Information processing program, information processing method, and information processing device

Abstract

Making it easier to train the model. [Solution] The information processing device 100 has a model 110 that generates output data representing the energy and force of a target molecule in response to input data representing the structure of the target molecule. For each structural data 101, the information processing device 100 acquires a plurality of first features relating to the structure of the target molecule in response to the input of the structural data 101 into the model 110. For each of the plurality of structural data 101, the information processing device 100 calculates a second feature relating to the interatomic distance in the structure of the target molecule represented by the structural data 101. Based on the distribution of coordinates in a multidimensional space 120 that represents a combination of the acquired plurality of first features and the calculated second features, the information processing device 100 selects one of the plurality of structural data 101 to be used when updating the model 110.

Description

Technical Field

[0001] The present invention relates to an information processing program, an information processing method, and an information processing apparatus.

Background Art

[0002] Conventionally, there is a technology of molecular dynamics simulation that uses the equations of motion to simulate the time evolution of the motion of each atom in an atomic system including N atoms that form a target molecule based on the potential related to the target molecule. Here, there is a technology for training a model representing the potential used in the molecular dynamics simulation. The model is a neural network. The model estimates, for example, the energy related to the target molecule and the force related to the target molecule from the structure of the target molecule.

[0003] As a prior art, for example, there is one that uses structure description information corresponding to a selected crystal structure candidate as input data, performs machine learning using information on the energy related to the selected crystal structure candidate as teacher data, and realizes a crystal form prediction device. [[ID=…]]

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, in the prior art, it may be difficult to efficiently train a model representing the potential. For example, if appropriate training data is not prepared, there is a problem that the accuracy of estimating the energy related to the target molecule and the force related to the target molecule from the structure of the target molecule in the trained model decreases.

[0006] [[ID=…]] In one aspect, the present invention aims to make it easier to train models. [Means for solving the problem]

[0007] According to one embodiment, an information processing program, information processing method, and information processing device are proposed that, for each of a plurality of structural data representing the structure of a target molecule, a plurality of first features relating to the structure of the target molecule are obtained in accordance with the result of inputting the structural data into a model that generates output data representing the energy of the target molecule in accordance with input data representing the structure of the target molecule, a second feature relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated for each of the plurality of structural data, and an information processing program, information processing method, and information processing device are proposed that, by referring to coordinates in a multidimensional space that represent a combination of the obtained plurality of first features and the calculated second features corresponding to each of the plurality of structural data, at least one of the plurality of structural data to be used when updating the model is selected based on the distribution of the coordinates. [Effects of the Invention]

[0008] In one embodiment, it becomes possible to make it easier to train the model. [Brief explanation of the drawing]

[0009] [Figure 1] Figure 1 is an explanatory diagram showing one embodiment of the information processing method according to the embodiment. [Figure 2] Figure 2 is an explanatory diagram showing an example of the information processing system 200. [Figure 3] Figure 3 is a block diagram showing an example of the hardware configuration of the information processing device 100. [Figure 4] Figure 4 is a block diagram showing an example of the functional configuration of the information processing device 100. [Figure 5] Figure 5 is an explanatory diagram showing an example of the operation of the information processing device 100. [Figure 6]Figure 6 is an explanatory diagram showing an example of MD calculation. [Figure 7] Figure 7 is an explanatory diagram illustrating an example of selecting structural data group 502. [Figure 8] Figure 8 is an explanatory diagram showing an example of the effect. [Figure 9] Figure 9 is a flowchart showing an example of the overall processing procedure. [Modes for carrying out the invention]

[0010] Embodiments of the information processing program, information processing method, and information processing apparatus according to the present invention will be described in detail below with reference to the drawings.

[0011] (An embodiment of the information processing method according to the embodiment) Figure 1 is an explanatory diagram showing one embodiment of the information processing method according to the embodiment. The information processing device 100 is a computer for facilitating the training of a model representing potential. The information processing device 100 is, for example, a server or a PC (Personal Computer). The model is, for example, a neural network. In the following description, a neural network may be referred to as "NN (Neural Network)".

[0012] Traditionally, molecular dynamics simulations have been performed using equations of motion to simulate the time evolution of the motion of each atom in an atomic system containing N atoms that make up a target molecule, based on the potential of the target molecule. In the following explanation, molecular dynamics simulations may be referred to as "MD (Molecular Dynamics) calculations."

[0013] Potential is a parameter that forms the equations of motion. The potential represents, for example, the interaction between atoms. Specifically, the potential realizes a function of estimating the energy and the force related to the target molecule from the structure of the target molecule. The potential is set by a user, for example. The potential set by the user is also called an empirical potential. On the other hand, it is conceivable to represent the potential using a model. The model is, for example, a neural network (NN). The model representing the potential is also called a machine learning potential, for example.

[0014] Here, the model representing the potential may be trained. For example, the model may be trained based on training data that has the structure of the target molecule as an input sample and the energy and the force related to the target molecule as labels indicating the correct answers of the output. The training is carried out by, for example, backpropagation of error.

[0015] Specifically, when training the model, it is conceivable to prepare training data by performing density functional theory calculations on the structural data representing the structure of the target molecule, identifying the energy and the force related to the target molecule, and assigning labels. The labels indicate the energy and the force related to the identified target molecule. In the following description, density functional theory may be denoted as "DFT (Density Functional Theory)".

[0016] By preparing training data through DFT calculations, it is expected that the accuracy of the model will approach the accuracy of DFT calculations. The accuracy of the model indicates how accurately the energy and the force related to the target molecule can be estimated from the structure of the target molecule in the model.

[0017] However, if appropriate training data is not prepared, there is a problem of reducing the accuracy of the trained model. Also, in order to improve the accuracy of the model, there is a problem that increasing the number of training data to be prepared leads to an increase in the required time for training the model. Therefore, it is desired to improve the accuracy of the model while reducing the required time for training the model by selectively preparing appropriate training data.

[0018] On the other hand, for example, based on two-dimensional feature amounts regarding each of a plurality of structure data, one or more structure data to be labeled are selected from the plurality of structure data, and a first method for facilitating the preparation of appropriate training data can be considered. The label indicates the correct answer of the output, and indicates the energy regarding the target molecule and the force regarding the target molecule. Specifically, it is conceivable to arrange a plurality of structure data in a two-dimensional space with the two-dimensional feature amounts as coordinates, and uniformly select one or more structure data in the two-dimensional space. The two-dimensional feature amounts can be obtained, for example, by performing principal component analysis on a plurality of feature amounts in an intermediate layer of the NN when the structure data is input to the NN. This first method is also called, for example, active learning. Regarding active learning, for example, Reference Document 1 below can be referred to.

[0019] Reference Document 1: Shimizu, Koji, et al. “Phase stability of Au-Li binary systems studied using neural network potential.” Physical Review B 103.9 (2021): 094112.

