Information processing device, model generation method, and information processing method

Abstract

[Problem] To propose an application field using a neural network model that forms an NNP. [Solution] This information processing device comprises one or more memories and one or more processors. The one or more processors input information concerning an atom of a substance to a first model, and acquire information concerning the substance from an output layer of the first model. The first model is a model that: is provided with layers ranging from an input layer to a predetermined layer in a second model for receiving input of information concerning atoms and outputting a value of at least one of energy or force; and is trained to output information concerning the substance.

Description

Information processing device, model generation method, and information processing method

[0001] The present disclosure relates to an information processing device, a model generation method, and an information processing method.

[0002] NNP (Neural Network Potential), which is a neural network model trained based on data obtained from quantum chemical calculations, etc., is now beginning to be used to calculate force fields (energy, force) in the field of atomic simulations.

[0003] International Publication No. 2021 / 054402

[0004] This disclosure proposes application areas using neural network models forming NNPs.

[0005] According to one embodiment, the information processing device includes one or more memories and one or more processors. The one or more processors input information about atoms of a substance to a first model and obtain information about the substance from an output layer of the first model. The first model is a model trained to output information about the substance, including layers from an input layer to a predetermined layer in a second model that receives information about atoms and outputs at least a value of either energy or force.

[0006] FIG. 1 is a diagram showing an example of a model forming an NNP according to an embodiment. FIG. 2 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 3 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 4 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 5 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 6 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 7 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 8 is a diagram showing an example of the configuration of a neural network model according to an embodiment. FIG. 9 is a diagram showing an example of the configuration of a neural network model according to an embodiment.

[0007] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The following describes embodiments of the present invention with reference to the drawings. The drawings and the description of the embodiments are given by way of example only and are not intended to limit the present invention.

[0008] FIG. 1 is a diagram showing an example of a network model configuration of an NNP (Neural Network Potential) according to an embodiment. The model forming the NNP may be configured using, for example, an MLP (Multi-Layer Perceptron). This MLP may be, for example, a graph neural network capable of inputting and outputting graphs. The intermediate layer of each model shown below may be configured to function as a part of the graph neural network, i.e., capable of inputting and outputting graph information.

[0009] The model that forms the NNP shown in Figure 1 is a model that has been trained using, for example, interatomic interactions (energy) obtained as a result of quantum chemical calculations as training data. When information about the atoms that make up a substance is input to the input layer, the model that forms the NNP outputs energy from the output layer.

[0010] For example, in this embodiment, the nodes in the input layer of the model that forms the NNP correspond to atoms that make up a substance, and each node receives information about the atoms of the substance. Similarly, the output layer of the model that forms the NNP uses the information for each node to output the energy in the input state. By backpropagating this energy, it is also possible to obtain the force that each atom receives.

[0011] The atomic information input into the model that forms the NNP includes, for example, information about the type and position of each atom. In this specification, atomic information may be referred to as "atomic information." Examples of atomic position information include information that directly indicates the position of atoms using coordinates, and information that directly or indirectly indicates the relative positions between atoms. This information is expressed, for example, by the distance, angle, dihedral angle, etc. between atoms.

[0012] For example, by calculating the distance between two atoms or the angle between three atoms from atomic coordinate information and using this as atomic position information to input into the model that forms the NNP, it is possible to ensure invariance to rotation and translation, thereby improving the accuracy of the NNP. For example, atomic information may be information that directly indicates position, or information calculated from that position information. Furthermore, atomic information may include information on the type and position of the atom, as well as information on charge and bonding.

[0013] Generally, neural network models are gradually updated from input information to output information at each layer. Therefore, the output from any intermediate layer in a model that forms an NNP can be considered to be a quantity that has the characteristics connecting atomic information and energy information.

[0014] In this embodiment, a neural network model capable of inferring the properties of a substance is formed using the output from the intermediate layer in the model forming the NNP.

[0015] The processing circuit of the information processing device that performs model training may perform training by changing the output from the output layer of the model that forms the NNP shown in Figure 1 to another feature. The processing circuit may use a model that has been trained in advance as a model that forms the NNP, or may form and train a new neural network model that constitutes the NNP.

[0016] FIG. 2 is a diagram showing a non-limiting example of model formation according to this embodiment. The information processing device of this embodiment acquires various physical property values ​​using a model having a predetermined intermediate layer of the model (hereinafter referred to as the second model) that forms the NNP shown in FIG. 1. This model may be a model (hereinafter referred to as the first model) that outputs physical property values ​​different from the information (energy, force) that can be acquired via the NNP. This disclosure will mainly describe the configuration of this first model.

[0017] Generally, each layer of the second model has nodes equal to the number of atoms. That is, the second model has nodes equal to the number of atoms from the input layer to the output layer. Therefore, it can be assumed that the output of each intermediate layer outputs some feature corresponding to each atom. In this disclosure, the output from the intermediate layer of the second model is mainly used to connect to a network that outputs a different characteristic, and further training is performed to obtain a model that infers the different characteristic.

