Learning device, learning method, trained model, and prediction device

Abstract

To provide a learning device, a learning method, a learned model and a predictor, capable of improving a prediction accuracy of complex physical properties.SOLUTION: A machine learning system 1 includes: an acquisition part 10 configured to acquire a molecular structure descriptor of a complex, a physical properties descriptor of a complex and complex physical properties showing physical properties of a complex; and a learning part 11 configured to learn a relation among the molecular structure descriptor, the physical properties descriptor and the complex physical properties, based on the molecular structure descriptor, the physical properties descriptor and the complex physical properties, and generate a first learning model that estimates complex physical properties according to an input of the molecular structure descriptor and the physical properties descriptor.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a learning device, a learning method, a trained model, and a prediction device. [Background Art]

[0002] Non-Patent Document 1 discloses a method for predicting the oxidation-reduction potential of a metal complex using a trained model generated by machine learning that uses molecular structure descriptors of metal complexes. [Prior Art Literature] [Non-Patent Literature]

[0003] [Non-Patent Document 1] Jon Paul Janet and Heather J.Kulik,“Resolving Transition Metal Chemical Space:Feature Selection for Machine Learning and Structure-Property Relationships”,J. Phys. Chem. A,2017,121,46,p.8939-8954 [Summary of the Invention] [Problem to be Solved by the Invention]

[0004] In the prediction method described in Non-Patent Document 1, an oxidation-reduction potential, which is a complex physical property indicating a physical property of a metal complex, is predicted using a trained model generated by machine learning that uses molecular structure descriptors depending on the structure and molecular weight of the metal complex. With this method, complex physical properties can be accurately predicted (estimated) for input data that has a distribution similar to that of the structure and molecular weight of metal complexes in the training data (teacher data) used for machine learning of the learning model. However, with conventional prediction methods, the prediction accuracy of complex physical properties may decrease for input data having a distribution different from the aforementioned distribution of the training data.

[0005] This disclosure aims to provide a learning device, a learning method, a trained model, and a prediction device that can improve the accuracy of predicting complex physical properties. [Means for solving the problem]

[0006] A learning device according to one embodiment of the present disclosure is a learning device that performs machine learning of a learning model and includes at least one processor, an acquisition unit that acquires a molecular structure descriptor of a complex, a physical property descriptor of a complex, and complex physical properties that indicate the physical properties of a complex, and a learning unit that learns the relationship between the molecular structure descriptor and the physical property descriptor and the complex physical properties based on the molecular structure descriptor and the physical property descriptor and generates a first learning model that estimates the complex physical properties in response to input of the molecular structure descriptor and the physical property descriptor. [Effects of the Invention]

[0007] According to this disclosure, the accuracy of predicting complex properties can be improved. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 shows an example of the computer hardware configuration used in a prediction system according to one embodiment. [Figure 2] Figure 2 shows an example of the functional configuration of the prediction system according to the embodiment. [Figure 3] Figure 3 is a flowchart showing an example of a learning method as a processing flow. [Figure 4] Figure 4 is a flowchart showing an example of a learning method as a processing flow. [Figure 5] Figure 5 is a flowchart showing an example of a prediction method as a processing flow. [Modes for carrying out the invention]

[0009] [Description of Embodiments in this Disclosure] First, the embodiments of this disclosure will be listed and explained. At least some of the embodiments described below may be combined in any way.

[0010] A learning device relating to one aspect of this disclosure is a learning device that performs machine learning of a learning model and includes at least one processor, an acquisition unit that acquires a molecular structure descriptor of a complex, a physical property descriptor of a complex, and complex physical properties that indicate the physical properties of a complex, and a learning unit that learns the relationship between the molecular structure descriptor and the physical property descriptor and the complex physical properties based on the molecular structure descriptor and the physical property descriptor and generates a first learning model that estimates the complex physical properties in response to inputs of the molecular structure descriptor and the physical property descriptor.

[0011] In one aspect of this disclosure, the learning device learns the relationship between molecular structure descriptors and property descriptors and complex properties, and generates a first learning model that estimates complex properties in response to input molecular structure descriptors and property descriptors. Thus, the learning device generates the first learning model using property descriptors that are less dependent on the structure of the complex metal, in addition to molecular structure descriptors. As a result, the first learning model can accurately estimate complex properties in response to input molecular structure descriptors and property descriptors, without depending on the structure of the complex metal. Therefore, the learning device can improve the accuracy of predicting complex properties.