[0020] Even with this first method, it can be difficult to efficiently train a model that represents the potential. For example, the model tends to have difficulty accurately estimating the energy and forces related to a target molecule for structures that are not represented by the training data. Here, in the first method, one or more structural data are uniformly selected in a two-dimensional space without considering the distance between atoms, so the accuracy of the model may be low for target molecule structures where the distance between atoms is small. Then, when performing the MD calculation, if the accuracy of the model is low, the distance between atoms may become too small, causing atoms to be ejected from the atomic system due to interatomic repulsion, which could lead to the MD calculation failing.

[0021] Furthermore, to account for cases where the distance between atoms is small, a second method can be considered, which extends the first method by preparing multiple structural data with reduced atomic system size, and then selecting one or more structural data from among them to assign labels. This second method is also called a Volume Scan. For more information on Volume Scans, see, for example, reference 2 below.

[0022] Reference 2: Magdau, Ioan-Bogdan, et al. “Machine learning force fields for molecular liquids: Ethylene Carbonate / Ethyl Methyl Carbonate binary solvent.” npj Computational Materials 9.1 (2023): 146.

[0023] Even with this second method, it can be difficult to efficiently train the model representing the potential. Similar to the first method, the second method uniformly selects one or more structural data points in a two-dimensional space without considering interatomic distances. Therefore, the model accuracy may be low for the structure of the target molecule where interatomic distances are small. Furthermore, during MD calculations, the low accuracy of the model can lead to interatomic distances becoming too small, causing atoms to be ejected from the atomic system due to interatomic repulsion, potentially resulting in the MD calculation failing.

[0024] Therefore, in this embodiment, we will describe an information processing method that can facilitate the training of a model representing potential.

[0025] In Figure 1, the information processing device 100 has a model 110. The model 110 generates output data representing the energy and force of a target molecule, based on input data representing the structure of the target molecule. The input data is, for example, structure data 101. The model 110 is, for example, a neural network. The model 110 is the object to be trained. Training corresponds to, for example, updating the model 110.

[0026] The information processing device 100 stores a plurality of structural data 101. Preferably, each of the plurality of structural data 101 represents, for example, the structure of the target molecule in a reduced system. Alternatively, each of the plurality of structural data 101 may represent, for example, the structure of the target molecule in a non-reduced system. The structural data 101 may further represent, for example, the size of the simulation box, the presence or absence of periodic boundary conditions, or the presence or absence of stress. The structural data 101 is generated, for example, using a Volume Scan. The structural data 101 may also be generated using methods other than Volume Scan. The structural data 101 may also be generated, for example, using a pressure setting for the system. The information processing device 100 may generate and store the plurality of structural data 101, for example, using a Volume Scan.

[0027] (1-1) For each structural data 101, the information processing device 100 obtains multiple first features related to the structure of the target molecule, depending on the result of inputting the structural data 101 into the model 110. For example, for each structural data 101, the information processing device 100 extracts multiple features in the intermediate layer of the model 110 when the structural data 101 is input into the model 110. For example, for each structural data 101, the information processing device 100 obtains multiple first features by performing principal component analysis on the extracted multiple features and calculating multiple first features. For example, there are two first features. This allows the information processing device 100 to obtain a guideline for selecting at least one of the multiple structural data 101 to be used when updating the model 110.

[0028] (1-2) The information processing device 100 calculates a second feature for each of the multiple structural data 101, relating to the interatomic distances in the structure of the target molecule represented by the structural data 101. The second feature is, for example, a statistical value of the interatomic distances in the structure of the target molecule represented by the structural data 101. The statistical value is, for example, the minimum value, maximum value, mean value, median value, or mode value. This allows the information processing device 100 to obtain a guideline for selecting at least one of the multiple structural data 101 to be used when updating the model 110. Specifically, the information processing device 100 can select structural data 101 that corresponds to cases where the interatomic distances are relatively close, which may be a factor in the failure of the MD calculation, as the structural data 101 to be used when updating the model 110.

[0029] (1-3) The information processing device 100 identifies coordinates in a multidimensional space 120 that represent a combination of acquired first features and calculated second features corresponding to each of the multiple structural data 101. The multidimensional space 120 is, for example, a three-dimensional space. The information processing device 100 refers to the identified coordinates and, based on the distribution of those coordinates in the multidimensional space 120, selects at least one of the multiple structural data 101 to be used when updating the model 110. For example, the information processing device 100 selects one or more structural data 101 uniformly in the multidimensional space 120, not only from the perspective of the first features but also from the perspective of second features relating to the distance between atoms.

[0030] This allows the information processing device 100 to select appropriate structural data 101 to use when updating the model 110 in order to improve the accuracy of the model 110. Here, for example, in conventional methods, structural data 101, shown by the thick circle, is selected based on multiple first features without considering the second feature, and it may not be possible to select structural data 101 corresponding to cases where the distance between atoms is relatively close. In contrast, the information processing device 100 can select structural data 101 corresponding to cases where the distance between atoms is relatively close, contained in region 121 of the multidimensional space 120. Therefore, the information processing device 100 can efficiently update the model 110. The information processing device 100 can, for example, reduce the time required when updating the model 110. The information processing device 100 can, for example, improve the accuracy of the updated model 110.

[0031] (1-4) The information processing device 100 updates the model 110 based on at least one of the selected structural data 101. The information processing device 100, for example, performs a DFT calculation on at least one of the selected structural data 101 and generates training data that associates the selected structural data 101 with labels indicating the correct output. The labels indicate, for example, the energy and the force relating to the target molecule.

[0032] The information processing device 100 updates the model 110, for example, based on the generated training data. Specifically, the information processing device 100 updates the parameters of the model 110 so that the output of the model 110 when the structured data 101 from the generated training data is input to the model 110 approaches the correct output indicated by the labels in the generated training data. In this way, the information processing device 100 can efficiently update the model 110 to improve its accuracy.

[0033] Furthermore, the information processing device 100 can prepare multiple structural data 101 such that each structural data 101 represents the structure of the target molecule in the reduced system. Therefore, the information processing device 100 can update the model 110 using structural data 101 that corresponds to irregular cases where the distance between atoms is relatively close, which can be a cause of MD calculation failure. As a result, the information processing device 100 can easily improve the accuracy of the model 110.

[0034] Here, we have described the case where the functions of the information processing device 100 are realized by a single computer, but this is not the only case. For example, the functions of the information processing device 100 may be realized through the collaboration of multiple computers. Also, for example, the functions of the information processing device 100 may be realized on the cloud.

[0035] Furthermore, for example, the information processing device 100 may collaborate with other computers. Specifically, the information processing device 100 may implement a function to select structured data 101 and collaborate with other computers that prepare the structured data 101, other computers that prepare training data, and other computers that update the model 110. The case in which the information processing device 100 collaborates with other computers will be described later, for example, using Figure 2.

[0036] (An example of information processing system 200) Next, using Figure 2, we will describe an example of an information processing system 200 to which the information processing device 100 shown in Figure 1 is applied.