[0018] FIG. 3 is a diagram showing an outline of an example of generating the first model in FIGS. 1 and 2. As shown in FIG. 3, the first model can output information different from that of the second model by copying the second model's input layer through a predetermined intermediate layer and connecting a new output layer. In this case, a neural network model that acquires desired physical property values ​​can be generated by training (transfer learning) parameters indicating the connection between the predetermined intermediate layer and the output layer using appropriate training data. In FIG. 3, the predetermined intermediate layer is the layer immediately preceding the output layer of the second model.

[0019] The configuration in Figure 3 has two hidden layers, but this is not limited to this. For example, the second model may have multiple layers up to the predetermined hidden layer, or the layer next to the input layer may be the predetermined hidden layer. The first model can be optimized (trained) by transfer learning by appropriately copying the parameters from the input layer to the predetermined hidden layer of the second model.

[0020] The first model is not limited to having a separate output layer connected to a predetermined intermediate layer of the second model, and the predetermined intermediate layer of the second model may be configured to be the output layer of the first model.

[0021] In this embodiment, the first model and the second model output different information, i.e., the first model outputs information other than energy and force, but this is not limited to this. As another non-limiting example, the first model may output the same type of information as the second model. By using the first model, the processing circuit is expected to obtain, for example, the same physical property values ​​(substantially the same physical property values) or similar physical property values ​​as the second model using a model with lower computational cost than output from the output layer of the second model.

[0022] The first and second models may also output different types of energy or force. For example, if the second model, which forms NNPs, outputs "total energy," the first model may infer physical properties such as adsorption energy or activation energy.

[0023] FIG. 4 is a diagram showing another example of first model generation. As shown in this figure, the first model is configured to include an additional intermediate layer between the predetermined intermediate layer and the output layer. In this way, the first model may include a different second intermediate layer rather than directly propagating information from the predetermined intermediate layer to the output layer. This allows more complex calculations to be performed on the features output from the predetermined intermediate layer, and depending on the target of inference, better results may be obtained compared to FIG. 3.

[0024] In the first model shown in Figure 4, there is one hidden layer between the predetermined hidden layer and the output layer of the first model, but this is not limited to this. The first model may have a configuration in which multiple hidden layers are provided between the predetermined hidden layer and the output layer.

[0025] Figure 5 shows another example of generating the first model. As shown in this figure, the first model may be configured so that information is propagated directly from other hidden layers in addition to the predetermined hidden layer. In this case, the parameters from the input layer to the predetermined hidden layer of the first model may be the same as the parameters from the input layer to the predetermined hidden layer of the second model.

[0026] Furthermore, information does not have to be directly propagated from the intermediate layer, for example, information may be propagated directly from the input layer to the output layer.

[0027] FIG. 6 is a diagram showing another example of first model generation. As shown in this figure, the first model may have one or more intermediate layers between a predetermined intermediate layer and an output layer. Information may be propagated directly from an intermediate layer other than the predetermined intermediate layer to an intermediate layer subsequent to the predetermined intermediate layer. Similarly to the above, information may be propagated directly from the input layer to an intermediate layer subsequent to the predetermined intermediate layer, rather than from an intermediate layer.

[0028] When there are multiple intermediate layers between a predetermined intermediate layer and the output layer, information may be directly propagated from an intermediate layer before the predetermined intermediate layer to at least one intermediate layer after the predetermined intermediate layer. Alternatively, information may be directly propagated from the predetermined intermediate layer to multiple intermediate layers after the predetermined intermediate layer. Furthermore, in a configuration where there are multiple intermediate layers after the predetermined intermediate layer, information may be directly propagated from an intermediate layer before the predetermined intermediate layer to the output layer, as shown in FIG. 5.

[0029] As described above, in the first model, the number of intermediate layers between the input layer and the predetermined intermediate layer (the intermediate layers before the predetermined intermediate layer) and the number of intermediate layers between the predetermined intermediate layer and the output layer (the intermediate layers after the predetermined intermediate layer) are arbitrary. Therefore, the propagation of information from intermediate layer to intermediate layer in FIG. 6 may be the propagation of information from any one or more intermediate layers before the predetermined intermediate layer to any one or more intermediate layers after the predetermined intermediate layer.

[0030] FIG. 7 is a diagram showing another example of generating the first model. As shown in this figure, the predetermined intermediate layer does not have to be the layer immediately preceding the output layer in the second model. Furthermore, in this case, the configuration of the first model may be the configuration shown in FIGS. 4 to 6. That is, the first model may have one or more intermediate layers between the predetermined intermediate layer and the output layer, or may be configured so that information can be propagated directly from an intermediate layer before the predetermined intermediate layer to an intermediate layer after the predetermined intermediate layer or to the output layer.

[0031] FIG. 8 is a diagram showing another example of generating the first model. The first model may be formed by using the output layer of the second model as a predetermined intermediate layer. In this case, the predetermined intermediate layer outputs an energy value, just like the second model. Training may be performed between the predetermined intermediate layer and the output layer so that this output is converted into desirable information.