[0012] In one embodiment, the learning unit may use a property descriptor estimated by a second learning model that estimates property descriptors based on molecular structure descriptors. In this configuration, by using the property descriptors estimated by the second learning model for learning, a first learning model capable of accurately estimating complex properties can be generated.

[0013] In one embodiment, the acquisition unit acquires complex data including coordinate data and property descriptors of the complex, and the learning unit converts the coordinate data included in the complex data into molecular structure descriptors and acquires the property descriptors included in the complex data. Based on the converted molecular structure descriptors and property descriptors, the learning unit learns the relationship between the molecular structure descriptors and property descriptors and generates a second learning model that estimates the property descriptors in response to the input of molecular structure descriptors. In this configuration, a second learning model that estimates the property descriptors in response to the input of molecular structure descriptors can be appropriately generated. Furthermore, the second learning model generated as described above can estimate the property descriptors with high accuracy.

[0014] A learning method relating to one aspect of this disclosure is a learning method that is executed on at least one processor and performs machine learning of a learning model, and includes: an acquisition step of acquiring a molecular structure descriptor of a complex, a physical property descriptor of a complex, and complex properties that indicate the physical properties of a complex; and a learning step of learning the relationship between the molecular structure descriptor and the physical property descriptor and the complex properties based on the molecular structure descriptor and the physical property descriptor and generating a first learning model that estimates the complex properties in response to inputs of the molecular structure descriptor and the physical property descriptor.

[0015] In one aspect of this disclosure, a learning method is used to learn the relationship between molecular structure descriptors and property descriptors and complex properties, and to generate a first learning model that estimates complex properties in response to input molecular structure descriptors and property descriptors. Thus, the learning method generates the first learning model using property descriptors that are less dependent on the structure of the complex metal, in addition to molecular structure descriptors. As a result, the first learning model can accurately estimate complex properties in response to input molecular structure descriptors and property descriptors, without depending on the structure of the complex metal. Therefore, the learning method can improve the accuracy of predicting complex properties.

[0016] A pre-trained model relating to one aspect of this disclosure is a pre-trained model for estimating complex properties that exhibit the physical properties of a complex, and is constructed by learning the relationship between the molecular structure descriptor and property descriptor of the complex and the complex properties, and estimates the complex properties in response to the input of the molecular structure descriptor and property descriptor.

[0017] In the trained model according to one aspect of the present disclosure, the relationship between a molecular structure descriptor of a complex, a descriptor of the complex and a physical property of the complex is learned, and the physical property of the complex is estimated in accordance with inputs of the molecular structure descriptor and a physical property descriptor. As described above, in the trained model, learning is performed using, in addition to the molecular structure descriptor, a physical property descriptor that hardly depends on the structure of the complex metal. Accordingly, the trained model can accurately estimate the physical property of the complex in accordance with inputs of the molecular structure descriptor and the physical property descriptor without depending on the structure of the complex metal. Therefore, in the trained model, the prediction accuracy of the complex physical property can be improved.

[0018] A prediction device according to one aspect of the present disclosure is a prediction device that includes at least one processor and predicts a complex physical property indicating a physical property of a complex using the above-described trained model, wherein the processor of the prediction device acquires input data and inputs the input data to the trained model, thereby outputting a prediction result related to the complex physical property.

[0019] In the prediction device according to one aspect of the present disclosure, the complex physical property is predicted using the above-described trained model. Therefore, in the prediction device, the prediction accuracy of the complex physical property can be improved.

[0020] [Details of Embodiments of the Present Disclosure] Specific examples of embodiments of the present disclosure will be described below with reference to the drawings. The present disclosure is not limited to these examples, is defined by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims. In the description of the drawings, the same reference numerals are given to the same elements, and overlapping descriptions will be omitted.

