Training data generation program, training data generation method, and information processing device

Abstract

An information processing device (100) generates a plurality of pieces of unknown structure data by executing a molecular dynamics simulation using a result obtained by inputting, to a machine learning potential, structure data in which the positions of a plurality of atoms are set. A computer executes screening on the plurality of pieces of unknown structure data using uncertainty of predicted values of forces acting on atoms estimated from the unknown structure data and structural feature quantities of the unknown structure data. The computer trains the machine learning potential using training data based on the structure data obtained by the screening.

Description

Training data generation program, training data generation method, and information processing device.

[0001] This invention relates to a training data generation program, etc.

[0002] There is a molecular dynamics (MD) simulation using machine learning interatomic potentials. MD simulation is a method that simulates the time evolution of a system consisting of N atoms by solving the equations of motion. N is a natural number greater than or equal to 1.

[0003] The equations of motion used in MD simulations are given by equation (1). In equation (1), r ｉ U(r) indicates the position (coordinate) of the i-th atom out of N atoms. １ ,. . . r Ｎ This represents an empirical potential whose functional form is defined in advance.

[0004]

[0005] Figure 9 is a diagram illustrating the procedure for MD simulation. In the explanation of Figure 9, the device that performs the MD simulation is referred to as computer 10. Computer 10 receives input of the initial structure 5. The initial structure 5 is the initial atomic arrangement and includes the positions (coordinates) of multiple atoms. Based on the positions r of the atoms, computer 10 calculates the force f acting on the atoms.

[0006] Computer 10 calculates the atomic velocity v and position r by integrating the equation of motion shown in equation (1) over time, and updates the atomic velocity v and position r based on the calculation results.

[0007] Computer 10 calculates the time evolution of N atoms by repeatedly performing the processes of calculating the force f and updating the atomic velocity v and position r.

[0008] Furthermore, the above MD simulations have limitations in accuracy because they use empirical potentials with predefined functional forms. To improve the accuracy of MD simulations, machine learning potentials are being used to predict energy and force from atomic arrangements.

[0009] Figure 10 is a diagram illustrating MD simulation using a machine learning potential. As shown in Figure 10, by inputting the atomic arrangement 15 into the machine learning potential 20, energy and force are output from the machine learning potential 20. In the MD simulation, material properties can be calculated using the energy and force output from the machine learning potential 20. These material properties include atomic density, diffusion coefficient, and elastic modulus.

[0010] To reliably perform MD simulations, it is necessary to use a highly accurate machine learning potential. To construct a highly accurate machine learning potential, it is necessary to create a dataset that corresponds to a variety of atomic arrangements and train the machine learning potential. For example, the dataset may contain multiple training data, where the input to the training data is the atomic arrangement and the labels are the energy and force corresponding to the atomic arrangement.

[0011] For example, active learning is a conventional method for creating datasets that correspond to diverse atomic arrangements. Figure 11 is a diagram illustrating conventional active learning. The conventional device that performs the process in Figure 11 is simply referred to as the "conventional device". The conventional device prepares an initial dataset 30 (step S10). The conventional device trains a machine learning potential 20 using multiple training data stored in the dataset 30 (step S11).

[0012] The conventional device performs MD using a machine learning potential 20 to generate structural data 25 in which the energy and force are unknown (step S12). For example, in the structural data 25, one white circle represents one structural data, and one structural data represents the structural data of a system consisting of multiple atoms.

[0013] Conventional devices screen structural data 25 using information such as force uncertainty (step S13). Through this screening, several structural data are selected from the structural data 25. For example, the black circles represent the selected structural data 26.

[0014] The conventional device performs electronic state calculations on the selected structural data 26 to obtain energy and force (step S14). The conventional device generates structural data 27 by assigning labels (energy and force) to the selected structural data 26.

[0015] Conventional devices update the dataset 30 by adding labeled structural data 27 as training data (step S15).

[0016] The conventional device can generate training data with unknown structures that are not present in the initial dataset 30 by repeatedly executing a cycle consisting of steps S10 to S15 shown in Figure 11.

[0017] Here, a crucial point that determines the success or failure of active learning is how to screen for unknown structural data that is not present in the dataset 30 from the structural data 25 generated in step S12 of Figure 11.

[0018] In the example described in step S13 of Figure 11, we explained the case where screening is performed based on force uncertainty, but there are also conventional techniques that perform screening using structural features of structural data 25.