[0037] Figure 2 is an explanatory diagram showing an example of an information processing system 200. In Figure 2, the information processing system 200 includes an information processing device 100, a training processing device 201, and a client device 202.

[0038] In the information processing system 200, the information processing device 100 and the training processing device 201 are connected via a wired or wireless network 210. The network 210 can be, for example, a LAN (Local Area Network), a WAN (Wide Area Network), or the Internet. Also in the information processing system 200, the information processing device 100 and the client device 202 are connected via a wired or wireless network 210.

[0039] The information processing device 100 is a computer designed to facilitate updating the model representing the potential. The information processing device 100 receives an update request that demands updating the model. In response to the update request, the information processing device 100 sends a generation request to the training processing device 201 that demands the generation of multiple structural data. The information processing device 100 receives the multiple structural data corresponding to the generation request from the training processing device 201.

[0040] The information processing device 100 obtains multiple first features related to the structure of the target molecule for each of the multiple structural data, assuming that the structural data is input into the model. The information processing device 100 sends a calculation request to the training processing device 201, for example, requesting the calculation of the first features. The information processing device 100 receives multiple first features corresponding to the calculation request from the training processing device 201.

[0041] The information processing device 100 calculates a second feature for each of the multiple structural data sets, relating to the interatomic distances in the structure of the target molecule represented by that structural data set. Based on the multiple first features obtained and the calculated second features corresponding to each of the multiple structural data sets, the information processing device 100 selects one or more structural data sets from the multiple structural data sets to be used when updating the model. The information processing device 100 sends an update request to the training processing device 201, which includes the one or more selected structural data sets and requests that the model be updated.

[0042] The information processing device 100 receives an execution request from the client device 202 requesting that it perform an MD calculation. The information processing device 100 transmits the received execution request to the training processing device 201. The information processing device 100 receives the result of the MD calculation from the training processing device 201. The information processing device 100 transmits the result of the MD calculation to the client device 202. The information processing device 100 is, for example, a server or a PC.

[0043] The training processing unit 201 is a computer with a model. Upon receiving a generation request, the training processing unit 201 uses the model to perform MD calculations on the target molecule, generating multiple structural data sets and transmitting them to the information processing unit 100. Upon receiving a calculation request, the training processing unit 201 calculates multiple first features for each of the multiple structural data sets and transmits them to the information processing unit 100.

[0044] When the training processing unit 201 receives an update request, it generates training data by performing a DFT calculation on each of the one or more structural data included in the update request and assigning labels to them. The training processing unit 201 updates the model based on the generated training data. When the training processing unit 201 receives an execution request, it performs an MD calculation using the model and sends the results to the information processing unit 100. The training processing unit 201 is, for example, a server or a PC.

[0045] The client device 202 is a computer owned by the system user. The client device 202 generates an execution request requesting to perform an MD calculation and sends it to the information processing device 100. The client device 202 receives the results of the MD calculation from the information processing device 100. The client device 202 outputs the results of the MD calculation so that the system user can refer to them. The client device 202 may be, for example, a PC, a tablet terminal, or a smartphone.

[0046] This explanation describes a case where the information processing device 100 is a different device from the training device, but it is not limited to this case. For example, the information processing device 100 may have the function of a training device and may operate as a training device. In this case, the information processing system 200 does not need to include a training device.

[0047] This explanation describes a case where the information processing device 100 is a different device from the client device 202, but it is not limited to this case. For example, the information processing device 100 may have the functionality of a client device 202 and may operate as a client device 202. In this case, the information processing system 200 does not need to include a client device 202.

[0048] (Example of hardware configuration of information processing device 100) Next, an example of the hardware configuration of the information processing device 100 will be described using Figure 3.

[0049] Figure 3 is a block diagram showing an example of the hardware configuration of the information processing device 100. In Figure 3, the information processing device 100 includes a CPU (Central Processing Unit) 301, a memory 302, and a network interface 303. The information processing device 100 also includes a recording medium interface 304, a recording medium 305, a display 306, and an input device 307. Each component is connected by a bus 300.

[0050] Here, the CPU 301 is responsible for the overall control of the information processing device 100. The memory 302 includes, for example, ROM (Read Only Memory), RAM (Random Access Memory), and flash ROM. Specifically, for example, flash ROM and ROM store various programs, and RAM is used as the work area for the CPU 301. Programs stored in memory 302 are loaded into the CPU 301, causing the CPU 301 to execute the coded processes.

[0051] The network interface 303 is connected to network 210 via a communication line, and then connects to other computers via network 210. The network interface 303 manages the internal interface with network 210 and controls the input and output of data from other computers. The network interface 303 is, for example, a modem or a LAN adapter.

[0052] The recording medium interface (I / F) 304 controls the reading and writing of data to the recording medium 305 according to the control of the CPU 301. The recording medium interface (I / F) 304 is, for example, a disk drive, an SSD (Solid State Drive), or a USB (Universal Serial Bus) port. The recording medium 305 is a non-volatile memory that stores the data written under the control of the recording medium interface (I / F) 304. The recording medium 305 is, for example, a disk, semiconductor memory, or USB memory. The recording medium 305 may be detachable from the information processing device 100.

[0053] Display 306 displays data such as cursors, icons, toolboxes, documents, images, or functional information. Display 306 is, for example, a CRT (Cathode Ray Tube), a liquid crystal display, or an organic EL (Electroluminescence) display. Input device 307 has keys for inputting characters, numbers, or various instructions, and performs data input. Input device 307 is, for example, a keyboard or a mouse. Input device 307 may also be, for example, a touch panel input pad or a numeric keypad.

[0054] The information processing device 100 may have, in addition to the components described above, a camera, for example. Furthermore, the information processing device 100 may have, in addition to the components described above, a printer, scanner, microphone, or speaker, for example. Also, the information processing device 100 may have multiple recording medium interfaces 304 and recording mediums 305, for example. Furthermore, the information processing device 100 does not necessarily have, for example, a display 306 or an input device 307. Also, the information processing device 100 does not necessarily have, for example, recording medium interfaces 304 and recording mediums 305.

[0055] (Example hardware configuration of training processing device 201) The hardware configuration example of the training processing device 201 is the same as the hardware configuration example of the information processing device 100 shown in Figure 3, so the explanation is omitted.

[0056] (Example hardware configuration for client device 202) The hardware configuration example for client device 202 is the same as the hardware configuration example for information processing device 100 shown in Figure 3, so the explanation is omitted.

[0057] (Example of the functional configuration of the information processing device 100) Next, an example of the functional configuration of the information processing device 100 will be described using Figure 4.

[0058] Figure 4 is a block diagram showing an example of the functional configuration of the information processing device 100. The information processing device 100 includes a storage unit 400, an acquisition unit 401, a generation unit 402, a first calculation unit 403, a second calculation unit 404, a selection unit 405, a training unit 406, and an output unit 407.

