Self-supervised learning program, method, and device

Abstract

This self-supervised learning device: generates data indicating a second molecule obtained by substituting each of a plurality of atoms included in a first molecule with any of the atoms included in the first molecule; inputs the data indicating the second molecule to a machine learning model that performs prediction regarding a molecular structure, to thereby acquire a prediction result; and updates a parameter of the machine learning model on the basis of a comparison result between ground-truth data corresponding to the first molecule and the prediction result.

Description

Self-supervised learning programs, methods, and apparatus

[0001] The disclosed technologies relate to self-supervised learning programs, self-supervised learning methods, and self-supervised learning devices.

[0002] Materials discovery, which involves predicting the properties of material structures and discovering or optimizing new materials suitable for various applications, is conducted in a wide range of fields and industries and is essential for technological advancement. Specific examples of the objectives of materials discovery include the search for high-performance materials for aerospace vehicles, materials for herovskite solar cells, new catalysts that efficiently promote chemical reactions, materials for biodegradable plastics, and highly biocompatible medical alloys.

[0003] In materials discovery, the properties of material structures are predicted through experiments and simulations. One type of simulation used is molecular dynamics (MD), which uses Newton's equations of motion to determine the forces and energies of atoms and molecules. MD simulations have the advantage of being able to calculate the time evolution of the system and having relatively low computational costs. Density functional theory (DFT) calculations, which use first-principles calculations based on quantum mechanics to determine the forces and energies of atoms and molecules, are also used. DFT calculations generally calculate the ground state and have relatively high computational costs.

[0004] Also, recently, material discovery has been carried out by MI (Materials Informatics) using a large amount of data. MI is a field in the material science area that clarifies the properties of materials or discovers new materials by making full use of AI (Artificial Intelligence) and the like. MI analyzes a large amount of data and provides a method for predicting the characteristics and behaviors of materials. As a technology related to MI, for example, there is a method called high-throughput screening that quickly selects materials satisfying criteria for a certain problem setting. High-throughput screening is used to quickly evaluate a large number of material candidates using AI or the like and identify the optimal materials. Also, there is a method of directly predicting the results obtained by MD simulation or DFT calculation using an AI surrogate model. The method using an AI surrogate model is used to replace simulations with high computational costs and can significantly shorten the calculation time.

[0005] Thus, in recent years, techniques for quickly predicting material properties by using an AI model have been developed. Various models, such as SchNet, PaiNN, GemNet-OC, EquiformerV2, etc., have been proposed as AI models for predicting material properties.

[0006] In order to train an AI model, correct label data is required. For example, in machine learning in the material field, the input data is a graph of a chemical structure considering atomic species and the positions of each atom, and the labels are force, energy, etc. This label data regarding the material structure is obtained by experiments or simulations. However, experiments require advanced equipment and expertise and are time-consuming and costly. Also, simulations, especially DFT, etc. have high computational costs and require large-scale computational resources. That is, both experiments and simulations have high implementation costs, and it is difficult to collect a large amount of label data.

[0007] Therefore, there is a method called self-supervised learning, in which an AI model is pre-trained using a large amount of unlabeled data, and then trained using a small amount of labeled data. This method is also used in the field of materials discovery.

[0008] For example, an information processing system has been proposed that generates a trained model by performing machine learning based on sequence information representing sequences that include part or all of the sequences of multiple antigen-binding molecules or proteins. Based on the trained model, this system generates virtual sequence information in which at least one of the constituent units that make up the sequence represented by the sequence information input to the trained model has been mutated.

[0009] Furthermore, a method has been proposed to pre-train a graph neural network using, for example, an untagged compound dataset. In this method, one of the tagged compound datasets tags one of the compound's properties. This method further obtains a compound property prediction model by performing multi-task training on a graph neural network pre-trained using multiple tagged compound datasets. The compound property prediction model is used to predict various properties of a compound.

[0010] International Publication No. 2020 / 246617, Japanese Patent Publication No. 2022-163051

[0011] Traditional self-supervised learning methods for training machine learning models still have room for improvement in terms of accuracy.

[0012] One aspect of the disclosure technology is that it aims to generate highly accurate machine learning models through self-supervised learning.

[0013] In one embodiment, the disclosed technology generates data representing a second molecule in which each of a plurality of atoms contained in a first molecule is replaced with any of the atoms contained in the first molecule. The disclosed technology then inputs the data representing the second molecule into a machine learning model that makes predictions about molecular structure to obtain prediction results, and updates the parameters of the machine learning model based on the comparison result between the ground truth data corresponding to the first molecule and the prediction results.

[0014] One aspect of this approach is that it allows for the generation of highly accurate machine learning models through self-supervised learning.

[0015] This is a diagram illustrating the NN for predicting molecular energy. This is a diagram illustrating transfer learning. This is a diagram illustrating reference method 1. This is a diagram illustrating reference method 2. This is a diagram illustrating fine tuning. This is a diagram illustrating the problems of the reference method. This is a diagram illustrating the overview of this embodiment. This is a functional block diagram of the self-supervised learning device according to the first to fourth embodiments. This is a diagram illustrating the processing of the generation unit according to the first embodiment. This is a block diagram illustrating the schematic configuration of a computer that functions as a self-supervised learning device according to the first to fourth embodiments. This is a flowchart illustrating an example of the self-supervised learning process according to the first embodiment. This is a diagram illustrating other processing of the generation unit according to the first embodiment. This is a diagram illustrating the processing of the generation unit according to the second embodiment. This is a flowchart illustrating an example of the self-supervised learning process according to the second embodiment. This is a flowchart illustrating an example of the self-supervised learning process according to the third embodiment. This is a flowchart illustrating an example of the self-supervised learning process according to the fourth embodiment.