[0032] Furthermore, as shown by the dotted line, information may be propagated from a hidden layer before a predetermined hidden layer of the first model to the output layer. Furthermore, without being limited to these examples, a hidden layer may be arranged after the predetermined hidden layer as shown in Figure 4, or information may be propagated directly from a hidden layer before the predetermined hidden layer to a hidden layer after the predetermined hidden layer as shown in Figure 6.

[0033] 3 to 8 show some examples, but the present invention is not limited to these. For example, a layer that performs dimensionality reduction or dimensionality expansion may be present between the input layer and a predetermined intermediate layer.

[0034] When the input layer and the output layer have the same number of nodes as the atoms, it is preferable that the predetermined intermediate layer has the same number of nodes as the input layer and the output layer, since data for each atom can be output from each node, but this is not limited to this. For example, the predetermined intermediate layer may be a layer that compresses or expands the nodes (in other words, dimensionality compression or dimensionality expansion).

[0035] Furthermore, various arrangements of intermediate layers are possible, but the parameters of the copied parts may be fixed, and there may be connections that allow information to propagate between the input layer, any intermediate layer, or the output layer.

[0036] Furthermore, although the parameters from the input layer to the predetermined hidden layer in the first model are the same as the parameters from the input layer to the predetermined hidden layer in the second model, this is not limiting. In other words, the first model may be a model that has been fine-tuned to produce a different output using parameters obtained from a model that forms an NNP.

[0037] Furthermore, the model formed below the predetermined intermediate layer as the first model is not limited to a neural network model. For example, other machine learning models such as a random forest may be connected to the predetermined intermediate layer. Furthermore, the first model is not limited to the layers and parameters configured in an MLP, and may include layers and parameters of other types of neural network models.

[0038] Figure 9 is a diagram showing yet another example of the first model. Although bias inputs are not shown in the model configurations of Figures 3 to 8, appropriate bias inputs may be applied as shown in Figure 9. For example, the first model may perform inference by applying bias inputs to the output of a predetermined hidden layer.

[0039] Figure 10 is a diagram showing yet another example of the first model. The second model may be optimized as the first model by separately acquiring a different value (feature) for the input atomic structure, and using this separately acquired value and values ​​from a predetermined intermediate layer to perform transfer learning and fine tuning. As an example, the first model may be a model that performs inference by weighting and adding values ​​from a predetermined intermediate layer to a fingerprint acquired from input atomic information.

[0040] The feature may be a predefined feature obtained based on a predetermined algorithm, such as the fingerprint described above. As another example, another neural network may be formed for the input of the input layer to calculate the feature. In this case, the other neural network may also be trained as part of transfer learning.

[0041] Figure 11 is a diagram showing another example of Figure 10. A configuration may be adopted in which feature amounts separately acquired from information about atoms input to the input layer of the first model can be input to intermediate layers after a predetermined intermediate layer of the first model. Alternatively, a neural network model may be separately formed to acquire feature amounts from information about atoms to be input, and the output of this neural network may be input to intermediate layers after the predetermined intermediate layer.

[0042] Figure 12 is a diagram showing yet another example of the first model. For the input atomic configuration, information about multiple chemical structures is used as input information for a parallel intermediate layer. For example, the first model is a model that branches into multiple (three in Figure 12) parallel layers using parameters acquired from the second model from the input layer to a predetermined intermediate layer, and then integrates and outputs the data output from each branch.

[0043] The dashed arrows in Figure 12 represent weighting parameters that can be set arbitrarily. For example, the first model may have a configuration in which outputs are propagated between intermediate layers after a predetermined intermediate layer in a parallel network structure. The dashed arrows are not required, and the connections indicated by the dashed arrows do not have to exist in the first model.

[0044] If this dashed arrow were omitted, the first model would show that data is propagated in parallel between intermediate layers related to multiple chemical structures. The dashed arrow in Figure 12 represents the input of information obtained from one chemical structure to an intermediate layer that processes another chemical structure. In this case, the connections between the intermediate layers related to multiple chemical structures include serial connections. For example, if multiple chemical structures correspond to structural changes following a chemical reaction, flowing information between chemical structures as shown in this first model is useful for estimating the ease of the chemical reaction.

[0045] Note that even if there are connections of dashed arrows, it is not necessary for all propagations indicated by dashed arrows to be implemented simultaneously in the first model. The first model in Figure 12 does not exclude the existence of input and output of data indicated by dashed arrows in at least one parallel propagation path among those indicated by dashed arrows.

[0046] The parallel and serial connections described here are as follows: In FIG. 12, the connection of the left input layer → hidden layer → predetermined hidden layer → hidden layer → hidden layer with the solid arrows in the same structure in the middle or right column is referred to as a parallel propagation path. On the other hand, the dashed arrow from the left hidden layer to the middle hidden layer, or the dashed arrow from the middle hidden layer to the left hidden layer, is referred to as a serial connection. The same applies to the middle path and the right path. There are also cases where the left path and the right path are connected in series.

[0047] The branches from the input layer to parallel paths include, for example, a branch that outputs the information input to the input layer as is to the hidden layer, and a branch that outputs the information with one or more small changes, which may correspond to small changes in the position, structure, etc. of atoms in the graph.