[0021] [Outline of System] The machine learning system 1 according to this embodiment is a computer system that predicts the physical properties (hereinafter referred to as "complex properties") of metal complexes (complex compounds). Metal complexes are formed by the bonding of ligands (molecules, ions) to metal ions. Metal complexes may include organometallic complexes and inorganic metal complexes. Examples of complex properties include magnetic properties (permeability, magnetic flux density, etc.), thermodynamic properties (oxidation-reduction potential, solubility, etc.), optical properties (infrared, ultraviolet, visible light spectra, etc.), and conductivity (ions, electrical conductivity, etc.). In this embodiment, the results predicted by the machine learning system 1 are referred to as "prediction results".

[0022] Machine learning system 1 utilizes machine learning to predict complex properties. Machine learning is a method that autonomously discovers laws or rules by iteratively learning based on given information. Machine learning system 1 performs machine learning using a machine learning model. For example, machine learning system 1 may perform machine learning using a Gradient Boosting Decision Tree (GBDT), machine learning using multiple regression analysis, or machine learning using a Convolutional Neural Network (CNN) which includes convolutional layers and pooling layers. A Convolutional Neural Network is a type of deep learning that uses a multi-layered neural network.

[0023] Machine learning system 1 trains a machine learning model by repeatedly performing learning, and obtains this machine learning model as a trained model. This corresponds to the learning phase. In the learning phase, machine learning system 1 functions as a learning device. The trained model is a machine learning model that is predicted to be optimal for predicting complex properties. Machine learning system 1 processes input data using the trained model and outputs prediction results for complex properties. This corresponds to the operation phase (prediction phase). In the operation phase, machine learning system 1 functions as a prediction device.

[0024] Trained models are portable across computer systems. Therefore, a trained model generated on one computer system can be used on another computer system. Of course, a single computer system may perform both the generation and use of the trained model. That is, machine learning system 1 may perform both the training phase and the operation phase, or it may perform only one of the training phases. In this embodiment, machine learning system 1 performs both the training phase and the operation phase.

[0025] In the learning phase, machine learning system 1 utilizes training data. The training data consists of complex data. Machine learning system 1 generates a trained model by performing machine learning using the training data. In the operation phase, machine learning system 1 obtains prediction results by providing input data to the trained model. The input data consists of data related to the molecular structure descriptor and physical property descriptor of the metal complex.

[0026] [System Configuration] Figure 1 shows an example of a typical hardware configuration of a computer 100 that constitutes a machine learning system 1. For example, the computer 100 includes a processor 101, a main memory unit 102, an auxiliary memory unit 103, a communication control unit 104, an input device 105, and an output device 106. The processor 101 executes the operating system and application programs. The main memory unit 102 consists of, for example, ROM and RAM. The auxiliary memory unit 103 consists of, for example, a hard disk or flash memory, and generally stores a larger amount of data than the main memory unit 102. The communication control unit 104 consists of, for example, a network card or a wireless communication module. The input device 105 consists of, for example, a keyboard, mouse, touch panel, etc. The output device 106 consists of, for example, a monitor and speakers.

[0027] Each functional element of the machine learning system 1 is realized by a learning program 110 and a prediction program 120 that are pre-stored in the auxiliary storage unit 103. Specifically, each functional element is realized by loading the learning program 110 or the prediction program 120 onto the processor 101 or the main memory unit 102 and executing the learning program 110 or the prediction program 120. The processor 101 operates the communication control unit 104, the input device 105, or the output device 106 according to the learning program 110 or the prediction program 120, and reads and writes data to the main memory unit 102 or the auxiliary storage unit 103. The data or database necessary for processing is stored in the main memory unit 102 or the auxiliary storage unit 103.

[0028] The learning program 110 and the prediction program 120 may be provided by being permanently recorded on a tangible recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory. Alternatively, the learning program 110 and the prediction program 120 may be provided via a communication network as data signals superimposed on a carrier wave.

[0029] The machine learning system 1 may consist of one computer 100 or multiple computers 100. When multiple computers 100 are used, these computers 100 are connected via a communication network such as the Internet or an intranet to logically construct a single machine learning system 1.

[0030] Figure 2 shows an example of the functional configuration of machine learning system 1. As shown in Figure 2, machine learning system 1 comprises an acquisition unit 10, a learning unit 11, a storage unit 12, a prediction unit 13, and a database 20.

