural formula, structural image, substance name, or other information other than a fingerprint notation is input, it is input to the prediction means in the calculation unit 24 after passing through an appropriate conversion means. The prediction means predicts the physical properties of the input organic compound based on the results learned in advance by the learning means.

[0155] The result of the prediction is output via the output unit 26 .

[0156] [Data Server 28] The data server 28 has a function of storing the program executed by the calculation unit 24. The data server 28 may also have a function of storing data generated by the calculation unit 24 (e.g., calculation results, analysis results, inference results), data input to the input unit 22, etc.

[0157] The data server 28 may have a database. Furthermore, the design system 20 may have a database separate from the data server 28. The design system 20 may have a function to retrieve data from a database that exists outside the data server 28 or outside the design system 20. Furthermore, the design system 20 may have a function to retrieve data from both its own database and an external database.

[0158] A file server may be used instead of the database. For example, when using files stored in a file server, it is preferable that the database has paths to files stored in the file server.

[0159] The data server 28 includes at least one of a volatile memory and a non-volatile memory. Examples of the volatile memory include a dynamic random access memory (DRAM) and a static random access memory (SRAM). Examples of the non-volatile memory include a resistive random access memory (ReRAM), a phase change random access memory (PRAM), a ferroelectric random access memory (FeRAM), a magnetoresistive random access memory (MRAM), and a flash memory. The data server 28 may also have a recording media drive, such as a hard disk drive (HDD) or a solid state drive (SSD).

[0160] [Calculation Unit 24] The calculation unit 24 may include, for example, an arithmetic circuit, a central processing unit (CPU), or a graphics processing unit (GPU).

[0161] The calculation unit 24 may have a microprocessor such as a DSP (Digital Signal Processor). The microprocessor may be implemented by a PLD (Programmable Logic Device) such as an FPGA (Field Programmable Gate Array) or an FPAA (Field Programmable Analog Array). The calculation unit 24 may also have a quantum processor. The calculation unit 24 can perform various data processing and program control by interpreting and executing instructions from various programs using the processor. Programs that can be executed by the processor are stored in at least one of the memory area of ​​the processor and the data server 28.

[0162] When the arithmetic unit includes a neural network circuit, the neural network circuit preferably includes a product-sum operation circuit capable of performing product-sum operation. The product-sum operation circuit preferably includes a memory circuit for storing weight data. The memory element included in the memory circuit includes a transistor and a capacitor. The transistor is preferably a transistor including an oxide semiconductor in a semiconductor layer having a channel formation region (hereinafter referred to as an OS transistor). The OS transistor has an extremely small leakage current when it is off. Therefore, data can be stored by utilizing the property of being able to retain charge by turning the OS transistor off.

[0163] Another aspect of the present invention is a recording medium on which a control program and control software are recorded that can generate fingerprints that are concatenated or parallelized using these multiple fingerprint types, perform machine learning, and predict physical properties.

[0164] F100 flowchart, F120 flowchart, F140 flowchart, F200 flowchart, F210 flowchart, F220 flowchart, S101: step, S102: step, S103: step, S104: step, S105: step, S106: step, S107: step, S108: step, S109: step, S110: step, S121: step, S122: step, S123: step, S124: step, S125: step, S126: step, S141: step, S142: step, S201: step, S202: step, S203: step, S204: step, S205: step, S206: step, S207: step, S208: step, S211: step, S212: step, S213: step, S214: step, S215: step, S216: step, S217: step, S218: step, S221: step, S222: step, S223: step, S224: step, S225: step, S226: step, S227: step, S228: step, S229: step Step, S230: Step, S231: Step, S232: Step, S233: Step, S237: Step, S238: Step, S239: Step, S240: Step, D101: Step, D121: Step, D122: Step, D141: Step, D201: Step, D211: Step, D212: Step, D213: Step, D214: Step, D215: Step, D216: Step, D221: Step, D222: Step, D223: Step Step, D224: Step, D225: Step, D226: Step, D227: Step, D228: Step, 10: Data model, 11: Main system, 12: Large-scale language model for molecule generation, 13: Database for learning data, 14: Physical property calculation system, 15: Large-scale language model for evaluating the degree of achievement of requests, 16: Learning management system, 17: Reward prediction system, 19: User, 20: Design system, 21: Information terminal, 22: Input unit, 24: Calculation unit, 26: Output unit, 28: Data server

Claims

generating a first set of organic compounds using the user text with a first large-scale language model; calculating physical property values ​​of the first organic compound included in the first organic compound group by scientific calculation; evaluating a degree of fulfillment of a request from a correlation between the user text, the molecular structure of the first organic compound, and the physical property value by a second large-scale language model; generating a plurality of learning data by combining the user text, the molecular structure of the first organic compound, and the degree of achievement of the request; constructing a reward model using the training data; creating a pseudo user sentence; generating a second set of organic compounds using the pseudo-user sentences through the first large-scale language model; generating pseudo-reward data by evaluating a molecular structure of a second organic compound included in the second organic compound group using the reward model; and updating parameters of the first large-scale language model using the pseudo-reward data. 　 generating a first set of organic compounds using the user text with a first large-scale language model; calculating physical property values ​​of the first organic compound included in the first organic compound group by scientific calculation; evaluating a degree of fulfillment of a request from a correlation between the user text, the molecular structure of the first organic compound, and the physical property value by a second large-scale language model; generating a plurality of learning data by combining the user text, the molecular structure of the first organic compound, and the degree of achievement of the request; updating parameters of the first large-scale language model using the training data; A method for learning a design model of an organic compound having the formula:

2. The method for learning an organic compound design model according to claim 1, wherein the reward model predicts a degree of pseudo-request achievement from the pseudo-user sentence and the molecular structure of the second organic compound.

4. The method for learning a design model of an organic compound according to claim 1, wherein the physical property value is a T1 level. 　 An input unit, a calculation unit, and an output unit, the input unit has a function of outputting input user text to the calculation unit; The calculation unit is generating a first group of organic compounds using the user text by a first large-scale language model; a function of calculating a physical property value of the first organic compound included in the first organic compound group by scientific calculation; a function of evaluating a degree of fulfillment of a request from a correlation between the user text, the molecular structure of the first organic compound, and the physical property value by a second large-scale language model; a function of listing molecular structures of organic compounds and physical property values ​​of the molecular structures in descending order of degree of achievement of the request by the second large-scale language model; The output unit has a function of outputting the listed molecular structures and physical property values.

6. The system for designing an organic compound according to claim 5, further comprising a function for generating learning data that combines the user text, the molecular structure of the first organic compound, and the degree of achievement of the request. 　 A function of generating learning data combining the user text, the molecular structure of the first organic compound, and the degree of request achievement according to claim 5, and updating parameters of the first large-scale language model using the training data.

8. The system for designing an organic compound according to claim 5, wherein the physical property value is a T1 level.

Images