[0048] The first model may use fixed parameters obtained from the second model up to a predetermined intermediate layer.

[0049] The information output from each parallel model in the first model is output from the output layer after the predetermined hidden layer in each branch via a hidden layer that integrates the outputs. The first model may further include a hidden layer for adjusting each output between the predetermined hidden layer and the hidden layer that integrates the outputs, on each path in the branch, as shown in Figure 12.

[0050] Parameters related to the intermediate layer and output layer after the integration process are tuned by transfer learning, etc., as described above with reference to Figure 11. Furthermore, as indicated by the dotted line brackets, if an additional intermediate layer is provided between the predetermined intermediate layer and the intermediate layer that integrates the outputs of the minutely changed structures, parameters related to this intermediate layer before integration may be tuned.

[0051] Examples of the multiple chemical structures and predicted information include, but are not limited to, the following:

[0052] For example, information on multiple chemical structures obtained by displacing some or all atoms of the original structure by small distances may be provided as input. This input may be used to obtain differential values ​​(e.g., Hessian matrices) with respect to the nuclear coordinates of the original structure. Furthermore, thermodynamic quantities (e.g., enthalpy at any temperature) that can be calculated from these differential values ​​can be predicted.

[0053] Furthermore, two different chemical structures may be given as input. Using this input, it is also possible to predict the χ parameter, which is a parameter indicating the anti-affinity between the two chemical structures. When predicting the χ parameter, each of the two chemical structures may be a molecule or a polymer component. In order to accurately predict the χ parameter, it is preferable that the volume of one of the two chemical structures is 0.125 to 8 times the volume of the other, and that the average volume of both is 1 nm or less. 3 It is preferable that:

[0054] In FIG. 12 and the above description, the first model has three parallel paths, but this is not limited to this and the first model may have two, or four or more parallel paths.

[0055] FIG. 13 is a diagram showing some non-limiting examples of variations in output from an intermediate layer between the input layer and the predetermined intermediate layer to an intermediate layer between the predetermined intermediate layer and the output layer.

[0056] As shown by the solid arrows on the left, multiple hidden layers may output to a single hidden layer.

[0057] As indicated by the dotted arrows, one hidden layer may output to multiple hidden layers.

[0058] As indicated by the dashed or dashed arrows, the data may be output from multiple hidden layers to multiple different hidden layers.

[0059] These are shown as examples, and as mentioned above, the connections between the hidden layers can be in any form. For example, the hidden layers between the input layer and a specific hidden layer, which are not connected in the second model, can be connected so that information can be transmitted directly, or the hidden layers between a specific hidden layer and an output layer can be connected so that information can be transmitted via a more complex network configuration.

[0060] 14 is a diagram showing yet another example of the first model. This first model separately obtains a different value (feature) that cannot be obtained from the input atomic configuration and inputs it into a parallel input layer separate from the input layer that inputs the atomic configuration. The network following the input layer corresponding to the feature may include one or more intermediate layers separate from the second model.

[0061] As in the above-described embodiments, the first model may have fixed parameters of the second model from the input layer corresponding to the atomic configuration to a predetermined intermediate layer.

[0062] The first model may propagate information output from an intermediate layer subsequent to the input layer corresponding to the feature amounts provided in parallel to an arbitrary layer subsequent to a predetermined intermediate layer from the input layer corresponding to the atomic configuration. After integrating the information obtained from the atomic configuration and the information obtained from the feature amounts, the first model can output information other than energy and force from the output layer.

[0063] Examples of feature quantities other than atomic structure include, but are not limited to, information such as temperature, pressure, time, and fraction. When data containing these non-limiting feature quantities is input, training data is prepared for the information desired to be acquired, and parameters related to layers in the areas indicated as transfer learning and learning in Figure 14 are optimized. As a result, a trained model can be created that can infer the desired physical property values ​​for various environmental information that can be quantified.

[0064] By allowing features other than atomic structure to be input to the input layer, it becomes possible to form a first model that predicts information other than energy and force, such as, but not limited to, viscosity and reaction rate constants.

[0065] FIG. 15 is a flowchart showing a process of training a first model in an information processing device according to an embodiment.

[0066] The processing circuit of the information processing device first acquires parameters of the second model (S100). These parameters may be acquired from a pre-trained model or may be trained by the same information processing device. The processing circuit acquires information about the second model, particularly about the layers and interlayers used to construct the first model.

[0067] Next, the processing circuitry forms a first model based on the parameters obtained from the second model (S102). The processing circuitry copies information such as parameters to the first model in common with the second model, and appropriately arranges additional layers to form the configuration of the first model.

[0068] Next, the processing circuit trains the first model (S104). The processing circuit trains the first model, for example, by transfer learning. The processing circuit trains the first model using, for example, atomic data constituting a substance and desired information, such as characteristic values ​​of the atomic data, as training data.

[0069] After the training is completed properly, the parameters etc. are output and the process ends.

[0070] FIG. 16 is a flowchart showing an inference process using the first model in an information processing device according to an embodiment.

[0071] The processing circuit of the information processing device first acquires atomic information of the substance whose value is to be acquired (S200). This atomic data may be graph information.