[0016] An example of an embodiment relating to the disclosed technology will be described below with reference to the drawings.

[0017] Before describing the details of the embodiment, we will explain self-supervised learning for machine learning models that predict material properties.

[0018] For example, as shown in Figure 1, a neural network (hereinafter also referred to as "NN") that predicts the energy of a molecule based on its structural data is trained by supervised learning using labeled data consisting of "structure" and "energy". That is, the molecular structural data is input to the NN, and the parameters of the NN are updated by backpropagating the difference between the predicted molecular energy (output from the NN) and the correct energy. This correct energy is calculated, for example, by simulation such as DFT. Since the computation time for molecular energy using DFT is very long, it is difficult to collect a large amount of labeled data for supervised learning.

[0019] Therefore, there is a method called self-supervised learning, which generates correct answers (labels) for a task from unlabeled data and performs supervised learning using the generated labels. Generally, self-supervised learning is applied to pre-training in transfer learning. As shown in Figure 2, transfer learning pre-trains the NN using self-supervised learning with unlabeled data. The NN includes a feature extraction unit that extracts features from the input data and an output unit that corresponds to the task. The output unit used during pre-training is a pre-training unit. In pre-training, machine learning is performed using a large amount of data similar to the data domain that is actually to be specialized. By using a large amount of data to train the NN on the essential features of the input molecular data, the initial state of the NN can be improved.

[0020] Furthermore, if the tasks for pre-training and fine-tuning (FT) are different, the output section is modified for FT. For example, if the pre-training task is 3-class classification but the fine-tuning task is 4-class classification, the output section is changed to a fully connected layer for 4-class classification. Subsequently, the NN is trained using supervised learning with labeled data for fine-tuning. In this case, the feature extraction section is the same as the pre-trained one. That is, the weights of the pre-trained NN are used as initial values, and machine learning is performed with actual data. By applying self-supervised learning as pre-training, the knowledge gained from pre-training can be utilized, and even when there is little labeled data needed for fine-tuning (supervised learning) after pre-training, the accuracy of the NN can be improved.

[0021] Two example methods are presented as reference techniques for self-supervised learning of neural networks (NNs) used to predict material properties. Reference technique 1 involves preparing molecular data, as shown in Figure 3, which includes information on the atomic species and three-dimensional position coordinates of each of the multiple atoms that make up the molecule. In the example in Figure 3, each atom is represented by a circle, and α, β, etc., within the circle represent the atomic species.

[0022] In Reference Method 1, a predetermined proportion of atoms in the prepared molecular data are replaced with other atoms. In the example in Figure 3, the shaded circles represent the atoms that have been replaced from the original molecular data. Then, the NN is made to predict whether each atom in the substituted molecular data has been replaced from each atom in the original molecular data. In other words, Reference Method 1 sets up an output unit (for pre-training) for performing a binary classification task. Since Reference Method 1 can generate ground truth data from the difference between the original molecular data and the substituted molecular data, self-supervised learning becomes possible. In the example in Figure 3, atoms that have not been replaced are represented by "0" and atoms that have been replaced are represented by "1" in the ground truth data and prediction results.

[0023] Reference Method 2, as shown in Figure 4, replaces a predetermined proportion of atoms in the prepared molecular data with other atoms, and then has the NN predict the original atom species before the substitution for each atom included in the substituted molecular data. In other words, Reference Method 2 sets up an output unit (for pre-training) for performing a multi-class classification task. In Reference Method 2, the original molecular data becomes the ground truth data, making self-supervised learning possible.

[0024] In the examples in Figures 3 and 4, when referring to the ground truth data, the prediction results by the neural network (NN) are incorrect for the atoms indicated by the shaded areas. The NN parameters are updated so that these prediction results match the ground truth data.

[0025] In both of the reference methods 1 and 2, as shown in Figure 5, the feature extraction unit of the NN, pre-trained by self-supervised learning, is used as the initial value, and fine tuning is performed with the actual training data. At the fine tuning stage, it is assumed that the feature extraction unit of the NN has learned essential features such as the structure of atoms and molecules. Also, since the output unit of the NN during pre-training is for predicting binary or multi-class classification tasks, in fine tuning, an output unit corresponding to the actual task, such as predicting force or energy, is set.

[0026] In reference methods 1 and 2, as shown in Figure 6, the atomic species of the atoms in the original molecular data (shaded circles in Figure 6) are replaced with atomic species randomly selected from a pre-defined group of atomic species. In this case, since the replacement is with a unique atom that has a different tendency from the other atoms in the molecule, it is easy to identify which atom has been replaced. Furthermore, it is easy to determine the original atomic species of the replaced atom from the atomic species of the atoms surrounding the unique atom (thick-lined circles in Figure 6). As a result, it becomes possible to predict the correct answer with the NN without learning the essential characteristics of the entire molecular structure, which reduces the accuracy of the NN as a tool for predicting material properties.