[0019] Japanese Patent Publication No. 2024-075646, Japanese Patent Publication No. 2023-059702, U.S. Patent Application Publication No. 2023 / 0359929

[0020] However, conventional techniques have the problem of not being able to train machine learning potential with high accuracy.

[0021] For example, when screening is performed based on force uncertainty, the diversity of the structural data is lost. Using machine learning potentials trained on structural data with lost diversity reduces the stability of MD simulations.

[0022] On the other hand, when screening is performed based on structural features, the resulting structural data may include non-physical structural data. If non-physical structural data is included, the electronic state calculations for such structural data will not converge, and it cannot be added to the dataset.

[0023] In one aspect, the present invention aims to provide a training data generation program, a training data generation method, and an information processing device that can accurately train machine learning potential.

[0024] In the first approach, the computer performs the following processes: The computer generates multiple unknown structural data by inputting structural data with the positions of multiple atoms into a machine learning potential and using the results to perform molecular dynamics simulations. The computer then screens the multiple unknown structural data using the uncertainty of the predicted forces acting on the atoms estimated from the unknown structural data and the structural features of the unknown structural data. The computer then trains the machine learning potential using training data based on the structural data obtained from the screening.

[0025] It allows for highly accurate training of machine learning potential.

[0026] Figure 1 is a diagram illustrating the first method. Figure 2 is a diagram illustrating the second method. Figure 3 is a functional block diagram showing the configuration of the information processing device according to this embodiment. Figure 4 is a diagram showing an example of the data structure of machine learning potential data. Figure 5 is a flowchart showing the processing procedure of the information processing device according to this embodiment. Figure 6 is a diagram (1) showing the results of effect verification. Figure 7 is a diagram (2) showing the results of effect verification. Figure 8 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of the embodiment. Figure 9 is a diagram illustrating the procedure of MD simulation. Figure 10 is a diagram illustrating MD simulation using machine learning potential. Figure 11 is a diagram illustrating conventional active learning.

[0027] Hereinafter, embodiments of the training data generation program, training data generation method, and information processing apparatus disclosed in the present application will be described in detail based on the drawings. Note that the present invention is not limited by these embodiments.

[0028] Before describing this embodiment, supplementary explanations will be given regarding the screening using the uncertainty of force and the screening using structural feature quantities described in the background art. In the following description, the screening using the uncertainty of force is referred to as the "first method". The screening using structural feature quantities is referred to as the "second method".

[0029] First, the first method will be described. The apparatus that executes the first method is referred to as the "first apparatus". The first apparatus uses a plurality of different machine learning potentials. The plurality of machine learning potentials are machine learning potentials with different initial values of parameters and are trained using the same dataset.

[0030] The first apparatus obtains predicted values of the forces acting on each atom by inputting the same structural data into the plurality of machine learning potentials. The predicted values of the forces are output as many as the number of machine learning potentials.

[0031] FIG. 1 is a diagram for explaining the first method. In FIG. 1, as an example, machine learning potentials 20a, 20b, and 20c are used for explanation, but it is not limited thereto, and machine learning potentials other than 20a to 20c may be further used.

[0032] Let the machine learning potential 20a be a machine learning potential trained by the dataset 30 with the initial value of the parameter being the initial value 20a-1. Let the machine learning potential 20b be a machine learning potential trained by the dataset 30 with the initial value of the parameter being the initial value 20b-1. Let the machine learning potential 20c be a machine learning potential trained by the dataset 30 with the initial value of the parameter being the initial value 20c-1. The initial values 20a-1, 20b-1, and 20c-1 are different values respectively.

[0033] When the machine learning potentials 20a to 20c receive an input of structural data (atomic arrangement), they output predicted values of energy and force. In the subsequent explanations regarding FIG. 1, the explanation regarding the energy output from the machine learning potentials 20a, 20b, and 20c is omitted.

[0034] The first device obtains predicted values of force 21a, 21b, and 21c by inputting the structural data 16 into the machine learning potentials 20a, 20b, and 20c. The first device calculates the standard deviation of the predicted values 21a to 21c obtained from the machine learning potentials 20a to 20c as the "uncertainty of force ε" (hereinafter, sometimes referred to as "uncertainty ε"). The first device calculates the uncertainty ε of each structural data by executing the above process for each structural data to be screened.