[0059] The storage unit 400 is implemented by a storage area such as the memory 302 or recording medium 305 shown in Figure 3. The following description will focus on the case where the storage unit 400 is included in the information processing device 100, but is not limited to this case. For example, the storage unit 400 may be included in a device different from the information processing device 100, and the contents of the storage unit 400 may be accessible from the information processing device 100.

[0060] The acquisition unit 401 to the output unit 407 function as an example of a control unit. Specifically, the acquisition unit 401 to the output unit 407 realize their functions, for example, by having the CPU 301 execute a program stored in a storage area such as the memory 302 or recording medium 305 shown in Figure 3, or by using the network I / F 303. The processing results of each functional unit are stored in a storage area such as the memory 302 or recording medium 305 shown in Figure 3.

[0061] The memory unit 400 stores various information that is referenced or updated during the processing of each functional unit. The memory unit 400 stores, for example, a model. The model has the function of generating output data representing the energy and forces of a target molecule, in response to input data representing the structure of the target molecule. The structure includes, for example, the positions of each atom of the multiple atoms that make up the target molecule. The structure also includes, for example, the velocities of each atom of the multiple atoms that make up the target molecule.

[0062] The model is, for example, a neural network (NN). Specifically, the model is an NN generated based on one or more training data. The training data represents, for example, structural data representing an input sample and a label that is the correct output. The structural data represents the structure of the target molecule. The label indicates the energy and force related to the target molecule. The model is, for example, pre-configured by the user. The model may be acquired, for example, by the acquisition unit 401. The model is updated, for example, by the training unit 406.

[0063] The storage unit 400 stores, for example, multiple structural data. Preferably, each of the multiple structural data represents the structure of the target molecule in a reduced system. The structure includes, for example, the position of each atom of the multiple atoms that make up the target molecule. The structure also includes, for example, the velocity of each atom of the multiple atoms that make up the target molecule. Each of the multiple structural data may represent the structure of the target molecule in a non-reduced system. The structural data is generated, for example, by the generation unit 402. The structural data may be acquired, for example, by the acquisition unit 401.

[0064] The memory unit 400 stores, for example, one or more training data. The training data is used when updating the model. Each of the one or more training data represents a structural data representing an input sample and a label representing the correct output. The structural data represents the structure of the target molecule. Specifically, the label indicates the energy and force related to the target molecule, corresponding to the structure of the target molecule represented by the structural data. The training data is generated, for example, by the training unit 406.

[0065] The acquisition unit 401 acquires various types of information used in the processing of each functional unit. The acquisition unit 401 stores the acquired information in the storage unit 400 or outputs it to each functional unit. The acquisition unit 401 may also output the information stored in the storage unit 400 to each functional unit. The acquisition unit 401 acquires various types of information, for example, based on user input. The acquisition unit 401 may also receive various types of information from a device other than the information processing device 100, for example.

[0066] The acquisition unit 401 acquires, for example, an update request that requests updating a model. The update request may specify, for example, a target molecule. The update request may include, for example, a model. The update request may include, for example, multiple structural data. Specifically, the acquisition unit 401 acquires an update request by accepting the input of an update request. Specifically, the acquisition unit 401 may acquire an update request by receiving an update request from another computer. The other computer is, for example, a client device 202.

[0067] The acquisition unit 401 acquires a model, for example. Specifically, the acquisition unit 401 acquires a model by accepting a model as input. Specifically, the acquisition unit 401 may acquire a model by receiving a model from another computer. The other computer may be, for example, a client device 202. Specifically, the acquisition unit 401 may acquire a model by extracting a model from an update request.

[0068] The acquisition unit 401 acquires, for example, multiple structural data. Specifically, the acquisition unit 401 acquires multiple structural data by receiving input of multiple structural data. Specifically, the acquisition unit 401 may acquire multiple structural data by receiving multiple structural data from another computer. The other computer is, for example, a client device 202. Specifically, the acquisition unit 401 may acquire multiple structural data by extracting multiple structural data from an update request.

[0069] The acquisition unit 401 may receive a start trigger to initiate processing in any of the functional units. A start trigger may be, for example, a predetermined operation input by a user. A start trigger may also be, for example, the receipt of predetermined information from another computer. A start trigger may also be, for example, the output of predetermined information by any of the functional units. The acquisition unit 401 accepts, for example, the acquisition of an update request as a start trigger to initiate processing in the generation unit 402, the first calculation unit 403, the second calculation unit 404, the selection unit 405, and the training unit 406.

[0070] The generation unit 402 obtains multiple structural data by generating multiple structural data. The generation unit 402 generates multiple structural data, for example, by performing MD calculations using a model after reducing the system. The MD calculations may be coarse-grained. Specifically, the generation unit 402 randomly sets multiple initial structures for the target molecule. Specifically, for each set initial structure, the generation unit 402 generates multiple structural data representing the structure of the target molecule by numerically integrating the equations of motion and simulating the time evolution of that initial structure. The generation unit 402 may also generate multiple structural data, for example, by performing MD calculations using a model without reducing the system. In this way, the generation unit 402 can prepare multiple structural data to serve as the basis for training data.

[0071] The first calculation unit 403 calculates multiple first features related to the structure of the target molecule for each of the multiple structural data, based on the results of inputting the structural data into the model. For example, the first calculation unit 403 calculates multiple first features related to the structure of the target molecule by performing principal component analysis on the output of the intermediate layer of the model when each of the multiple structural data is input into the model. There are, for example, two first features. This allows the first calculation unit 403 to obtain guidance on selecting at least one of the multiple structural data to use when updating the model.

[0072] The second calculation unit 404 calculates a second feature quantity for each of the multiple structural data, relating to the interatomic distances in the structure of the target molecule represented by that structural data. The second feature quantity is, for example, a statistical value of the interatomic distances in the structure of the target molecule represented by the structural data. The statistical value is, for example, the minimum value, maximum value, mean value, median value, or mode value. This allows the second calculation unit 404 to obtain guidance for selecting at least one of the multiple structural data to use when updating the model.

[0073] The selection unit 405 sets coordinates in a multidimensional space that represent the combination of multiple acquired first features and calculated second features corresponding to each of the multiple structural data. Based on the distribution of the set coordinates, the selection unit 405 selects at least one of the multiple structural data to be used when updating the model.

[0074] The selection unit 405 calculates, for example, the distance between the coordinates corresponding to each of the multiple structural data and the coordinates corresponding to other structural data, based on the set coordinate distribution. The distance is, for example, the Euclidean distance. The selection unit 405 selects a predetermined number of structural data from among the multiple structural data, starting with those whose calculated distances from other structural data are relatively large.

[0075] Specifically, the selection unit 405 selects a predetermined number of structural data from among multiple structural data, in order from those with the largest statistical distance values ​​from other calculated structural data. These statistical values ​​may include, for example, the minimum, maximum, mean, median, or mode. This allows the selection unit 405 to select appropriate structural data to use when updating the model in order to improve its accuracy.