[0027] Therefore, in this embodiment, as shown in Figure 7, molecular data for self-supervised learning is generated without creating unique atoms within the molecule by shuffling the atoms contained in the original molecular data. In the lower part of Figure 7, the shaded atoms represent atomic species α in the original molecular data being replaced by atomic species β, and the atoms with thick lines represent atomic species β in the original molecular data being replaced by atomic species α. In Figures 6 and 7, α and β represent atomic species such as atomic number and element symbol. The same applies to α, β, and γ in the following figures. The self-supervised learning apparatus according to each embodiment will be described in detail below.

[0028] <First Embodiment> As shown in Figure 8, the self-supervised learning device 10 according to the first embodiment functionally includes a generation unit 12, an acquisition unit 14, and an update unit 16. A neural network (NN) 24 is stored in a predetermined memory area of ​​the self-supervised learning device 10. The NN 24 is an example of a "machine learning model" of the disclosed technology.

[0029] As shown in Figure 8, the self-supervised learning device 10 receives first molecular data. The first molecular data includes, for example, information on the atomic species and three-dimensional position coordinates of each atom contained in the first molecule. The first molecular data may also include information indicating the bonding relationships between atoms.

[0030] The generation unit 12 acquires the first molecular data input to the self-supervised learning device 10. The generation unit 12 generates second molecular data representing a second molecule in which each of the multiple atoms contained in the first molecule is replaced with any of the atoms contained in the first molecule. Specifically, the generation unit 12 generates the second molecular data by shuffling the atoms contained in the first molecule in a complete permutation.

[0031] For example, the generation unit 12 stores the first molecular data as an atomic species list and a coordinate list. The atomic species list is a list of the length (number of elements) corresponding to the number of atoms contained in the first molecule, and the value of each element is a symbol representing the atomic species, such as the atomic number or element symbol. The coordinate list is a list of the size (number of atoms × 3) representing the three-dimensional position coordinates of each atom contained in the first molecule. The three-dimensional position coordinates of the atom corresponding to the i-th element of the atomic species list are stored in the i-th element of the coordinate list.

[0032] As shown in Figure 9A, the generation unit 12 assigns indices 0, 1, 2..., N-1 (where N is the number of atoms in the first molecule) to the original atom list in order from the beginning. In Figure 9, the numbers written next to each element in the list are the indices corresponding to that element. The generation unit 12 generates random integers from 0 to N-1, rearranges the integers assigned as indices to the original atom list in a complete permutation, and associates them in order from the beginning of the substituted atom list. The substituted atom list is a list with the same number of elements as the original atom list, and at this stage, each element is empty. As shown in Figure 9B, the generation unit 12 generates a shuffled atom list by storing the atom species stored in the elements of the original atom list that correspond to the index of that element in each element of the substituted atom list. The generation unit 12 generates the second molecule data by combining the shuffled atom list with the coordinate list of the first molecule data.

[0033] NN24 includes a feature extraction unit that extracts features representing the molecular structure of a second molecule from second molecular data, and an output unit that outputs prediction results according to the pre-training task based on the features extracted by the feature extraction unit. The pre-training task may be a binary classification task, as shown in Figure 3, which predicts whether each atom in the second molecule has been replaced by each atom in the first molecule. Alternatively, the pre-training task may be a multi-class classification task, as shown in Figure 4, which predicts the original atom before replacement for each atom in the second molecular data.

[0034] The acquisition unit 14 inputs the second molecular data generated by the generation unit 12 to the NN 24 and acquires the prediction results output from the NN 24. In the case of the binary classification task described above, the prediction results are information indicating whether or not each atom in the second molecule has been substituted from the first molecule, and in the case of the multi-class classification task, they are information indicating the original atomic species of each atom in the second molecule before substitution.

[0035] The update unit 16 updates the parameters of the NN24 based on the comparison result between the correct data corresponding to the first molecule and the prediction result obtained by the acquisition unit 14. Specifically, in the case of the binary classification task described above, the update unit 16 generates correct data in which, for corresponding atoms with the same 3D position coordinates in the first molecular data and the second molecular data, the atomic species is the same and the value is [0], and if they are different, the value is [1]. In the case of the multi-class classification task described above, the update unit 16 uses the first molecular data as the correct data. The update unit 16 then backpropagates the difference between the prediction result and the correct data to the NN24 and updates the parameters of the NN24.

[0036] The update unit 16 repeatedly updates the parameters of the NN24 until a predetermined termination condition is met. The predetermined termination condition may be, for example, when the number of repetitions reaches a predetermined number, when the difference between the prediction result and the correct data becomes less than or equal to a predetermined value, or when the difference between the previous difference and the current difference becomes less than or equal to a predetermined value (when the difference converges).

[0037] The self-supervised learning device 10 may be implemented, for example, by the computer 40 shown in Figure 10. The computer 40 includes a CPU (Central Processing Unit) 41, a GPU (Graphics Processing Unit) 42, a memory 43 as a temporary storage area, and a non-volatile storage device 44. The computer 40 also includes input / output devices 45 such as input devices and display devices, and an R / W (Read / Write) device 46 that controls the reading and writing of data to and from the storage medium 49. The computer 40 also includes a communication interface 47 that connects to a network such as the Internet. The CPU 41, GPU 42, memory 43, storage device 44, input / output devices 45, R / W device 46, and communication interface 47 are connected to each other via a bus 48.