[0035] The first device is the lower limit value σ of the uncertainty of force ｌｏ and the upper limit value σ ｈｉ and uses σ ｌｏ ≦ε<σ ｈｉ to screen the structural data that satisfies the condition. The lower limit value σ ｌｏ and the upper limit value σ ｈｉ are set in advance.

[0036] For example, the structural data for which ε<σ ｌｏ can be said to be predicted with sufficient accuracy by the current machine learning potential, so there is no need to add it to the dataset.

[0037] For the structural data for which σ ｈｉ <ε, the prediction accuracy by the current machine learning potential is very low, and it is likely to be non-physical structural data, so it is not included in the dataset.

[0038] For the structural data for which σ ｌｏ ≦ε<σ ｈｉ is a physically reasonable structure and the prediction accuracy by the current machine learning potential is low, so it is included in the dataset.

[0039] The first device randomly selects a predetermined number of structural data from the screened structural data. As explained in step S14 of Figure 11, an electronic state calculation is performed on the selected structural data. Since the electronic state calculation is computationally expensive, it is not practical to perform the calculation on all the screened structural data.

[0040] In the first method described above, the diversity of the structural data is lost during the process of randomly selecting a predetermined number of elements after force uncertainty-based screening. This reduces the stability of MD simulations using machine learning potentials.

[0041] Next, the second method will be explained. The device that performs the second method will be referred to as the "second device." The second method calculates the structural features of the unknown structural data generated in step S12 of Figure 11. The unknown structural data is structural data for which the force and energy applied to the structural data are unknown. The second device projects the unknown structural data onto a two-dimensional feature map using a dimensionality reduction method.

[0042] Figure 2 is a diagram illustrating the second method. In Figure 2, the horizontal axis of graphs G1 and G2 corresponds to the first principal component. The vertical axis of graphs G1 and G2 corresponds to the second principal component. For example, graph G1 is obtained by using a dimensionality reduction method. ３ The results of projecting each structural data of Li onto a two-dimensional feature map are shown. One point on graph G1 is Au ３ This corresponds to the projection result of one structure data of Li.

[0043] The second device obtains the result of graph G2 by broadly sampling a predetermined number of points on graph G1, while removing similar structures based on the distance between two points in the result projected onto the two-dimensional feature map (graph G1).

[0044] In the second method described above, sampling is performed broadly, including sparse regions on the feature map, which means that the resulting structural data may include non-physical structures. If non-physical structural data is included, the electronic state calculations for such structural data will not converge, and it cannot be added to the dataset.

[0045] The above provides supplementary explanations regarding the first and second methods.

[0046] Next, the information processing device according to this embodiment will be described. The information processing device according to this embodiment will be referred to as "information processing device 100". The information processing device 100 uses active learning to perform screening in two stages during the process of generating a dataset. For example, the information processing device 100 performs a first-stage screening on unknown structural data based on the uncertainty of the predicted force value. The information processing device 100 then performs a second-stage screening on the structural data screened in the first stage, based on structural features.

[0047] The information processing device 100 samples physically valid structural data in the first screening stage and samples diverse structures with different structural features in the second screening stage. This ensures diversity in the dataset and improves the stability of MD simulations using machine learning potentials. In other words, the information processing device 100 allows for highly accurate training of machine learning potentials.

[0048] An example of the configuration of the information processing device 100 according to this embodiment will be described. Figure 3 is a functional block diagram showing the configuration of the information processing device according to this embodiment. As shown in Figure 3, the information processing device 100 has a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0049] The communication unit 110 performs data communication with an external device via a network. The communication unit 110 is implemented by a NIC (Network Interface Card) or the like. For example, the information processing device 100 may acquire a dataset 141 from an external device.

[0050] The input unit 120 is an input device that inputs various types of information into the information processing device 100. The input unit 120 is compatible with keyboards, mice, touch panels, etc. The user inputs various types of information by operating the input unit 120.

[0051] The display unit 130 is a display device that displays information output from the control unit 150. The display unit 130 is compatible with liquid crystal displays, organic EL (Electro Luminescence) displays, touch panels, etc.

[0052] The storage unit 140 includes a dataset 141, machine learning potential data 142, and a structured data table 143. The storage unit 140 is a memory or the like.

[0053] Dataset 141 is used when training the machine learning potential. Dataset 141 contains multiple training datasets, where the input (explanatory variables) of the training dataset are structural data (atomic arrangements), and the labels (dependent variables) are the energy and force corresponding to the structural data.