[0072] The processing circuit inputs the acquired atomic data into the first model (S202). The processing circuit infers and acquires the desired data by forward propagating the data input from the input layer (S204). In this way, the desired quantity can be inferred using the first model.

[0073] As described above, according to this embodiment, by performing transfer learning using a model that forms an NNP, it becomes possible to obtain diverse and highly accurate other information about atoms and substances.

[0074] The intermediate layer of the model that forms the NNP outputs a multidimensional quantity (e.g., 100 values ​​per atom) with a value for each atom. These quantities are expected to contain information representing the state of each atom based on its surrounding environment (e.g., bonding state, oxidation number, etc.) through the function of the neural network.

[0075] Furthermore, NNP can use data based on physical simulations as training data, making it easy to generate models with excellent generalization performance. Therefore, by using such models to infer other information, highly accurate results can be expected. Furthermore, by using a predetermined intermediate layer with the same number of nodes as the input and output layers, it is possible to obtain features for each atom or bond that constitutes a substance. As a result, the features for each atom can be appropriately reused to obtain other values.

[0076] The energy that can be obtained from the model that forms the NNP has a clear physical definition. This allows for highly accurate calculations, such as theoretical values. Generally, when dealing with atoms, molecules, etc., it is necessary to define quantities such as electric charge, but such quantities are difficult to clearly define. Energy is also extensive and can be superimposed. Therefore, in intermediate layers close to the output layer of the model that forms the NNP, such as the intermediate layer immediately before the second model, each node is expected to contain appropriate information about each atom. Therefore, according to the model disclosed herein, it is expected that the output from such intermediate layers can be used to appropriately obtain data about various objects or atoms. Note that the information contained in the intermediate layer may include information about each atom as well as information about bonds or substances not associated with specific atoms.

[0077] The output of the first model may be, for example, various physical properties of molecules and the environment, optical properties, mechanical properties, or effects on living organisms. As a representative example, the first model may be configured to output either the HOMO (Highest Occupied Molecular Orbital) energy, the LUMO (Lowest Unoccupied Molecular Orbital) energy, the χ parameter, or a fingerprint. As a result, the solubility and pH of a substance can be inferred. As another example, the first model may be configured to perform clustering and visualization. As a result, it can be used as an indicator of whether a molecule belongs to a crystal, whether it resembles a crystal, or the like. Furthermore, information about a substance may be output from layers other than the output layer of the first model.

[0078] The χ parameter is a dimensionless expression of the energy when two atomic groups come into contact, and while methods for calculating it based on the Monte Carlo method or molecular dynamics method are known, these calculations are expensive. By using the first model formed by this disclosure, it is expected that calculation costs can be reduced.

[0079] In the above-described embodiment, the output layer of the model (second model) that forms the NNP may be configured to output at least one of the energy of the system, the energy of the atoms, or the force acting on the atoms.

[0080] The trained model in the above-described embodiment may be a concept that includes, for example, a model that has been trained as described and then further distilled using a general method.

[0081] In addition, a model generation method for training and generating a first model using the above-mentioned information processing device is naturally included in the scope of the present disclosure.

[0082] To summarize the above explanation, in this disclosure, the expression that the first model comprises layers from the input layer to a predetermined intermediate layer in the second model and other layers includes at least one of the following two concepts.

[0083] [1] The first model is a model that is composed of (1) the layers in the second model from the input layer to the specified intermediate layer (specified layer), and (2) other layers, and is then trained by transfer learning with the value of (1) fixed.

[0084] [2] The first model is a model that is composed of (1) layers from the input layer to a predetermined hidden layer (predetermined layer) in the second model, and (2) other layers, and then trained by fine-tuning, which updates the values ​​of (1) and (2) through learning. This includes cases where at least some of the values ​​of (1) are updated. For example, it includes cases where the values ​​from the input layer to a certain hidden layer in the second model are fixed, and other parameters in the second model are updated.

[0085] Some or all of the devices (information processing devices) in the above-described embodiments may be configured as hardware, or may be configured as software (programs) executing information processing by a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), etc. When configured as software information processing, software that realizes at least some of the functions of each device in the above-described embodiments may be stored on a non-transitory storage medium (non-transitory computer-readable medium) such as a CD-ROM (Compact Disc-Read Only Memory) or a USB (Universal Serial Bus) memory, and the software information processing may be executed by loading the software into a computer. The software may also be downloaded via a communications network. Furthermore, all or part of the software processing may be implemented in a circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array), thereby allowing the software information processing to be executed by hardware.

[0086] The storage medium that stores the software may be a removable medium such as an optical disk, or a fixed medium such as a hard disk or memory. The storage medium may be located inside the computer (such as a main memory or auxiliary memory), or may be located outside the computer.

[0087] 17 is a block diagram showing an example of the hardware configuration of each device (information processing device) in the above-described embodiment. Each device may be realized as a computer 7 including, for example, a processor 71, a main storage device 72 (memory), an auxiliary storage device 73 (memory), a network interface 74, and a device interface 75, all of which are connected via a bus 76.