[0076] The training unit 406 updates the model. The training unit 406 calculates the energy and force related to the target molecule by, for example, performing DFT calculations on each structural data selected by the selection unit 405, and sets a label corresponding to the structural data. The training unit 406 generates training data that associates each structural data selected by the selection unit 405 with the set label. The training unit 406 updates the model based on the generated training data.

[0077] Specifically, the training unit 406 updates the model parameters so that the model's output, when the structural data from the generated training data is input to the model, approaches the labels from the generated training data. This allows the training unit 406 to efficiently update the model to improve its accuracy.

[0078] Furthermore, the information processing device 100 may repeatedly perform a series of processes to update the model using the generation unit 402, the first calculation unit 403, the second calculation unit 404, the selection unit 405, and the training unit 406. The information processing device 100 may, for example, repeatedly perform the series of processes until a predetermined termination condition is met. The predetermined termination condition may be, for example, that the series of processes is performed a predetermined number of times. The predetermined termination condition may also be, for example, that the amount of change in the model parameters before and after the update falls below a threshold. In this case, the generation unit 402 generates, for example, a plurality of new structural data representing the structure of the target molecule in the reduced system.

[0079] The first calculation unit 403 obtains, for example, multiple first feature quantities relating to the structure of the target molecule for each of the new multiple structural data, according to the result of inputting the structural data into the previously updated model. The second calculation unit 404 calculates, for example, second feature quantities relating to the interatomic distances in the structure of the target molecule represented by the structural data for each of the new multiple structural data.

[0080] The selection unit 405 sets coordinates in a multidimensional space that represent, for example, combinations of acquired first features and calculated second features corresponding to each of the new multiple structural data. Based on the distribution of the set coordinates, the selection unit 405 selects at least one of the new multiple structural data to be used when updating the model.

[0081] The training unit 406 sets labels corresponding to each structural data selected by the selection unit 405, for example, by performing a DFT calculation on each of the structural data selected by the selection unit 405. The training unit 406 generates training data that associates each of the structural data selected by the selection unit 405 with the set labels. The training unit 406 updates the model based on the generated training data, for example. This allows the information processing device 100 to repeatedly update the model to improve its accuracy.

[0082] The output unit 407 outputs the processing result of at least one of the functional units. The output format can be, for example, display on a screen, print to a printer, transmit to an external device via the network interface 303, or store in a storage area such as the memory 302 or recording medium 305. This allows the output unit 407 to notify the user of the processing result of at least one of the functional units, thereby improving the usability of the information processing device 100.

[0083] The output unit 407 outputs, for example, the structural data selected by the selection unit 405. Specifically, the output unit 407 outputs the structural data selected by the selection unit 405 in a way that is accessible to the user. Specifically, the output unit 407 transmits the structural data selected by the selection unit 405 to another computer. The other computer is, for example, the client device 202. This allows the output unit 407 to provide appropriate structural data to be used when updating the model, and enables the model update process to be performed externally.

[0084] The output unit 407 outputs, for example, the training data generated by the training unit 406. Specifically, the output unit 407 outputs the training data generated by the training unit 406 so that it can be viewed by the user. Specifically, the output unit 407 transmits the training data generated by the training unit 406 to another computer. The other computer is, for example, the client device 202. This allows the output unit 407 to provide appropriate training data to be used when updating the model, and enables the model update process to be performed externally.

[0085] The output unit 407 outputs, for example, the model updated by the training unit 406. Specifically, the output unit 407 outputs the model updated by the training unit 406 so that it can be viewed by the user. Specifically, the output unit 407 transmits the model updated by the training unit 406 to another computer. The other computer is, for example, the client device 202. This allows the output unit 407 to make an improved model, which has improved accuracy in estimating the energy of the target molecule, available externally.

[0086] Here, we have described a case in which the information processing device 100 includes a generation unit 402, a first calculation unit 403, a second calculation unit 404, a selection unit 405, and a training unit 406, but it is not limited to this. For example, the information processing device 100 may not include any of the functional units. Specifically, the information processing device 100 may not include the generation unit 402. In this case, the information processing device 100 may cooperate with another computer that includes the generation unit 402.

[0087] (An example of the operation of the information processing device 100) Next, an example of the operation of the information processing device 100 will be explained using Figures 5 to 8.

[0088] Figure 5 is an explanatory diagram showing an example of the operation of the information processing device 100. In Figure 5, the information processing device 100 stores a machine learning potential 500. The machine learning potential 500 is a neural network (NN). The machine learning potential 500 has the function of generating output data representing the energy and forces related to the target molecule, etc., in response to input data representing the structure of the target molecule.

[0089] (5-1) The information processing device 100 generates a set of structural data 501 which includes multiple structural data for which the energy and forces of the target molecule are unknown, by performing an MD calculation using the machine learning potential 500. The set of structural data 501 may also include multiple structural data for which the energy and forces of the target molecule are known, for example. Now, we will move on to the explanation of Figure 6 and describe an example of an MD calculation.

[0090] Figure 6 is an explanatory diagram showing an example of MD calculation. In Figure 6, the information processing device 100 stores the equations of motion using the machine learning potential 500. (6-1) The information processing device 100 sets the initial structure of the target molecule with a reduced system size using Volume Scan. The system size may be reduced, for example, by setting the pressure on the system. The initial structure includes, for example, the coordinates r of each atom and the velocity v of each atom of the multiple atoms that make up the target molecule. The information processing device 100 sets the initial structure of the target molecule randomly, for example.

[0091] The information processing device 100 performs the following series of processes on the initial structure of the target molecule: (6-2) The information processing device 100 calculates the force f acting on the atoms based on the structure of the target molecule and a constant relating to the force f acting on the atoms. (6-3) The information processing device 100 updates the coordinate r of each atom and the velocity v of each atom by numerically integrating the equation of motion, and updates the structure of the target molecule.

[0092] The information processing device 100 simulates the time evolution of the target molecule by repeatedly performing the series of processes described in (6-2) and (6-3) above on the structure of the target molecule that was updated immediately before, a predetermined number of times. By repeatedly performing the series of processes a predetermined number of times, the information processing device 100 generates structural data representing the structure of the target molecule that was last updated. The information processing device 100 sets the generated structural data, which is obtained when each of several randomly set different structures is adopted as the initial structure, into the structural data group 501.

[0093] Returning to the explanation of Figure 5, (5-2) the information processing device 100 selects structure data group 502 from structure data group 501 to be used when updating the machine learning potential 500. Now, moving on to the explanation of Figure 7, we will describe an example in which the information processing device 100 selects structure data group 502 from structure data group 501.

[0094] Figure 7 is an explanatory diagram illustrating an example of selecting structural data group 502. In the example in Figure 7, structural data group 501 is assumed to contain multiple structural data for which, for example, the energy and forces related to the target molecule are unknown and unlabeled. The labels indicate the energy and forces related to the target molecule.