[0031] The acquisition unit 10 is a functional element that acquires data. In the learning phase, the acquisition unit 10 acquires complex data as training data from the database 20. The acquisition unit 10 acquires one or more complex data at predetermined timings. In the operation phase, the acquisition unit 10 acquires input data.

[0032] The acquisition unit 10 can access a database 20 that stores complex data. The database 20 may be used to train a learning model. The database 20 may be, for example, a component of the machine learning system 1, or it may be built in a separate computer system from the machine learning system 1. The machine learning system 1 and the database 20 may be connected via a communication network, or both the machine learning system 1 and the database 20 may be built in a single computer.

[0033] The method for preparing the complex data to be stored in database 20 is not limited. For example, the complex data may be stored in database 20 through manual input by an operator, or it may be automatically collected and stored in database 20 by machine learning system 1 or another computer system.

[0034] The complex data stored in database 20 includes data for multiple metal complexes. In the complex data, each metal complex is associated with structure-optimized coordinate data and multiple physical property descriptors. Examples of physical property descriptors include energy (electron energy, dispersion energy, etc.), dipole moment and polarizability, molecular vibration, charge information, ionization energy (electron affinity), HOMO (Highest Occupied Molecular Orbital) / LUMO (Lowest Unoccupied Molecular Orbital), and dielectric constant.

[0035] The learning unit 11 is a functional element that performs machine learning. The learning unit 11 repeatedly performs the learning process to generate a trained model. The learning unit 11 generates a preliminary training model (second training model) and a predictive training model (first training model) as trained models.

[0036] (Pre-trained model) The generation of a preliminary learning model in the learning unit 11 will now be explained. The learning unit 11 converts coordinate data into molecular structure descriptors in the complex data acquired by the acquisition unit 10 and obtains molecular structure descriptors. Molecular structure descriptors are numerical representations of feature quantities such as molecular properties. Examples of molecular structure descriptors include Finger Print descriptors, RACs descriptors, and CoMFA descriptors.

[0037] The learning unit 11 performs a learning process using molecular structure descriptors as explanatory variables and physical property descriptors as target variables to generate a preliminary learning model. That is, the learning unit 11 learns the relationship between molecular structure descriptors and physical property descriptors to generate a preliminary learning model. The learning unit 11 may use one or more molecular structure descriptors as explanatory variables. The preliminary learning model outputs physical property descriptors for the input molecular structure descriptors. That is, the preliminary learning model estimates physical property descriptors based on molecular structure descriptors. The learning unit 11 stores the preliminary learning model obtained by repeating the multiple learning processes a predetermined number of times in the storage unit 12.

[0038] (Predictive learning model) The generation of a predictive learning model in the learning unit 11 is described below. The learning unit 11 performs a learning process using molecular structure descriptors and physical property descriptors as explanatory variables and complex properties as the target variable to generate a predictive learning model. That is, the learning unit 11 learns the relationship between molecular structure descriptors, physical property descriptors and complex properties to generate (construct) a predictive learning model. The learning unit 11 may use one or more molecular structure descriptors and one or more physical property descriptors as explanatory variables. The learning unit 11 uses the physical property descriptors output from the pre-learning model as the target variable. Specifically, the learning unit 11 inputs molecular structure descriptors into the pre-learning model and obtains the physical property descriptors output from the pre-learning model.

[0039] The predictive learning model outputs complex properties based on the input molecular structure descriptor and property descriptor. That is, the predictive learning model estimates complex properties based on the molecular structure descriptor and property descriptor. The learning unit 11 stores the predictive learning model obtained by repeating multiple learning processes a predetermined number of times in the storage unit 12.

[0040] The prediction unit 13 is a functional element that predicts the properties of complexes using a pre-trained predictive learning model. The prediction unit 13 inputs the input data into the pre-trained model. In response to the input data being input into the pre-trained model, the prediction unit 13 obtains a prediction result that includes the output values ​​output from the pre-trained model.

[0041] [How the prediction system works] (Learning Phase) The learning method (method for generating a preliminary learning model) will be explained with reference to Figure 3. Figure 3 is a flowchart showing an example of the learning method as processing flow S1. Processing flow S1 corresponds to the learning phase and is an example of the learning method according to this disclosure.