[0088] Although the computer 7 in FIG. 17 includes one of each component, it may also include multiple of the same component. Also, while FIG. 17 shows one computer 7, the software may be installed on multiple computers, and each of the multiple computers may execute the same or different parts of the software. In this case, a distributed computing configuration may be used in which each computer communicates via a network interface 74 or the like to execute the processing. In other words, each device (information processing device) in the above-described embodiment may be configured as a system in which one or more computers execute instructions stored in one or more storage devices to realize its functions. Furthermore, it may also be configured such that information transmitted from a terminal is processed by one or more computers provided on a cloud, and the processing results are transmitted to the terminal.

[0089] The various calculations of each device (information processing device) in the above-described embodiments may be executed in parallel using one or more processors, or using multiple computers via a network. Furthermore, the various calculations may be distributed to multiple processor cores within a processor and executed in parallel. Furthermore, some or all of the processes, means, etc. disclosed herein may be implemented by at least one processor and storage device provided on a cloud that can communicate with a computer 7 via a network. Thus, each device in the above-described embodiments may be implemented in the form of parallel computing using one or more computers.

[0090] The processor 71 may be an electronic circuit (processing circuit, processing circuitry, CPU, GPU, FPGA, ASIC, etc.) that at least controls or performs calculations on a computer. The processor 71 may also be a general-purpose processor, a dedicated processing circuit designed to perform a specific calculation, or a semiconductor device that includes both a general-purpose processor and a dedicated processing circuit. The processor 71 may also include an optical circuit, or may include a calculation function based on quantum computing.

[0091] The processor 71 may perform arithmetic processing based on data or software input from each device or the like configured internally by the computer 7, and may output arithmetic results or control signals to each device or the like. The processor 71 may control each component constituting the computer 7 by executing the OS (Operating System) of the computer 7, applications, etc.

[0092] Each device (information processing device) in the above-described embodiments may be realized by one or more processors 71. Here, the processor 71 may refer to one or more electronic circuits arranged on one chip, or may refer to one or more electronic circuits arranged on two or more chips or two or more devices. When multiple electronic circuits are used, the electronic circuits may communicate with each other via wire or wirelessly.

[0093] The main memory device 72 may store instructions executed by the processor 71, various data, etc., and information stored in the main memory device 72 may be read by the processor 71. The auxiliary memory device 73 is a memory device other than the main memory device 72. Note that these memory devices refer to any electronic component capable of storing electronic information, and may be semiconductor memory. The semiconductor memory may be either volatile memory or non-volatile memory. The memory device for saving various data, etc. in each device (information processing device) in the above-described embodiments may be realized by the main memory device 72 or the auxiliary memory device 73, or may be realized by an internal memory built into the processor 71. For example, the memory unit in the above-described embodiments may be realized by the main memory device 72 or the auxiliary memory device 73.

[0094] When each device (information processing device) in the above-described embodiments is configured with at least one storage device (memory) and at least one processor connected (coupled) to this at least one storage device, at least one processor may be connected to one storage device. At least one storage device may be connected to one processor. A configuration in which at least one processor among multiple processors is connected to at least one storage device among multiple storage devices may also be included. This configuration may also be realized by storage devices and processors included in multiple computers. Furthermore, a configuration in which a storage device is integrated with a processor (for example, a cache memory including an L1 cache and an L2 cache) may also be included.

[0095] The network interface 74 is an interface for connecting to the communication network 8 wirelessly or via a wire. The network interface 74 may be an appropriate interface, such as one that conforms to an existing communication standard. Information may be exchanged with an external device 9A connected via the communication network 8 via the network interface 74. The communication network 8 may be any one of a WAN (Wide Area Network), a LAN (Local Area Network), a PAN (Personal Area Network), etc., or a combination thereof, as long as information is exchanged between the computer 7 and the external device 9A. An example of a WAN is the Internet, an example of a LAN is IEEE 802.11 or Ethernet (registered trademark), and an example of a PAN is Bluetooth (registered trademark) or NFC (Near Field Communication), etc.

[0096] The device interface 75 is an interface such as a USB that directly connects to the external device 9B.

[0097] The external device 9A is a device connected to the computer 7 via a network. The external device 9B is a device connected directly to the computer 7.

[0098] For example, the external device 9A or the external device 9B may be an input device. The input device may be, for example, a camera, a microphone, a motion capture device, various sensors, a keyboard, a mouse, a touch panel, or other device, and provides acquired information to the computer 7. Alternatively, the external device 9A or the external device 9B may be a device equipped with an input unit, a memory, and a processor, such as a personal computer, a tablet terminal, or a smartphone.

[0099] Furthermore, the external device 9A or the external device 9B may be, for example, an output device. The output device may be, for example, a display device such as an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) panel, or a speaker that outputs sound or the like. Alternatively, the output device may be a device including an output unit, a memory, and a processor, such as a personal computer, a tablet terminal, or a smartphone.

[0100] Furthermore, the external device 9A or the external device 9B may be a storage device (memory). For example, the external device 9A may be a network storage or the like, and the external device 9B may be a storage such as an HDD.