[0095] Furthermore, the structural data group 501 includes multiple structural data for which the energy and forces related to the target molecule are known and have been labeled. In Figure 7, unlabeled structural data corresponds to pg10 (unlabeled). Labeled structural data corresponds to pg10 (labeled).

[0096] (7-1) For each structural data in the structural data group 501, the information processing device 100 acquires multiple features that are output by the intermediate layer of the machine learning potential 500 when the structural data is input to the machine learning potential 500. For each structural data in the structural data group 501, the information processing device 100 calculates feature 1 and feature 2 related to the structure of the target molecule by performing principal component analysis based on the acquired multiple features.

[0097] Graph 700 in Figure 7 shows a two-dimensional space with axes corresponding to feature quantity 1 and feature quantity 2, respectively, which relate to the structure of the target molecule. Graph 700 represents the distribution of each structural data in the two-dimensional space, where feature quantity 1 and feature quantity 2, corresponding to each structural data, are used as coordinate values ​​on each axis.

[0098] (7-2) For each structural data in the structural data group 501, the information processing device 100 calculates the minimum interatomic distance in the structure of the target molecule represented by the structural data as feature quantity 3. It is preferable that feature quantity 3 represents the minimum distance between atom O and atom H. Graph 710 in Figure 7 shows a three-dimensional space having axes corresponding to feature quantity 1, feature quantity 2, and feature quantity 3, respectively. Graph 710 shows the distribution of each structural data in the three-dimensional space when feature quantity 1, feature quantity 2, and feature quantity 3, corresponding to each structural data, are used as coordinate values ​​on each axis.

[0099] (7-3) The information processing device 100 calculates a statistical value of the Euclidean distance between each structural data in the structural data group 501 and other structural data in three-dimensional space. The statistical value may be, for example, the minimum value, maximum value, mean value, median value, or mode value. In the example in Figure 7, the statistical value is specifically the maximum value. The information processing device 100 selects a predetermined number of structural data from the structural data group 501, in order of the structural data with the relatively large calculated statistical value, thereby selecting a structural data group 502 containing a predetermined number of structural data.

[0100] This allows the information processing device 100 to select a set of structure data 502 containing appropriate structure data to improve the accuracy of the machine learning potential 500. For example, the information processing device 100 can select the set of structure data 502 to include a balanced amount of structure data that addresses cases where the distance between atoms is relatively close, which can be a factor in the failure of MD calculations. Therefore, the information processing device 100 can efficiently update the machine learning potential 500 to improve its accuracy.

[0101] Returning to the explanation of Figure 5, (5-3) the information processing device 100 performs a DFT calculation on each of the selected structural data in the structural data group 502 to calculate the energy and force related to the target molecule and set a label. The label indicates the energy and force related to the target molecule. The information processing device 100 may also calculate the energy and force related to the target molecule and set a label using a method other than DFT calculation.

[0102] (5-4) For each of the selected structural data in the structural data group 502, the information processing device 100 generates training data in which the structural data is represented as an input sample and the set label is represented as the output correct answer. The information processing device 100 then compiles the generated training data to prepare a training data group 503. The information processing device 100 registers the training data group 503 into the dataset 510. The dataset 510 is implemented by the memory 302.

[0103] (5-5) The information processing device 100 refers to the dataset 510 and updates the machine learning potential 500 based on the training data set 503. The update is achieved, for example, by backpropagation. The information processing device 100 updates the parameters of the machine learning potential 500 so that the output of the machine learning potential 500 corresponding to the input of structured data contained in each training data approaches the label contained in that training data. In this way, the information processing device 100 can improve the accuracy of the machine learning potential 500.

[0104] The information processing device 100 may repeatedly perform the series of processes described in (5-1), (5-2), (5-3), (5-4), and (5-5) above. This allows the information processing device 100 to improve the accuracy of the machine learning potential 500. Next, we will move on to the explanation of Figure 8 and describe an example of the effects of the information processing device 100.

[0105] Figure 8 is an explanatory diagram illustrating an example of the effect. In Figure 8, Table 800 shows the results of comparing the method by the information processing device 100 with the first conventional method, the second conventional method, and the third conventional method. In the following explanation, the method by the information processing device 100 may be referred to as the "proposed method." The first conventional method selects 11,597 structural data to be labeled from among multiple structural data based on the two-dimensional feature quantities of each structural data, without reducing the size of the system, and prepares 11,597 training data.

[0106] The second conventional method, like the first conventional method, selects 11,597 structure data and prepares 11,597 training data. Then, based on the 2D features of each structure data from the multiple structure data with reduced system size, the second conventional method selects an additional 1,000 structure data to be labeled and prepares an additional 1,000 training data. Therefore, the second conventional method prepares a total of 12,597 training data, including the 11,597 training data prepared in the same way as the first conventional method and the new 1,000 training data.

[0107] The third conventional method, like the first conventional method, selects 11,597 structure data and prepares 11,597 training data. Then, based on the 2D features of each of the multiple structure data sets with reduced system size, the third conventional method selects 500 structure data sets to be labeled and prepares an additional 500 training data sets. Therefore, the third conventional method prepares a total of 12,097 training data sets, including the 11,597 training data sets prepared in the same way as the first conventional method and the new 500 training data sets.

[0108] The information processing device 100 selects 11,597 structural data and prepares 11,597 training data, similar to the first conventional method. Then, based on the three-dimensional features of each of the multiple structural data with reduced system size, the information processing device 100 selects 500 structural data to be labeled and prepares an additional 500 training data. As mentioned above, the three-dimensional features include features related to the distance between atoms. Specifically, the distance between atoms is the OH-OH distance. O is oxygen, and H is hydrogen. Therefore, the information processing device 100 prepares a total of 12,097 training data, including the 11,597 training data prepared in the same way as the first conventional method and the new 500 training data.

[0109] Table 800 shows the number of structured data (number of additional data), the histogram of OH-to-OH distance [distance: Å], and the MD computation time [ns] for the proposed method, the first conventional method, the second conventional method, and the third conventional method. The number of structured data is the total number of selected structured data. The number of structured data corresponds to the total number of training data prepared. The number of additional data is the total number of additionally selected structured data. The number of additional data corresponds to the total number of additionally prepared training data. The histogram of OH-to-OH distance [distance: Å] is a frequency distribution table based on the OH-to-OH distance for multiple structured data.

[0110] The MD computation time [ns] is calculated by performing the MD calculation four times, until it completes for 20 ns or collapses, and the cumulative time for each calculation is shown. Specifically, if any of the MD calculations collapse before completing 20 ns, the MD computation time [ns] will be less than 80 ns. A smaller MD computation time [ns] indicates that the MD calculation is more prone to collapse and that the accuracy of the model based on the prepared training data is low.

[0111] As shown in Table 800, the proposed method can select structural data with smaller OH-to-OH distances compared to the first, second, and third conventional methods. Therefore, the proposed method can improve the accuracy of the model for target molecules with small OH-to-OH distances based on the selected structural data.