[0042] In step S11, the acquisition unit 10 acquires complex data from the database 20. In step S12, the learning unit 11 converts the coordinate data contained in the complex data into molecular structure descriptors and acquires molecular structure descriptors.

[0043] In step S13, the learning unit 11 performs machine learning using molecular structure descriptors and property descriptors. The learning unit 11 inputs the molecular structure descriptors and property descriptors into a machine learning model and obtains prediction results output from the machine learning model. Based on the error between the prediction results and the metal complex shown in the complex data (i.e., the correct answer), the learning unit 11 updates the parameters in the machine learning model. For example, methods such as backpropagation can be used to update the parameters.

[0044] In step S14, the learning unit 11 determines whether or not to terminate the learning process. If the machine learning termination conditions are met, the learning unit 11 terminates the learning process; otherwise, it continues the machine learning process. The termination conditions can be set arbitrarily. For example, the termination conditions may be set based on the error, or based on the number of complex data to be processed, i.e., the number of learning iterations.

[0045] If learning is to continue (step S14: NO), the learning unit 11 retrieves the next complex data from the database 20 and performs the processing from step S13 onwards for that complex data. If learning is to be terminated (step S14: YES), in step S15, the learning unit 11 retrieves the trained model. In this way, during the learning phase, the machine learning system 1 generates a preliminary trained model by performing machine learning using molecular structure descriptors and physical property descriptors.

[0046] The learning method (method for generating a predictive learning model) will be explained with reference to Figure 4. Figure 4 is a flowchart showing an example of the learning method as processing flow S2. Processing flow S2 corresponds to the learning phase and is an example of the learning method related to this disclosure.

[0047] In step S21, the acquisition unit 10 acquires complex data from the database 20 (acquisition step). In step S22, the learning unit 11 acquires molecular structure descriptors from the output of the pre-learned model.

[0048] In step S23, the learning unit 11 performs machine learning using molecular structure descriptors, physical property descriptors, and complex properties (learning step). The learning unit 11 inputs molecular structure descriptors, physical property descriptors, and complex properties into a machine learning model and obtains prediction results output from the machine learning model.

[0049] In step S24, the learning unit 11 determines whether or not to terminate the learning process. If the machine learning termination conditions are met, the learning unit 11 terminates the learning process; otherwise, it continues the machine learning process.

[0050] If learning is to continue (step S24: NO), the learning unit 11 retrieves the next complex data from the database 20 and performs the processing from step S23 onwards for that complex data. If learning is to be terminated (step S24: YES), in step S25, the learning unit 11 retrieves the trained model. In this way, during the learning phase, the machine learning system 1 generates a predictive learning model by performing machine learning using molecular structure descriptors, physical property descriptors, and complex properties.

[0051] (Operational Phase) The prediction method will be explained with reference to Figure 5. Figure 5 is a flowchart showing an example of the prediction method as processing flow S3. Processing flow S3 corresponds to the operation phase and is an example of the prediction method related to this disclosure.

[0052] In step S31, the acquisition unit 10 acquires the input data. In step S32, the prediction unit 13 inputs the input data into the predictive learning model and outputs the prediction result obtained by the predictive learning model. The method of outputting the prediction result by the prediction unit 13 is not particularly limited. For example, the prediction unit 13 may output the prediction result to the output device 106, store it in a predetermined database, or transmit it to another computer system.

[0053] [effect] As described above, the machine learning system 1 according to this embodiment learns the relationship between molecular structure descriptors and property descriptors and complex properties, and generates a predictive learning model that estimates complex properties in response to input of molecular structure descriptors and property descriptors. In this way, the machine learning system 1 generates a predictive learning model using property descriptors that are less dependent on the structure of the complex metal, in addition to molecular structure descriptors. As a result, the predictive learning model can accurately estimate complex properties in response to input of molecular structure descriptors and property descriptors, without depending on the structure of the complex metal. Therefore, the machine learning system 1 can improve the accuracy of predicting complex properties. For example, the machine learning system 1 can achieve a Mean Absolute Error (MAE) of 0.199V for predicting complex properties.