[0054] The machine learning potential data 142 contains data for multiple machine learning potentials. Figure 4 shows an example of the data structure of the machine learning potential data. In the example shown in Figure 4, the machine learning potential data 142 has machine learning potentials 50a, 50b, and 50c. In Figure 4, machine learning potentials 50a to 50c are used as an example for explanation, but the explanation is not limited to these, and other machine learning potentials besides machine learning potentials 50a to 50c may also be used.

[0055] Machine learning potential 50a is defined as the machine learning potential trained using dataset 141, with the initial parameter value set to initial value 50a-1. Machine learning potential 50b is defined as the machine learning potential trained using dataset 141, with the initial parameter value set to initial value 50b-1. Machine learning potential 50c is defined as the machine learning potential trained using dataset 141, with the initial parameter value set to initial value 50c-1. Initial values ​​50a-1, 50b-1, and 50c-1 are all different values. Machine learning potentials 50a to 50c are Neural Networks (NN), etc.

[0056] In the following explanation, machine learning potentials 50a to 50c will be collectively referred to as machine learning potential 50. When machine learning potential 50 receives structural data (atomic arrangement) as input, it outputs predicted values ​​for energy and force.

[0057] The structure data table 143 is a table that holds unknown structure data. As will be described later, the unknown structure data in the structure data table 143 is subject to screening.

[0058] The control unit 150 includes a training unit 151, a generation unit 152, a first screening unit 153, a second screening unit 154, an electronic state calculation unit 155, and a physical property calculation unit 156. For example, the control unit 150 is a CPU (Central Processing Unit), etc.

[0059] The training unit 151 trains the machine learning potential 50 based on the dataset 141. For example, the training unit 151 updates the parameters of the machine learning potential 50 so that the output when the structural data of the training data from the dataset 141 is input to the machine learning potential 50 approaches the label. The training unit 151 trains the machine learning potential 50 using backpropagation.

[0060] The training unit 151 repeatedly executes the above process each time the dataset 141 is updated by the process described later.

[0061] The generation unit 152 generates multiple unknown structural data by performing an MD simulation using the machine learning potential 50 (any one of the machine learning potentials 50a, 50b, or 50c), and registers the generated multiple unknown structural data in the structural data table 143.

[0062] For example, the generation unit 152 estimates the energy and force of the structural data by inputting the pre-specified structural data into the machine learning potential 50a. Using the estimated force, the generation unit 152 integrates the equation of motion described in equation (1) over time to calculate the atomic velocity v and position r, and then calculates the time evolution of the specified structural data by updating the atomic velocity v and position r based on the calculation results. This generates multiple unknown structural data.

[0063] The first screening unit 153 performs a screening on unknown structural data included in the structural data table 143 based on the uncertainty of the predicted values ​​of the forces acting on atoms estimated from the unknown structural data. The first screening unit 153 keeps the screened unknown structural data and deletes the other structural data from the structural data table 143.

[0064] The processing of the first screening unit 153 will be explained in more detail. For example, the first screening unit 153 obtains unknown structural data from the structural data table 143 and inputs the obtained unknown structural data into the machine learning potentials 50a to 50c, respectively, to obtain multiple force prediction values. The first screening unit 153 calculates the standard deviation of the multiple force prediction values ​​as the "force uncertainty ε".

[0065] The first screening unit 153 determines that the calculated force uncertainty ε is σ ｌｏ ≦ε<σ ｈｉ If the conditions are met, the corresponding unknown structural data is left in the structural data table 143. On the other hand, the first screening unit 153 determines that the calculated force uncertainty ε is σ ｌｏ ≦ε<σ ｈｉ If the conditions are not met, the corresponding unknown structural data is deleted from the structural data table 143. Note that the lower limit of force uncertainty σｌｏ and upper limit σ ｈｉ These are set in advance. For example, the lower limit σ ｌｏ It is "0.05 eV / Å", and the upper limit σ ｈｉ This is "0.2 eV / Å".

[0066] The first screening unit 153 performs screening by repeatedly executing the above process on each unknown structural data in the structural data table 143.

[0067] The second screening unit 154 performs a screening on the unknown structural data contained in the structural data table 143 based on the structural features of the unknown structural data. The second screening unit 154 keeps the screened unknown structural data and deletes the other structural data from the structural data table 143.

[0068] The processing of the second screening unit 154 will be explained in more detail. For example, the second screening unit 154 calculates structural features for each unknown structural data contained in the structural data table 143. The second screening unit 154 can calculate structural features using a symmetric function.