[0112] Therefore, as shown in Table 800, unlike the first and third conventional methods, the proposed method can perform MD calculations without disrupting the model by improving the accuracy of the model for molecules with small OH-to-OH distances. Furthermore, as shown in Table 800, unlike the second conventional method, the proposed method can reduce the number of additional structural data to be selected, thereby reducing the time required to update the model.

[0113] (Overall processing procedure) Next, an example of the overall processing procedure executed by the information processing device 100 will be described using Figure 9. The overall processing is realized, for example, by the CPU 301 shown in Figure 3, storage areas such as memory 302 and recording medium 305, and network I / F 303.

[0114] Figure 9 is a flowchart showing an example of the overall processing procedure. In Figure 9, the information processing device 100 uses Volume Scan to generate multiple structural data by contracting the system (step S901).

[0115] Next, the information processing device 100 inputs each structural data into the neural network (NN) and extracts two first features related to each structural data from the NN (step S902). Then, for each structural data, the information processing device 100 calculates a second feature representing the minimum distance between atoms (step S903).

[0116] Next, the information processing device 100 calculates the distance between coordinate points based on the coordinate points of the structural data, which are represented by a combination of two first features and one second feature for each structural data in three-dimensional space (step S904). Then, based on the distance between coordinate points, the information processing device 100 selects a predetermined number of coordinates in order from those with relatively long distances to other coordinates (step S905).

[0117] Next, the information processing device 100 performs a DFT calculation on the structure data corresponding to each coordinate of a predetermined number of selected coordinates and assigns labels to them (step S906). Then, the information processing device 100 updates the NN based on the labeled structure data (step S907).

[0118] The information processing device 100 performs MD calculation using the updated NN (step S908). The information processing device 100 then completes the entire process. This allows the information processing device 100 to improve the accuracy of the NN. Here, the information processing device 100 may execute some of the steps in the flowchart of Figure 9 in a different order. For example, the order of steps S902 and S903 can be changed.

[0119] Furthermore, the information processing device 100 may omit processing in some steps of the flowchart in Figure 9. For example, the processing in step S908 can be omitted. Specifically, the information processing device 100 may further improve the accuracy of the neural network by performing the overall processing multiple times without performing the processing in step S908. Specifically, the information processing device 100 may perform the processing in step S908 after performing the overall processing multiple times.

[0120] (Examples of applications of the information processing device 100) Next, examples of applications of the information processing device 100 will be described. The information processing device 100 can be applied, for example, to the field of drug discovery or the field of materials development. Specifically, the information processing device 100 can be applied when it is desirable to perform MD calculations in order to investigate structural changes of proteins.

[0121] As explained above, the information processing device 100 can store a model that generates output data representing the energy of a target molecule in response to input data representing the structure of the target molecule. The information processing device 100 can obtain multiple first features related to the structure of the target molecule for each of the multiple structural data representing the structure of the target molecule, in accordance with the result of inputting the structural data into the model. The information processing device 100 can calculate a second feature related to the interatomic distance in the structure of the target molecule represented by the structural data for each of the multiple structural data. The information processing device 100 can select at least one of the multiple structural data to be used when updating the model, based on the distribution of coordinates in a multidimensional space that represents a combination of the multiple first features and the second features. As a result, the information processing device 100 can select appropriate structural data to be used when updating the model in order to improve the accuracy of the model.

[0122] According to the information processing device 100, multiple structural data representing the structure of the target molecule in the reduced system can be used. This allows the information processing device 100 to prepare corresponding structural data for cases where the distances between atoms are relatively small, thereby improving the accuracy of the model.

[0123] According to the information processing device 100, the model can be updated based on training data that associates any selected structural data with the energy for the target molecule obtained by performing a DFT calculation on that structural data. This allows the information processing device 100 to prepare appropriate training data that includes the energy accurately determined by the DFT calculation.

[0124] According to the information processing device 100, for each of the new structural data sets, multiple first features relating to the structure of the target molecule can be obtained according to the result of inputting the structural data into the updated model. According to the information processing device 100, for each of the new structural data sets, a second feature relating to the interatomic distances in the structure of the target molecule represented by the structural data can be calculated. According to the information processing device 100, based on the distribution of coordinates in a multidimensional space that represent the combination of the multiple first features and the second features, at least one of the new structural data sets can be selected to be used when updating the model. As a result, the information processing device 100 can further update the model to improve its accuracy.

[0125] According to the information processing device 100, based on the distribution of coordinates, a predetermined number of structural data can be selected from among multiple structural data, in order from those with relatively large Euclidean distances to other structural data. This allows the information processing device 100 to uniformly select a predetermined number of structural data so as to select structural data corresponding to cases where the interatomic distances are relatively close, which could be a cause of MD calculation failure, from the perspective of the minimum interatomic distance.

[0126] According to the information processing device 100, multiple structural data can be generated by performing MD calculations using the model. This allows the information processing device 100 to prepare multiple structural data within its own system.

[0127] According to the information processing device 100, a neural network (NN) can be used in the model. According to the information processing device 100, for each of the multiple structural data, multiple primary features related to the structure of the target molecule can be obtained by performing principal component analysis on the output of the intermediate layer of the model when the structural data is input to the model. As a result, the information processing device 100 can efficiently obtain multiple primary features related to the structure of the target molecule, making it easier to uniformly select structural data from the viewpoint of the structure of the target molecule.

[0128] According to the information processing device 100, the model can employ a neural network (NN) generated based on multiple training data sets, each representing a sample of input data indicating the structure of the target molecule and the correct output data indicating the energy of the target molecule. This allows the information processing device 100 to utilize a model in a format that can accurately represent the potential.

[0129] According to the information processing device 100, a second feature quantity can be calculated that indicates the minimum interatomic distance in the structure of the target molecule. This makes it easier for the information processing device 100 to select structural data corresponding to cases where the interatomic distances are relatively close, which can be a cause of MD calculation failure, from the perspective of the minimum interatomic distance.

[0130] The information processing method described in this embodiment can be implemented by executing a pre-prepared program on a computer such as a PC or workstation. The information processing program described in this embodiment is recorded on a computer-readable recording medium and executed by being read from the recording medium by the computer. The recording medium can be a hard disk, flexible disk, CD (Compact Disc)-ROM, MO (Magneto Optical Disc), DVD (Digital Versatile Disc), etc. Furthermore, the information processing program described in this embodiment may be distributed via a network such as the Internet.

[0131] With regard to the embodiments described above, the following additional information is disclosed.

[0132] (Note 1) For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule according to the input data representing the structure of the target molecule acquires multiple first features related to the structure of the target molecule according to the result of inputting the structural data. For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. Referencing coordinates in a multidimensional space that represent combinations of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, and based on the distribution of the coordinates, select at least one of the multiple structural data to be used when updating the model. An information processing program characterized by having a computer perform the processing.