[0038] The storage device 44 is, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), or flash memory. The storage device 44 stores a self-supervised learning program 50 that causes the computer 40 to function as a self-supervised learning device 10. The self-supervised learning program 50 includes a generation process control instruction 52, an acquisition process control instruction 54, and an update process control instruction 56. The storage device 44 also has an information storage area 60 in which information constituting the NN24 is stored.

[0039] The CPU 41 reads the self-supervised learning program 50 from the storage device 44, loads it into memory 43, and sequentially executes the control instructions contained in the self-supervised learning program 50. The CPU 41 operates as the generation unit 12 shown in Figure 8 by executing the generation process control instruction 52. The CPU 41 also operates as the acquisition unit 14 shown in Figure 8 by executing the acquisition process control instruction 54. The CPU 41 also operates as the update unit 16 shown in Figure 8 by executing the update process control instruction 56. The CPU 41 also reads information from the information storage area 60 and loads the NN 24 into memory 43. As a result, the computer 40 that executed the self-supervised learning program 50 functions as a self-supervised learning device 10. The CPU 41 that executes the program is hardware. Also, part of the program may be executed by the GPU 42.

[0040] Furthermore, the functions realized by the self-supervised learning program 50 may be implemented, for example, by semiconductor integrated circuits, more specifically by ASICs (Application Specific Integrated Circuits), FPGAs (Field-Programmable Gate Arrays), etc.

[0041] Next, the operation of the self-supervised learning device 10 according to the first embodiment will be described. When the first molecular data is input to the self-supervised learning device 10 and the execution of self-supervised learning of NN24 is instructed, the self-supervised learning process shown in Figure 11 is executed in the self-supervised learning device 10. Note that the self-supervised learning process is an example of the self-supervised learning method of the disclosed technology.

[0042] In step S10, the generation unit 12 acquires the first molecular data input to the self-supervised learning device 10 and creates copies, one for generating the second molecular data and the other for generating the correct answer data. Next, in step S12, the generation unit 12 generates the second molecular data by shuffling the atoms contained in the first molecule in a complete permutation.

[0043] Next, in step S14, the acquisition unit 14 inputs the generated second molecular data into NN24 to obtain a prediction result regarding the molecular structure. Next, in step S16, the update unit 16 generates correct answer data based on the first molecular data. Next, in step S18, the update unit 16 backpropagates the difference between the prediction result and the correct answer data to NN24 to update the parameters of NN24.

[0044] Next, in step S20, the update unit 16 determines whether the termination condition for parameter update is satisfied. If the termination condition is satisfied, the process proceeds to step S22; if not, the process returns to step S10. In step S22, the update unit 16 stores NN24 with the finally set parameters in a predetermined storage area, and the self-supervised learning process ends.

[0045] As described above, the self-supervised learning device according to the first embodiment generates second molecular data in which each of the plurality of atoms included in the first molecule is replaced with any one of the atoms included in the first molecule. Further, the self-supervised learning device inputs the second molecular data into an NN, which is a machine learning model for predicting the molecular structure, to obtain a prediction result, and updates the parameters of the NN based on the comparison result between the correct answer data corresponding to the first molecule and the prediction result. Thus, instead of replacing an atom using an atomic species randomly selected from a preset atomic species group, the atoms included in the original molecular data are shuffled to generate second molecular data. As a result, there are no specific atoms in the second molecule, so it becomes a problem of predicting the original atoms from the molecular structure such as the interatomic distance and the angle of the position of other atoms viewed from a certain atom. That is, it becomes a problem that requires knowledge of the entire molecular structure rather than information in the vicinity of a certain atom, accelerating the learning of the essential feature quantities of the molecular structure, and enabling the generation of a highly accurate machine learning model by self-supervised learning.

[0046] In the first embodiment, the case of generating the second molecular data by shuffling the atoms included in the first molecule in a complete permutation has been described, but the present invention is not limited to this. In the case of shuffling in a complete permutation as in the first embodiment, before and after the shuffling, since there is no change in the composition ratio, which is the ratio of the number of atoms of each atomic species to the number of atoms included in the first molecule, the composition ratio of the atomic species can be a hint for prediction and may facilitate the prediction. Therefore, for example, by generating random numbers from 0 to N-1 allowing duplicates, the atoms included in the first molecule may be shuffled allowing duplicates of the atoms to be replaced.

[0047] For example, in the example of FIG. 12, the index "5" of the original atomic list is selected repeatedly in the atomic list after substitution. As a result, the atomic species "γ" that appears only once in the original atomic list appears twice in the shuffled atomic list. In this way, by changing the composition ratio before and after shuffling to make the problem more difficult to predict, a more accurate machine learning model can be generated.