[0054] In the machine learning system 1 according to this embodiment, the learning unit 11 uses as property descriptors property descriptors property descriptors property descriptors estimated by a preliminary learning model that estimates property descriptor

[0055] In the machine learning system 1 according to this embodiment, the acquisition unit 10 acquires complex data including coordinate data and property descriptors of the complex. The learning unit 11 converts the coordinate data included in the complex data into molecular structure descriptors and acquires the property descriptors included in the complex data. Based on the converted molecular structure descriptors and property descriptors, the learning unit 11 learns the relationship between molecular structure descriptors and property descriptors and generates a preliminary learning model that estimates property descriptors in response to molecular structure descriptor input. With this configuration, a preliminary learning model that estimates property descriptors in response to molecular structure descriptor input can be appropriately generated. Furthermore, the preliminary learning model generated as described above can estimate property descriptors with high accuracy.

[0056] While embodiments of this disclosure have been described above, this disclosure is not necessarily limited to the embodiments described above, and various modifications are possible without departing from its essence.

[0057] In the above embodiment, a configuration in which the learning unit 11 acquires physical property descriptors using a pre-learned model was described as one example. However, physical property descriptors may also be acquired based on molecular structure descriptors by simulation, calculation formulas (quantum computation), etc.

[0058] In the above embodiment, an example was described in which the acquisition unit 10 acquires complex data from the database 20 and the learning unit 11 converts the coordinate data included in the complex data into molecular structure descriptors. However, the acquisition unit 10 may also acquire molecular structure descriptors from the database 20. That is, the database 20 may include molecular structure descriptors in the complex data. [Explanation of Symbols]

[0059] 1…Machine learning systems 10…Acquisition part 11…Learning Department 12...Storage section 13…Prediction Department 20…Database 100... Computer 101… Processor 102...Main memory section 103…Auxiliary storage unit 104...Communication Control Unit 105...Input device 106…Output device 110...Learning Program 120…Prediction Program S1…Processing flow S2…Processing flow S3…Processing flow S11...Step S12...Step S13...Step S14...Step S15... Step S21...Step S22... Step S23...Step S24... Step S25... Step S31... Step S32...Step

Claims

1. A learning device equipped with at least one processor, which performs machine learning on a learning model, An acquisition unit that acquires a molecular structure descriptor of the complex, a physical property descriptor of the complex, and complex properties that represent the physical properties of the complex, A learning device comprising: a learning unit that learns the relationship between the molecular structure descriptor, the physical property descriptor and the complex properties based on the molecular structure descriptor, the physical property descriptor and the complex properties, and generates a first learning model that estimates the complex properties in response to input of the molecular structure descriptor and the physical property descriptor.

2. The learning device according to claim 1, wherein the learning unit uses the physical property descriptor estimated by a second learning model that estimates the physical property descriptor based on the molecular structure descriptor as the physical property descriptor.

3. The acquisition unit acquires complex data including the coordinate data and the physical property descriptor of the complex, The learning device according to claim 2, wherein the learning unit converts the coordinate data included in the complex data into molecular structure descriptors, obtains the physical property descriptors included in the complex data, learns the relationship between the molecular structure descriptors and the physical property descriptors based on the converted molecular structure descriptors and the physical property descriptors, and generates a second learning model that estimates the physical property descriptors in response to the input of the molecular structure descriptors.

4. A learning method that runs on at least one processor and performs machine learning on a learning model, An acquisition step to obtain a molecular structure descriptor of the complex, a physical property descriptor of the complex, and complex properties that represent the physical properties of the complex, A learning method comprising: a learning step of generating a first learning model that learns the relationship between the molecular structure descriptor and the property descriptor and the complex properties based on the molecular structure descriptor, the property descriptor and the complex properties, and estimates the complex properties in response to inputs of the molecular structure descriptor and the property descriptor.

5. A trained model that causes at least one processor to estimate the properties of a complex, A trained model constructed by learning the relationship between the molecular structure descriptor and the property descriptor of the complex and the properties of the complex, which causes the processor to estimate the properties of the complex in response to the input of the molecular structure descriptor and the property descriptor.

6. A prediction device comprising at least one processor, which predicts the properties of a complex using the trained model described in claim 5, The processor of the prediction device Get the input data, A prediction device that outputs prediction results related to the properties of a complex by inputting the aforementioned input data into the trained model.

Images