[0069] The second screening unit 154 projects the structural features of the unknown structural data onto a feature map based on a dimensionality reduction method. The second screening unit 154 removes similar points based on the distance between any two points (points corresponding to the unknown structural data) projected onto the feature map. The second screening unit 154 removes similar points on the feature map until the number of points remains at a predetermined number. The second screening unit 154 leaves the unknown structural data corresponding to the points that finally remain on the feature map in the structural data table 143, and removes the other structural data from the structural data table 143.

[0070] Unknown structural data screened by the first screening unit 153 and the second screening unit 154 remains in the structural data table 143.

[0071] The electronic state calculation unit 155 calculates the energy and force for each unknown structural data by performing electronic state calculations on the unknown structural data in the structural data table 143. The electronic state calculation unit 155 adds information (training data) that associates "structural data" with "energy and force" to the dataset 141. The electronic state calculation unit 155 performs electronic state calculations based on Density Functional Theory (DFT).

[0072] The training unit 151, generation unit 152, first screening unit 153, second screening unit 154, and electronic state calculation unit 155 described above repeatedly execute the above process until the accuracy of the machine learning potential 50 is sufficient. The information processing device 100 may calculate a score for the machine learning potential 50 using the test data, and if the score is above a threshold, it may determine that the accuracy of the machine learning potential 50 is sufficient. Alternatively, the information processing device 100 may determine that the accuracy of the machine learning potential 50 is sufficient when the number of times the above repeated process has exceeded a predetermined number of times.

[0073] The physical property calculation unit 156 utilizes the machine learning potential 50, which has been determined to have sufficient accuracy. The physical property calculation unit 156 performs an MD simulation using the machine learning potential 50 and calculates the physical properties. The physical property calculation unit 156 may output the calculation results of the physical properties to the display unit 130 for display.

[0074] Next, the processing procedure of the information processing device 100 according to this embodiment will be described. Figure 5 is a flowchart showing the processing procedure of the information processing device according to this embodiment. As shown in Figure 5, the information processing device 100 prepares the data set 141 (step S101).

[0075] The training unit 151 of the information processing device 100 trains the machine learning potential 50 based on the dataset 141 (step S102). If the accuracy of the machine learning potential 50 is sufficient (step S103, Yes), the information processing device 100 proceeds to step S104. The physical property calculation unit 156 of the information processing device 100 calculates the physical properties by performing an MD simulation using the machine learning potential 50 (step S104).

[0076] On the other hand, if the accuracy of the machine learning potential 50 is insufficient (step S103, No), the information processing device 100 proceeds to step S105. The generation unit 152 of the information processing device 100 performs an MD simulation using the machine learning potential 50 and generates multiple unknown structural data (step S105).

[0077] The first screening unit 153 of the information processing device 100 performs screening on unknown structural data based on the uncertainty of the force (step S106). The second screening unit 154 of the information processing device 100 performs screening on unknown structural data based on structural features (step S107).

[0078] The electronic state calculation unit 155 of the information processing device 100 calculates energy and force for the unknown structural data which is the result of a two-stage screening (step S108). The electronic state calculation unit 155 adds training data to the dataset 141 (step S109) and proceeds to step S102.

[0079] Next, the effects of the information processing device 100 according to this embodiment will be described. The information processing device 100 generates multiple unknown structural data by inputting structural data into a machine learning potential and performing an MD simulation using the results obtained. The information processing device 100 performs a screening of the multiple unknown structural data using the uncertainty of the predicted force acting on the atoms estimated from the unknown structural data and the structural features of the structural data, and trains the machine learning potential 50 using training data based on the structural data obtained from the screening. This makes it possible to train the machine learning potential with high accuracy.

[0080] For example, the information processing device 100 performs a first-stage screening on unknown structural data based on the uncertainty of the force prediction value, and then performs a second-stage screening on the structural data screened in the first stage based on structural features. The computational cost of the screening based on structural features is greater than the computational cost of the screening based on the uncertainty of the force prediction value. Therefore, by performing the screening in the order described above, the target of the screening based on structural features can be reduced, thereby reducing the computational cost.

[0081] The information processing device 100 may perform a first-stage screening based on structural features for unknown structural data, and then perform a second-stage screening based on the uncertainty of the force prediction value for the structural data screened in the first stage.