[0048] Also, in the first embodiment, the case of shuffling all the atoms included in the first molecule has been described, but the present invention is not limited to this. For example, the ratio of the number of atoms to be replaced with other atomic species to the number of atoms included in the first molecule (hereinafter referred to as "substitution rate") is set in advance as a hyperparameter, and only some of the atoms selected based on the substitution rate may be shuffled. However, when shuffling all the atoms as in the first embodiment, hyperparameters such as the substitution rate are not necessary. Therefore, there is no need to tune the hyperparameters, and a pre-trained model with a certain accuracy can be generated for any machine learning model and dataset.

[0049] <Second Embodiment> Next, the second embodiment will be described. In the self-supervised learning device according to the second embodiment, the same components as those in the self-supervised learning device 10 according to the first embodiment are denoted by the same reference numerals, and for functional parts having some common functions, reference numerals having the same last two digits are given, and detailed description thereof is omitted.

[0050] As shown in Figure 8, the self-supervised learning device 210 according to the second embodiment functionally includes a generation unit 212, an acquisition unit 14, and an update unit 16. Furthermore, NN24 is stored in a predetermined storage area of ​​the self-supervised learning device 210.

[0051] Here, atoms with a small proportion have a high probability of being replaced by other atoms after shuffling, while atoms with a large proportion have a low probability of being replaced by other atoms after shuffling. For example, in the original atom list shown in Figure 9A, the proportion of atom species α is 3 / 6 = 1 / 2, the proportion of atom species β is 2 / 6 = 1 / 3, and the proportion of atom species γ is 1 / 6. The probability that atom species α is replaced by another atom species is 3 / 6 = 1 / 2, the probability that atom species β is replaced by another atom species is 4 / 6 = 2 / 3, and the probability that atom species γ is replaced by another atom species is 5 / 6.

[0052] Thus, the fact that the probability of substitution by other atomic species differs depending on the composition ratio can serve as a hint during self-supervised learning and may facilitate prediction. Therefore, in the second embodiment, shuffling is performed in a way that does not cause variation in the probability of substitution by other atomic species for each atomic species.

[0053] The generation unit 212 selects atoms to be substituted so that the ratio of the number of atoms substituted with a different atomic species to the number of atoms of each atomic species contained in the first molecule (hereinafter referred to as the "substitution rate for each atomic species") satisfies predetermined conditions. The generation unit 212 then generates second molecular data by substituting the atomic species of the atoms to be substituted with other atomic species. The predetermined conditions may be, for example, that the substitution rate for each atomic species is a predetermined ratio for each composition ratio, or that it is a ratio randomly generated from a Gaussian distribution.

[0054] For example, suppose the generating unit 212 selects atoms corresponding to the elements shown by the thick lines in the original atom list shown in Figure 13A as the atoms to be replaced. In this case, the generating unit 212 may select atomic species β or γ in a 2:1 ratio for atomic species α, atomic species α or γ in a 3:1 ratio for atomic species β, and atomic species α or β in a 3:2 ratio for atomic species γ as the replacement atomic species.

[0055] Furthermore, the generation unit 212 may generate second molecular data by shuffling the atoms to be substituted in a complete permutation, similar to the generation unit 12 of the first embodiment. Figure 13B shows an example of the shuffled atom list in this case. Alternatively, the generation unit 212 may generate second molecular data by shuffling the atoms to be substituted, allowing for duplication. Figure 13C shows an example of the shuffled atom list in this case.

[0056] The self-supervised learning device 210 may be implemented, for example, by the computer 40 shown in Figure 10. The storage device 44 of the computer 40 stores a self-supervised learning program 250 that causes the computer 40 to function as the self-supervised learning device 210. The self-supervised learning program 250 includes a generation process control instruction 252, an acquisition process control instruction 54, and an update process control instruction 56. The storage device 44 also has an information storage area 60 in which information constituting the NN24 is stored.

[0057] The CPU 41 reads the self-supervised learning program 250 from the storage device 44, loads it into memory 43, and sequentially executes the control instructions contained in the self-supervised learning program 250. By executing the generation process control instruction 252, the CPU 41 operates as the generation unit 212 shown in Figure 8. The other control instructions are the same as those for the self-supervised learning program 50 according to the first embodiment. As a result, the computer 40 that has executed the self-supervised learning program 250 functions as a self-supervised learning device 210.

[0058] Furthermore, the functions realized by the self-supervised learning program 250 may be implemented, for example, using semiconductor integrated circuits, more specifically ASICs, FPGAs, etc.

[0059] Next, the operation of the self-supervised learning device 210 according to the second embodiment will be described. When the first molecular data is input to the self-supervised learning device 210 and the execution of self-supervised learning of NN24 is instructed, the self-supervised learning process shown in Figure 14 is executed in the self-supervised learning device 210. In the self-supervised learning process in the second embodiment, the same process as the self-supervised learning process in the first embodiment (Figure 11) is given the same step number and a detailed explanation is omitted.

[0060] After step S10, in the next step S210, the generation unit 212 selects atoms to be substituted so that the substitution rate for each atom species satisfies predetermined conditions. Next, in step S212, the generation unit 212 generates second molecular data by substituting the atom species of the atoms to be substituted with other atom species. Thereafter, steps S14 to S22 are executed in the same manner as in the first embodiment, and the self-supervised learning process is completed.