[0101] Furthermore, the external device 9A or the external device 9B may be a device having some of the functions of the components of each device (information processing device) in the above-described embodiments. That is, the computer 7 may transmit some or all of the processing results to the external device 9A or the external device 9B, or may receive some or all of the processing results from the external device 9A or the external device 9B.

[0102] In this specification (including the claims), when the expression "at least one of a, b, and c" or "at least one of a, b, or c" (including similar expressions) is used, it includes any of a, b, c, a-b, a-c, b-c, or a-b-c. It may also include multiple instances of any element, such as a-a, a-b-b, a-a-b-b-c-c, etc. It also includes the addition of elements other than the listed elements (a, b, and c), such as having d, as in a-b-c-d.

[0103] In this specification (including claims), when expressions such as "using / using data as input / based on / according to / in response to" (including similar expressions) are used, unless otherwise specified, this includes cases where the data itself is used, or where data that has been processed in some way (e.g., data with noise added, normalized data, features extracted from data, intermediate representations of data, etc.) is used. Furthermore, when a statement is made that a result is obtained "using data as input / based on / according to / in response to" (including similar expressions), this includes cases where the result is obtained based solely on the data in question, or where the result is influenced by other data, factors, conditions, and / or states other than the data in question. Furthermore, when a statement is made that "data is output" (including similar expressions), this includes cases where the data itself is used as output, or where data that has been processed in some way (e.g., data with noise added, normalized data, features extracted from data, intermediate representations of data, etc.) is used as output, unless otherwise specified.

[0104] When the terms "connected" and "coupled" are used in this specification (including the claims), they are intended as open-ended terms that encompass any of direct connection / coupling, indirect connection / coupling, electrically connection / coupling, communicatively connection / coupling, functionally connection / coupling, and physically connection / coupling. These terms should be interpreted appropriately according to the context in which they are used, but any connection / coupling form that is not intentionally or naturally excluded should be interpreted as being included in these terms without limitation.

[0105] In this specification (including the claims), the expression "A configured to B" may include the physical structure of element A having a configuration capable of performing operation B, and the permanent or temporary setting / configuration of element A being configured / set to actually perform operation B. For example, if element A is a general-purpose processor, it is sufficient that the processor has a hardware configuration capable of performing operation B, and is configured to actually perform operation B by setting a permanent or temporary program (instruction). Also, if element A is a dedicated processor or dedicated arithmetic circuit, it is sufficient that the circuit structure of the processor is implemented to actually perform operation B, regardless of whether control instructions and data are actually attached to it.

[0106] When used in this specification (including the claims), terms implying containing or possessing (e.g., "comprising" or "including" and "having"), they are intended to be open-ended terms that include containing or possessing things other than the object designated by the object of the term. When the object of such terms implies no quantity or a singular number (e.g., expressions using the articles "a" or "an"), the expression should be construed as not being limited to a specific number.

[0107] In this specification (including the claims), although expressions such as "one or more" or "at least one" are used in some places and expressions that do not specify a quantity or that imply a singular number (expressions using the articles a or an) are used in other places, the latter expressions are not intended to mean "one." In general, expressions that do not specify a quantity or that imply a singular number (expressions using the articles a or an) should be interpreted as not necessarily being limited to a specific number.

[0108] In this specification, when a particular advantage / result is described as being obtained from a particular configuration of an embodiment, it should be understood that the same advantage / result can also be obtained from one or more other embodiments having the same configuration, unless otherwise stated. However, it should be understood that the presence or absence of the effect generally depends on various factors, conditions, and / or circumstances, and that the effect is not necessarily obtained by the configuration. The effect is merely obtained by the configuration described in the embodiment when various factors, conditions, and / or circumstances are satisfied, and the effect does not necessarily occur in a claimed invention that defines the same or a similar configuration.

[0109] When terms such as "maximize" and "maximization" are used in this specification (including the claims), they include finding a global maximum, finding an approximation of a global maximum, finding a local maximum, and finding an approximation of a local maximum, and should be interpreted accordingly according to the context in which the term is used. They also include finding approximations of these maxima probabilistically or heuristically. Similarly, when terms such as "minimize" and "minimization" are used, they include finding a global minimum, finding an approximation of a global minimum, finding a local minimum, and finding an approximation of a local minimum, and should be interpreted accordingly according to the context in which the term is used. They also include finding approximations of these minima probabilistically or heuristically. Similarly, when terms such as "optimize" and "optimization" are used, they include finding a global optimum, finding an approximation of a global optimum, finding a local optimum, and finding an approximation of a local optimum, and should be interpreted accordingly according to the context in which the term is used. It also includes finding approximations of these optimum values ​​probabilistically or heuristically.

[0110] In this specification (including claims), when multiple pieces of hardware perform a predetermined process, the pieces of hardware may cooperate to perform the predetermined process, or some of the hardware may perform all of the predetermined process. Furthermore, some of the hardware may perform part of the predetermined process, and other hardware may perform the rest of the predetermined process. In this specification (including claims), when expressions such as "one or more pieces of hardware perform a first process, and the one or more pieces of hardware perform a second process" (including similar expressions) are used, the hardware performing the first process and the hardware performing the second process may be the same or different. In other words, it is sufficient that the hardware performing the first process and the hardware performing the second process are included in the one or more pieces of hardware. Note that hardware may also include an electronic circuit or a device including an electronic circuit.