[0133] (Note 2) The information processing program according to Note 1, characterized in that the plurality of structural data each represent the structure of the target molecule in the reduced system.

[0134] (Note 3) The model is updated based on training data that associates any of the selected structural data with the energy for the target molecule obtained by performing density functional theory calculations on any of the selected structural data. The information processing program described in Appendix 2, characterized in that it causes the computer to perform the processing.

[0135] (Note 4) For each of the new structural data representing the structure of the target molecule, a plurality of first features relating to the structure of the target molecule are obtained in the updated model according to the result of inputting the structural data. For each of the aforementioned new sets of structural data, a second feature quantity relating to the interatomic distances in the structure of the target molecule represented by the structural data is calculated. By referring to coordinates in a multidimensional space that represent the combination of the acquired multiple first features and the calculated second features corresponding to each of the multiple new structural data, at least one of the multiple new structural data to be used when updating the model is selected based on the distribution of the coordinates. The information processing program described in Appendix 3, characterized in that it causes the computer to perform the processing.

[0136] (Note 5) The process to be selected above is The information processing program according to Appendix 1, characterized in that, based on the distribution of the coordinates, a predetermined number of structural data are selected from the plurality of structural data in order from the structural data with a relatively large Euclidean distance to other structural data.

[0137] (Note 6) By performing a dynamic simulation using the above model, the above-mentioned multiple structural data are generated. An information processing program according to any one of the appendices 1 to 5, characterized in that it causes the computer to perform the processing.

[0138] (Note 7) The above model is a neural network, The aforementioned acquisition process is, An information processing program according to any one of the appendices 1 to 5, characterized in that, for each of the multiple structural data, a plurality of first features relating to the structure of the target molecule are obtained by performing principal component analysis on the output of the intermediate layer of the model when the structural data is input to the model.

[0139] (Note 8) The information processing program according to any one of Notes 1 to 5, characterized in that the model is a neural network generated based on a plurality of training data sets that associate samples of input data representing the structure of the target molecule with the correct output data representing the energy of the target molecule.

[0140] (Note 9) The information processing program according to any one of Notes 1 to 8, characterized in that the second feature quantity indicates the minimum value of the interatomic distance in the structure of the target molecule.

[0141] (Note 10) For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule according to the input data representing the structure of the target molecule acquires multiple first features relating to the structure of the target molecule according to the result of inputting the structural data, For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. Referencing coordinates in a multidimensional space that represent combinations of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, and based on the distribution of the coordinates, select at least one of the multiple structural data to be used when updating the model. An information processing method characterized in that the processing is performed by a computer.

[0142] (Note 11) For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule according to the input data representing the structure of the target molecule acquires multiple first features relating to the structure of the target molecule according to the result of inputting the structural data, For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. Referencing coordinates in a multidimensional space that represent combinations of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, and based on the distribution of the coordinates, select at least one of the multiple structural data to be used when updating the model. An information processing device characterized by having a control unit. [Explanation of symbols]

[0143] 100 Information Processing Devices 101 Structural Data 110 Model 120 Multidimensional space 121 areas 200 Information Processing Systems 201 Training Processing Device 202 Client Devices 210 Network 300 bus 301 CPU 302 memory 303 Network I / F 304 Recording medium interface 305 Recording media 306 displays 307 Input device 400 Storage section 401 Acquisition Department 402 Generator 403 First Calculation Unit 404 Second Calculation Unit 405 Selection Section 406 Training Department 407 Output section 500 Machine Learning Potential 501, 502 Structural Data Sets 503 Training Datasets 510 datasets 700,710 graph 800 tables

Claims

1. For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule in accordance with the input data representing the structure of the target molecule acquires multiple first features related to the structure of the target molecule in accordance with the result of inputting the structural data. For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. By referring to coordinates in a multidimensional space that represent the combination of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, at least one of the multiple structural data to be used when updating the model is selected based on the distribution of the coordinates. An information processing program characterized by having a computer perform the processing.

2. The information processing program according to claim 1, characterized in that the plurality of structural data each represent the structure of the target molecule in the reduced system.

3. The model is updated based on training data that associates any of the selected structural data with the energy for the target molecule obtained by performing density functional theory calculations on any of the structural data. The information processing program according to claim 2, characterized in that it causes the computer to perform the processing.

4. For each of the new structural data representing the structure of the target molecule, a plurality of first features relating to the structure of the target molecule are obtained in the updated model according to the result of inputting the structural data. For each of the aforementioned new sets of structural data, a second feature quantity relating to the interatomic distances in the structure of the target molecule represented by the structural data is calculated. By referring to coordinates in a multidimensional space that represent the combination of the acquired multiple first features and the calculated second features corresponding to each of the multiple new structural data, at least one of the multiple new structural data to be used when updating the model is selected based on the distribution of the coordinates. The information processing program according to claim 3, characterized in that it causes the computer to perform the processing.

5. The process to be selected is, The information processing program according to claim 1, characterized in that, based on the distribution of the coordinates, a predetermined number of structural data are selected from the plurality of structural data in order from the structural data with a relatively large Euclidean distance from other structural data.

6. By performing a dynamics simulation using the aforementioned model, the multiple structural data are generated. An information processing program according to any one of claims 1 to 5, characterized in that it causes the computer to perform the processing.

7. The aforementioned model is a neural network, The aforementioned acquisition process is, An information processing program according to any one of claims 1 to 5, characterized in that, for each of the plurality of structural data, a plurality of first features relating to the structure of the target molecule are obtained by performing principal component analysis on the output of the intermediate layer of the model when the structural data is input to the model.

8. The information processing program according to any one of claims 1 to 5, characterized in that the model is a neural network generated based on a plurality of training data sets that associate samples of input data representing the structure of the target molecule with correct output data representing the energy of the target molecule.

9. The information processing program according to any one of the appendices 1 to 5, characterized in that the second feature quantity indicates the minimum value of the interatomic distance in the structure of the target molecule.

10. For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule in accordance with the input data representing the structure of the target molecule acquires multiple first features related to the structure of the target molecule in accordance with the result of inputting the structural data. For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. By referring to coordinates in a multidimensional space that represent the combination of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, at least one of the multiple structural data to be used when updating the model is selected based on the distribution of the coordinates. An information processing method characterized in that the processing is performed by a computer.

11. For each of the multiple structural data representing the structure of the target molecule, a model that generates output data representing the energy of the target molecule in accordance with the input data representing the structure of the target molecule acquires multiple first features related to the structure of the target molecule in accordance with the result of inputting the structural data. For each of the aforementioned plurality of structural data, a second feature quantity relating to the interatomic distance in the structure of the target molecule represented by the structural data is calculated. By referring to coordinates in a multidimensional space that represent the combination of the acquired multiple first features and the calculated second features corresponding to each of the multiple structural data, at least one of the multiple structural data to be used when updating the model is selected based on the distribution of the coordinates. An information processing device characterized by having a control unit.

Images