[0061] As described above, the self-supervised learning device according to the second embodiment selects atoms to be substituted when generating second molecular data such that the ratio of atoms to be substituted with different atomic species to the number of atoms of each atomic species contained in the first molecule satisfies a predetermined condition. The self-supervised learning device then generates second molecular data by substituting the atomic species of the atoms to be substituted with other atomic species. This makes the problem difficult to predict, as the substitution rate for each atomic species according to the composition ratio does not serve as a hint during self-supervised learning, and a highly accurate machine learning model can be generated by self-supervised learning.

[0062] <Third Embodiment> Next, a third embodiment will be described. In the self-supervised learning device according to the third embodiment, the same reference numerals are used for components similar to those in the self-supervised learning device 10 according to the first embodiment, and for functional parts that share some functions, the last two digits of the reference numerals are shared, and detailed explanations are omitted.

[0063] As shown in Figure 8, the self-supervised learning device 310 according to the third embodiment functionally includes a generation unit 312, an acquisition unit 14, and an update unit 16. Furthermore, NN24 is stored in a predetermined storage area of ​​the self-supervised learning device 310.

[0064] The generation unit 312 includes adjusting the substitution rate, which is the ratio of the number of atoms substituted with other atomic species to the number of atoms contained in the first molecule, so that it falls within a predetermined range. For example, when the generation unit 312 generates second molecule data by shuffling the atom list, as in the first embodiment, it compares the original atom list with the shuffled atom list and calculates the substitution rate. If the calculated substitution rate is above a predetermined upper threshold, i.e., if it has been shuffled too much, the generation unit 312 returns some of the atoms whose atomic species have changed back to their original atomic species. Also, if the calculated substitution rate is below a predetermined lower threshold, i.e., if it has not been shuffled enough, the generation unit 312 replaces some of the atoms whose atomic species have not changed with other atomic species.

[0065] The self-supervised learning device 310 may be implemented, for example, by the computer 40 shown in Figure 10. The storage device 44 of the computer 40 stores a self-supervised learning program 350 that causes the computer 40 to function as the self-supervised learning device 310. The self-supervised learning program 350 includes a generation process control instruction 352, an acquisition process control instruction 54, and an update process control instruction 56. The storage device 44 also has an information storage area 60 in which information constituting the NN24 is stored.

[0066] The CPU 41 reads the self-supervised learning program 350 from the storage device 44, loads it into memory 43, and sequentially executes the control instructions contained in the self-supervised learning program 350. By executing the generation process control instruction 352, the CPU 41 operates as the generation unit 312 shown in Figure 8. The other control instructions are the same as those for the self-supervised learning program 50 according to the first embodiment. As a result, the computer 40 that has executed the self-supervised learning program 350 functions as a self-supervised learning device 310.

[0067] Furthermore, the functions realized by the self-supervised learning program 350 may be implemented, for example, by semiconductor integrated circuits, more specifically by ASICs, FPGAs, etc.

[0068] Next, the operation of the self-supervised learning device 310 according to the third embodiment will be described. When the first molecular data is input to the self-supervised learning device 310 and the execution of self-supervised learning of NN24 is instructed, the self-supervised learning process shown in Figure 15 is executed in the self-supervised learning device 310. In the self-supervised learning process in the third embodiment, the same process as the self-supervised learning process in the first embodiment (Figure 11) is given the same step number and a detailed explanation is omitted.

[0069] After going through steps S10 to S12, in the next step S310, the generating unit 312 calculates the substitution rate, which is the ratio of the number of atoms substituted with other atomic species to the number of atoms contained in the first molecule. The generating unit 312 then determines whether the calculated substitution rate is within a predetermined range. If the substitution rate is within the predetermined range, the process proceeds to step S14; if the substitution rate is above the upper threshold or below the lower threshold of the predetermined range, the process proceeds to step S312.

[0070] In step S312, if the substitution rate is above the upper threshold, the generation unit 312 restores some of the atoms whose atomic species have changed back to their original atomic species. If the substitution rate is below the lower threshold, the generation unit 312 replaces some of the atoms whose atomic species have not changed with other atomic species. In this way, the generation unit 312 adjusts the substitution. Thereafter, steps S14 to S22 are executed in the same manner as in the first embodiment, and the self-supervised learning process is completed.

[0071] As described above, the self-supervised learning device according to the third embodiment adjusts the substitution rate, which is the ratio of the number of atoms substituted with other atomic species to the number of atoms contained in the first molecule, to a predetermined range when generating the second molecular data. This allows for adjustment of the degree of unpredictability during self-supervised learning and enables the generation of an appropriate machine learning model.

[0072] In the third embodiment, a case was described in which the first molecular data is shuffled to generate the second molecular data, and then the substitution adjustments are made, but the invention is not limited to this. Before shuffling, atoms to be substituted may be selected so that the substitution rate falls within a predetermined range, and the atomic species of the selected atoms may be substituted with other atomic species.

[0073] <Fourth Embodiment> Next, a fourth embodiment will be described. In the self-supervised learning device according to the fourth embodiment, the same reference numerals are used for components similar to those in the self-supervised learning device 10 according to the first embodiment, and for functional parts that share some functions, the last two digits of the reference numerals are shared, and detailed explanations are omitted.

[0074] As shown in Figure 8, the self-supervised learning device 410 according to the fourth embodiment functionally includes a generation unit 412, an acquisition unit 14, and an update unit 16. Furthermore, NN24 is stored in a predetermined storage area of ​​the self-supervised learning device 410.