[0111] In this specification (including the claims), when multiple storage devices (memories) store data, each of the multiple storage devices may store only a portion of the data, or may store the entire data. Also, a configuration in which only some of the multiple storage devices store data may be included.

[0112] Although the embodiments of the present disclosure have been described in detail above, the present disclosure is not limited to the individual embodiments described above. Various additions, modifications, substitutions, and partial deletions are possible within the scope of the conceptual idea and spirit of the present disclosure, which is derived from the content defined in the claims and their equivalents. For example, when numerical values ​​or formulas are used in the above-described embodiments, they are shown for illustrative purposes and do not limit the scope of the present disclosure. Furthermore, the order of each operation shown in the embodiments is also illustrative and does not limit the scope of the present disclosure.

Claims

1. one or more memories; one or more processors; the one or more processors: inputting information about atoms of a substance into a first model, thereby obtaining information about the substance from the first model; The first model is Layers from an input layer to a predetermined layer in a second model that is a model forming an NNP (Neural Network Potential) that receives information about atoms as input and outputs at least one value of energy or force; a machine learning model located after the predetermined layer, The first model is a model trained to output information about the substance. Information processing device.

2. The first model is a model trained by transfer learning using from the input layer to the predetermined layer of the second model. The information processing device according to claim 1.

3. the first model is a model fine-tuned using the input layer to the predetermined layer of the second model; The information processing device according to claim 1.

4. the predetermined layer is an intermediate layer of the second model; The information processing device according to claim 1.

5. the predetermined layer is a layer immediately before the output layer of the second model or the output layer of the second model, The information processing device according to claim 1.

6. The machine learning model receives as input an output from the predetermined layer and an output from one or more layers from the input layer to the predetermined layer that are different from the predetermined layer. The information processing device according to claim 1.

7. The machine learning model receives the output from the predetermined layer and the feature amount as input. The information processing device according to claim 1.

8. The feature is a feature obtained using another neural network model. The information processing device according to claim 7.

9. the one or more processors: acquiring the feature quantity by inputting information different from the input to the input layer to the other neural network model; The information processing device according to claim 8.

10. the one or more processors: inputting information on the atomic structure of the substance into the input layer; inputting at least one of temperature, pressure, time, or fraction into the other neural network model; The information processing device according to claim 9.

11. The first model includes parallel propagation paths for information about a plurality of chemical structures, Equipped with inputting information for each of the parallel propagation paths from the input layer; The information processing device according to claim 1.

12. The parallel propagation paths are connected to at least one intermediate layer. The information processing device according to claim 11.

13. the one or more processors: acquiring information about the substance by inputting at least a first output from the predetermined layer for a first chemical structure and a second output from the predetermined layer for a second chemical structure different from the first chemical structure into the machine learning model; The information processing device according to claim 1.

14. The second chemical structure is a chemical structure in which the positions of at least some of the atoms contained in the first chemical structure are changed by a small distance. The information processing device according to claim 13.

15. The machine learning model includes a neural network model.

15. An information processing device according to claim 1.

16. The machine learning model comprises one or more intermediate layers and an output layer.

15. An information processing device according to claim 1.

17. The machine learning model comprises a random forest.

15. An information processing device according to claim 1.

18. the information about the substance is a physical property value of the substance; 15. An information processing device according to claim 1.

19. The physical property values ​​of the substance include at least one of HOMO (Highest Occupied Molecular Orbital) energy, LUMO (Lowest Unoccupied Molecular Orbital) energy, χ parameter, or fingerprint; 19. The information processing device according to claim 18.

20. The information about the substance is information used for clustering or visualization of the substance.

15. An information processing device according to claim 1.

21. The information about the atoms of the substance includes information about the chemical structure of the substance.

15. An information processing device according to claim 1.

22. Acquiring information about a substance using an information processing device according to any one of claims 1 to 14. Information processing methods.

23. 23. The information processing method according to claim 22, wherein one or more information processing devices are configured to execute the information processing method. program.

24. one or more memories; one or more processors; the one or more processors: training a first model to input information about atoms of a substance and output information about the substance; The first model includes layers from an input layer to a predetermined layer of the second model, which is a trained model, and a machine learning model located after the predetermined layer, The second model is a model that forms an NNP that outputs at least one value of energy or force when information about atoms is input. Information processing device.

25. the one or more processors train the first model by transfer learning using from the input layer to the predetermined layer of the second model; 25. The information processing device according to claim 24.

26. the one or more processors train the first model by fine tuning using from the input layer to the predetermined layer of the second model; 25. The information processing device according to claim 24.

27. The machine learning model receives an output from the predetermined layer and a feature acquired using another neural network model, the one or more processors train the other neural network model; 25. The information processing device according to claim 24.

28. generating the first model using an information processing device according to any one of claims 24 to 27; Model generation method.

29. The model generation method according to claim 28, for execution by one or more information processing devices, program.