[0082] Here, we will show the results of the effectiveness verification using Figures 6 and 7. Figure 6 is Figure (1) showing the results of the effectiveness verification. Figure 6 shows the number of training data added to the dataset in each cycle of active learning. The maximum number of training data added in each cycle was set to "200".

[0083] In the first method, the number of training data added in cycle 1 is "200", the number of training data added in cycle 2 is "200", and the number of training data added in cycle 3 is "200". In the first method, the number of training data added in cycle 4 is "197", and the number of training data added in cycle 5 is "200".

[0084] In the second method, the number of training data added in cycle 1 is "3", the number of training data added in cycle 2 is "3", and the number of training data added in cycle 3 is "1". In the first method, the number of training data added in cycle 4 is "31", and the number of training data added in cycle 5 is "137".

[0085] In the present invention (information processing device 100), the number of training data added in cycle number 1 becomes "200", the number of training data added in cycle number 2 becomes "200", the number of training data added in cycle number 3 becomes "200", in the present invention, the number of training data added in cycle number 4 becomes "200", the number of training data added in cycle number 5 becomes "200".

[0086] Referring to Figure 6, even though the maximum number of training data points added in each cycle is set to 200, the second method adds very few training data points to the dataset in each cycle. This is likely due to non-physical structural data being sampled, causing the electronic state calculation to fail to converge.

[0087] On the other hand, referring to Figure 6, it can be seen that both the first method and the present invention are able to add training data to the dataset almost exactly as set.

[0088] Figure 7 is Figure (2) showing the results of the effect verification. Figure 7 shows the results of evaluating the stability and density of the MD simulation using the machine learning potential. For example, the density (g / cm³) at a temperature of 298.15 K. ３ A comparison was made between the two values. For example, the experimental value was "1.120".

[0089] In the first method, the MD simulation using the machine learning potential collapsed and became uncomputable at cycles 1, 3, and 4. In the second method, the MD simulation using the machine learning potential collapsed and became uncomputable at cycles 1, 2, and 3. In the present invention (information processing device 100), the MD simulation using the machine learning potential is executed stably at cycles 1 to 5, and a wider variety of structural data can be added to the training data.

[0090] The conditions for performing the verification described in Figures 6 and 7 were, for example, as follows: Polyethylene glycol was used as the target, and the machine learning potential was set to NNP. A Deep Potential (DP) model was created using the following method: Number of atoms during training = 155, number of training data in the initial dataset = 2000, number of active learning cycles = 5, maximum number of structures added in each cycle = 200, labeling was performed using DFT calculation, and the uncertainty of the force prediction was evaluated using four DP models. An NNP-MD simulation using the DP model was performed using the following method: Number of atoms = 3100, calculation time = 11 ns, temperature = 298.15 K, pressure = 1 bar.

[0091] Next, an example of a computer hardware configuration that realizes the same functions as the information processing device 100 described above will be explained. Figure 8 is a diagram showing an example of a computer hardware configuration that realizes the same functions as the information processing device of the embodiment. As shown in Figure 8, the computer 200 has a CPU 201 that performs various calculations, an input device 202 that receives data input from the user, and a display 203. The computer 200 also has a communication device 204 that exchanges data with external devices via a wired or wireless network, and an interface device 205. The interface device 205 may have a microphone, speaker, etc., connected to it. The computer 200 also has a RAM 206 that temporarily stores various information and a hard disk drive 207. Each of the devices 201 to 207 is connected to a bus 208.

[0092] The hard disk drive 207 includes a training program 207a, a generation program 207b, a first screening program 207c, a second screening program 207d, an electronic state calculation program 207e, and a physical property calculation program 207f. The CPU 201 reads each of the programs 207a to 207f and loads them into the RAM 206.

[0093] The training program 207a functions as the training process 206a. The generation program 207b functions as the generation process 206b. The first screening program 207c functions as the first screening process 206c. The second screening program 207d functions as the second screening process 206d. The electronic state calculation program 207e functions as the electronic state calculation process 206e. The physical property calculation program 207f functions as the physical property calculation process 206f.

[0094] The processing of training process 206a corresponds to the processing of training unit 151. The processing of generation process 206b corresponds to the processing of generation unit 152. The processing of the first screening process 206c corresponds to the processing of first screening unit 153. The processing of second screening process 206d corresponds to the processing of second screening unit 154. The processing of electronic state calculation process 206e corresponds to the processing of electronic state calculation unit 155. The processing of physical property calculation process 206f corresponds to the processing of physical property calculation unit 156.