[0075] The generation unit 412 generates multiple candidate second molecular data by repeatedly substituting each of the multiple atoms contained in the first molecule with any of the atoms contained in the first molecule. Then, the generation unit 412 selects the best candidate as the second molecular data based on an evaluation according to the substitution rate of each atom species. The generation unit 412 may perform an evaluation according to the substitution rate of each atom species based on how well the substitution rate of each atom species satisfies predetermined conditions. For example, the generation unit 412 may evaluate candidates as better the smaller the variation in the substitution rate of each atom species, or it may evaluate candidates as better the more the substitution rate of each atom species follows a Gaussian distribution.

[0076] The self-supervised learning device 410 may be implemented, for example, by the computer 40 shown in Figure 10. The storage device 44 of the computer 40 stores a self-supervised learning program 450 that causes the computer 40 to function as the self-supervised learning device 410. The self-supervised learning program 450 includes a generation process control instruction 452, an acquisition process control instruction 54, and an update process control instruction 56. The storage device 44 also has an information storage area 60 in which information constituting the NN24 is stored.

[0077] The CPU 41 reads the self-supervised learning program 450 from the storage device 44, loads it into memory 43, and sequentially executes the control instructions contained in the self-supervised learning program 450. By executing the generation process control instruction 452, the CPU 41 operates as the generation unit 412 shown in Figure 8. The other control instructions are the same as those for the self-supervised learning program 50 according to the first embodiment. As a result, the computer 40 that has executed the self-supervised learning program 450 functions as a self-supervised learning device 410.

[0078] Furthermore, the functions realized by the self-supervised learning program 450 may be implemented, for example, by semiconductor integrated circuits, more specifically by ASICs, FPGAs, etc.

[0079] Next, the operation of the self-supervised learning device 410 according to the fourth embodiment will be described. When the first molecular data is input to the self-supervised learning device 410 and the execution of self-supervised learning of NN24 is instructed, the self-supervised learning process shown in Figure 16 is executed in the self-supervised learning device 410. In the self-supervised learning process in the fourth embodiment, the same process as the self-supervised learning process in the first embodiment (Figure 11) is given the same step number and a detailed explanation is omitted.

[0080] After step S10, in the next step S410, the generation unit 412 performs a process multiple times in which it replaces each of the multiple atoms contained in the first molecule with any of the atoms contained in the first molecule, thereby generating multiple candidates for second molecular data. Next, in step S412, the generation unit 412 selects the best candidate as the second molecular data based on an evaluation according to the substitution rate for each atom species. Thereafter, steps S14 to S22 are executed in the same manner as in the first embodiment, and the self-supervised learning process is completed.

[0081] As described above, the self-supervised learning device according to the fourth embodiment generates a plurality of candidate second molecular data sets and selects the best candidate as the second molecular data set based on an evaluation according to the substitution rate of each atomic species. This makes it possible to generate second molecular data suitable for self-supervised learning and to generate a machine learning model with high accuracy.

[0082] Here, we will explain the verification results of the effects of each of the above embodiments. In this verification, the following methods were tested.

[0083] Comparison Method 1: A method for supervising NN training using only labeled data (no self-supervised training). Comparison Method 2: A method for fine-tuning an NN pre-trained using self-supervised training (predicting whether atoms have been substituted or not) as described in Reference Method 1, using labeled data. This Method 1: A method for fine-tuning an NN pre-trained using the method of the first embodiment (predicting the original atoms) using labeled data.

[0084] Furthermore, in comparison method 2, after confirming that it is more accurate than the task of predicting the original atoms, the task of predicting whether or not atoms have been substituted is adopted.

[0085] In this validation, assuming that a small amount of data is used for fine-tuning, evaluation was performed using a portion of the publicly available dataset. For self-supervised learning, 460,328 training data points and 24,943 validation data points were used. For fine-tuning, 1,120 training data points and 2,624 validation data points were extracted from the total of 45,890 data points. Furthermore, a pre-trained model was generated by performing self-supervised learning for 4 epochs, and a machine learning model that predicts energy from molecular structure was generated by performing fine-tuning for 200 epochs. Note that in comparison method 1, only the machine learning equivalent to fine-tuning was performed.

[0086] The mean absolute error between the energy predicted for the validation data by the machine learning model generated through fine tuning and the energy label was used as an evaluation value for comparison. In other words, a lower evaluation value indicates higher accuracy. The table below shows the evaluation values ​​after fine tuning. As shown in the table below, an improvement in accuracy was observed with this method.

[0087]

[0088] In each of the above embodiments, the self-supervised learning program is pre-stored (installed) in the storage device, but the invention is not limited thereto. The program relating to the disclosed technology may be provided in a form stored on a storage medium such as a CD-ROM, DVD-ROM, or USB memory.

[0089] 10, 210, 310, 410 Self-supervised learning device 12, 212, 312, 412 Generation unit 14 Acquisition unit 16 Update unit 22 Atomic library 24 NN 40 Computer 41 CPU 42 GPU 43 Memory 44 Storage device 45 Input / output device 46 R / W device 47 Communication I / F 48 Bus 49 Storage medium 50, 250, 350, 450 Self-supervised learning program 52, 252, 352, 452 Generation process control instruction 54 Acquisition process control instruction 56 Update process control instruction 60 Information storage area

Claims

1. A self-supervised learning program for causing a computer to perform the following processes: generate data representing a second molecule in which each of a plurality of atoms contained in a first molecule is replaced with any of the atoms contained in the first molecule; input the data representing the second molecule into a machine learning model that makes predictions about molecular structure and obtain prediction results; and update the parameters of the machine learning model based on the comparison results between the correct data corresponding to the first molecule and the prediction results.

2. The self-supervised learning program according to claim 1, wherein the process for generating data representing the second molecule includes shuffling the atoms contained in the first molecule in a complete permutation.

3. The self-supervised learning program according to claim 1, wherein the process for generating data representing the second molecule includes shuffling the atoms contained in the first molecule, allowing for duplication of the atoms to be substituted.

4. The self-supervised learning program according to claim 1, wherein the process for generating data representing the second molecule includes selecting atoms to be substituted such that the ratio of the number of atoms to be substituted with a different atomic species to the number of atoms of each atomic species contained in the first molecule satisfies a predetermined condition, and substituting the atomic species of the atoms to be substituted with other atomic species.

5. The self-supervised learning program according to any one of claims 1 to 4, wherein the process for generating data representing the second molecule includes adjusting the ratio of the number of atoms substituted with other atomic species to the number of atoms contained in the first molecule to be within a predetermined range.

6. The self-supervised learning program according to any one of claims 1 to 4, wherein the process for generating data representing the second molecule involves performing a process multiple times in which each of the multiple atoms contained in the first molecule is replaced with any of the atoms contained in the first molecule, thereby generating multiple candidate data representing the second molecule, and selecting the best candidate as data representing the second molecule based on an evaluation according to the ratio of the number of replaced atoms for each atom species.

7. The self-supervised learning program according to any one of claims 1 to 4, wherein the machine learning model predicts whether each of the atoms contained in the second molecule is substituted for each of the atoms contained in the first molecule.

8. The self-supervised learning program according to any one of claims 1 to 4, wherein the machine learning model predicts the original atoms of the atoms that have been substituted from the atoms in the first molecule among the atoms in the second molecule.

9. A self-supervised learning method in which a computer performs the following steps: generate data representing a second molecule in which each of a plurality of atoms contained in a first molecule is replaced with any of the atoms contained in the first molecule; input the data representing the second molecule into a machine learning model that makes predictions about molecular structure to obtain prediction results; and update the parameters of the machine learning model based on the comparison results between the correct data corresponding to the first molecule and the prediction results.

10. The self-supervised learning method according to claim 9, wherein the process for generating data representing the second molecule includes shuffling the atoms contained in the first molecule in a complete permutation.

11. The self-supervised learning method according to claim 9, wherein the process for generating data representing the second molecule includes shuffling the atoms contained in the first molecule, allowing for duplication of the atoms to be substituted.

12. The self-supervised learning method according to claim 9, wherein the process for generating data representing the second molecule includes selecting atoms to be substituted such that the ratio of the number of atoms to be substituted with a different atomic species to the number of atoms of each atomic species contained in the first molecule satisfies a predetermined condition, and substituting the atomic species of the atoms to be substituted with other atomic species.

13. The self-supervised learning method according to any one of claims 9 to 12, wherein the process for generating data representing the second molecule includes adjusting the ratio of the number of atoms substituted with other atomic species to the number of atoms contained in the first molecule to be within a predetermined range.

14. The self-supervised learning method according to any one of claims 9 to 12, wherein the process for generating data representing the second molecule involves performing a process multiple times in which each of the multiple atoms contained in the first molecule is replaced with any of the atoms contained in the first molecule to generate multiple candidates for data representing the second molecule, and selecting the best candidate as data representing the second molecule based on an evaluation according to the ratio of the number of replaced atoms for each atom species.

15. The self-supervised learning method according to any one of claims 9 to 12, wherein the machine learning model predicts whether each of the atoms contained in the second molecule is substituted for each of the atoms contained in the first molecule.

16. The self-supervised learning method according to any one of claims 9 to 12, wherein the machine learning model predicts the original atom of the atom that was replaced from the atom contained in the first molecule among the atoms contained in the second molecule.

17. A self-supervised learning device comprising: a generation unit that generates data representing a second molecule obtained by substituting each of a plurality of atoms contained in a first molecule with any of the atoms contained in the first molecule; an acquisition unit that inputs the data representing the second molecule into a machine learning model that makes predictions about molecular structure and obtains prediction results; and an update unit that updates the parameters of the machine learning model based on a comparison result between the correct data corresponding to the first molecule and the prediction results.

18. The self-supervised learning device according to claim 17, wherein the generating unit shuffles the atoms contained in the first molecule in a complete permutation.

19. The self-supervised learning device according to claim 17, wherein the generating unit shuffles the atoms contained in the first molecule, allowing for duplication of the atoms to be substituted.

20. The self-supervised learning device according to claim 17, wherein the generating unit selects atoms to be substituted such that the ratio of the number of atoms to be substituted with a different atomic species to the number of atoms of each atomic species contained in the first molecule satisfies a predetermined condition, and replaces the atomic species of the atoms to be substituted with other atomic species.

Images