[0095] Furthermore, it is not necessary to store each program 207a to 207f on the hard disk drive 207 from the beginning. For example, each program can be stored on a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card inserted into the computer 200. Then, the computer 200 can read and execute each program 207a to 207f.

[0096] 100 Information Processing Unit 110 Communication Unit 120 Input Unit 130 Display Unit 140 Storage Unit 141 Dataset 142 Machine Learning Potential Data 143 Structural Data Table 150 Control Unit 151 Training Unit 152 Generation Unit 153 First Screening Unit 154 Second Screening Unit 155 Electronic State Calculation Unit 156 Physical Property Calculation Unit

Claims

1. A training data generation program characterized by generating multiple unknown structural data by inputting structural data with the positions of multiple atoms into a machine learning potential and performing molecular dynamics simulations using the results obtained; performing screening on the multiple unknown structural data using the uncertainty of the predicted force acting on the atoms estimated from the unknown structural data and the structural features of the unknown structural data; and having a computer perform the process of training the machine learning potential using training data based on the structural data obtained from the screening.

2. The training data generation program according to claim 1, characterized in that the process of performing the screening involves performing a first screening on the plurality of unknown structural data using the uncertainty of the predicted force values, and then performing a second screening on the results of the first screening using the structural features.

3. The training data generation program according to claim 1 or 2, characterized in that the process of inputting the unknown structural data into multiple machine learning potentials, causing the computer to further perform a process of calculating multiple force prediction values, and performing the screening, utilizes the standard deviation of the multiple force prediction values ​​as the uncertainty of the force prediction values.

4. The training data generation program according to claim 1 or 2, characterized in that the process of performing the screening involves projecting the structural features of the plurality of unknown structural data onto a feature map based on a dimensionality reduction method, and performing screening on the plurality of unknown structural data based on the distance of the plurality of unknown structural data projected onto the feature map.

5. A method for generating training data, characterized in that a computer performs the following processes: inputting structural data in which the positions of multiple atoms are set into a machine learning potential and performing molecular dynamics simulations using the results obtained; generating multiple unknown structural data; performing screening on the multiple unknown structural data using the uncertainty of the predicted force acting on the atoms estimated from the unknown structural data and the structural features of the unknown structural data; and training the machine learning potential using training data based on the structural data obtained from the screening.

6. The training data generation method according to claim 5, characterized in that the process of performing the screening involves performing a first screening on the plurality of unknown structural data using the uncertainty of the predicted force values, and then performing a second screening on the results of the first screening using the structural features.

7. The training data generation method according to claim 5 or 6, characterized in that the computer further performs a process of calculating multiple force prediction values ​​by inputting the unknown structural data into multiple machine learning potentials, and the screening process uses the standard deviation of the multiple force prediction values ​​as the uncertainty of the force prediction values.

8. The training data generation method according to claim 5 or 6, characterized in that the process of performing the screening involves projecting the structural features of the plurality of unknown structural data onto a feature map based on a dimensionality reduction method, and performing screening on the plurality of unknown structural data based on the distance of the plurality of unknown structural data projected onto the feature map.

9. An information processing device having a control unit that generates multiple unknown structural data by inputting structural data in which the positions of multiple atoms are set into a machine learning potential and performing a molecular dynamics simulation using the results obtained; performs a screening of the multiple unknown structural data using the uncertainty of the predicted force acting on the atoms estimated from the unknown structural data and the structural features of the unknown structural data; and trains the machine learning potential using training data based on the structural data obtained from the screening.

10. The information processing apparatus according to claim 9, characterized in that the process of performing the screening involves performing a first screening on the plurality of unknown structural data using the uncertainty of the predicted force values, and performing a second screening on the results of the first screening using the structural features.

11. The information processing apparatus according to claim 9 or 10, wherein the control unit further performs a process of calculating multiple force prediction values ​​by inputting the unknown structural data into multiple machine learning potentials, and the screening process uses the standard deviation of the multiple force prediction values ​​as the uncertainty of the force prediction values.

12. The information processing apparatus according to claim 9 or 10, wherein the process for performing the screening involves projecting the structural features of the plurality of unknown structural data onto a feature map based on a dimensionality reduction method, and performing a screening of the plurality of unknown structural data based on the distance of the plurality of unknown structural data projected onto the feature map.

Images