JP7828859B2 - Prediction device and prediction method - Google Patents

Description

This disclosure relates to a predictive apparatus and method for predicting the physical properties of compounds.

In materials development, there is a need to accurately predict the physical properties of materials before development begins. For example, Patent Document 1 discloses a technique for predicting material properties by obtaining the material's compositional formula and structural information. Non-Patent Document 1 discloses a technique for predicting the physical properties of inorganic compounds by calculating the contribution of each element.

International Patent Publication 2020/031671

Anthony Yu-Tung Wang, npj Computational Materials (2021),77.

One aspect of this disclosure is the realization of a prediction device and method for accurately predicting specific physical properties of inorganic compounds.

A prediction device according to one aspect of this disclosure comprises an input acquisition unit that acquires input information of a first inorganic compound whose specific physical properties are unknown, and a prediction unit that predicts the physical properties of the first inorganic compound from the input information using a trained model. The trained model has been trained to learn the relationship between the physical properties and reference information of a second inorganic compound whose physical properties are known. The input information and the reference information each include elemental information indicating the elements of the inorganic compound, compositional information indicating the ratio of the elements, and first structural information indicating the unit cell, translation, and symmetry of the crystal structure of the inorganic compound.

A prediction method according to one aspect of this disclosure includes an input acquisition step of acquiring input information of a first inorganic compound whose specific physical properties are unknown, and a prediction step of predicting the physical properties of the first inorganic compound from the input information using a trained model, wherein the trained model has been trained to learn the relationship between the physical properties and reference information of a second inorganic compound whose physical properties are known, and the input information and the reference information each include elemental information indicating the elements of the inorganic compound, compositional information indicating the ratio of the elements, and first structural information indicating the unit cell, translation, and symmetry of the crystal structure of the inorganic compound.

Each aspect of the prediction device described herein may be implemented by a computer. In this case, the control program for the prediction device, which enables the computer to implement the prediction device by operating the computer as each component (software element) of the prediction device, and the computer-readable recording medium on which the program is recorded, are also included in the scope of this disclosure.

According to one aspect of this disclosure, it is possible to predict specific physical properties of inorganic compounds with high accuracy.

This is a block diagram showing an example of the configuration of a prediction device according to an embodiment of this disclosure. This figure shows an example of the data structure of the input information. This figure shows an example of the data structure of the output information. This is a functional block diagram of the prediction unit, which performs prediction processing based on a pre-trained model. This is a schematic diagram illustrating some of the processes performed using a pre-trained model. This is a schematic diagram illustrating some of the processes performed using a pre-trained model. This figure shows an example of the displayed information. This figure shows an example of the displayed information. This figure shows an example of the displayed information. This flowchart shows an example of the flow of prediction processing performed by the prediction device. This block diagram shows an example of the configuration of a prediction device related to a modified version of the present disclosure. This figure shows an example of the displayed information. This graph shows the correlation between the measured and predicted magnetic moment values related to the training data of the example. This graph shows the correlation between the measured and predicted magnetic moment values related to the verification data of the example. This graph shows the correlation between the measured and predicted magnetic moment values related to the training data of the comparative example. This graph shows the correlation between the measured and predicted magnetic moment values related to the verification data of the comparative example.

<Embodiment 1>
[Overview of the prediction device]
Traditionally, developing novel materials with target physical properties relied on the experience and intuition of researchers. The inventors of this invention have established a method for predicting specific physical properties of unknown inorganic compounds by training a prediction device on the relationship between elemental information, compositional information, first structural information indicating the unit cell, translation and symmetry of the inorganic compound's crystal structure, and physical properties.

The prediction device described herein can be used, for example, in research facilities where research and development of new materials are conducted. The prediction device uses, as input information, elemental information indicating the elements of an inorganic compound, compositional information indicating the ratio of elements, and first structural information indicating the unit cell, symmetry, and translation of the crystalline structure of the inorganic compound. This allows the prediction device to accurately predict specific physical properties of inorganic compounds.

The prediction device, by employing first structural information—which indicates the unit cell, translation, and symmetry of the crystalline structure of inorganic compounds—as input information, can accurately predict the physical properties of inorganic compounds, even those where the crystalline structure is dominant over the physical properties.

Furthermore, by adopting information indicating the unit cell, translation, and symmetry of the crystalline structure of inorganic compounds as the first structural information, users can easily input structural information without having to meticulously set the spatial arrangement of each atom of the inorganic compound. This also has the effect of making it easier for users to prepare a dataset of structural information, not only for known compounds but also when predicting the physical properties of unknown compounds.

[Configuration of prediction device 1]
Next, the configuration of the prediction device 1 will be explained using Figure 1. Figure 1 is a block diagram showing an example of the configuration of the prediction device 1.

The prediction device 1 is a computer that acquires elemental information indicating the elements of an inorganic compound, compositional information indicating the ratio of elements, and first structural information indicating the unit cell, symmetry, and translation of the crystalline structure of the inorganic compound as input information 31, and predicts specific physical properties of the inorganic compound.

As shown in Figure 1, the prediction device 1 comprises a control unit 10 that comprehensively controls each part of the prediction device 1, a storage unit 20 that stores various data used by the control unit 10, and a communication unit 30 that communicates with external devices.

The memory unit 20 stores the trained model 213, the first inorganic compound database 210, and the second inorganic compound database 211.

The trained model 213 is a model that has been trained to learn the relationship between the reference information 32 of a second inorganic compound whose physical properties are known and specific physical properties. The reference information 32 includes elemental information indicating the elements of the second inorganic compound, compositional information indicating the ratio of elements, and first structural information indicating the unit cell, translation, and symmetry of the crystal structure of the second inorganic compound. Details of the elemental information, compositional information, first structural information, and physical properties will be described later.

The first inorganic compound database 210 is a database that stores elemental information, compositional information, and first structural information of inorganic compounds whose physical properties are unknown, with each information associated with the corresponding data. Here, if the physical properties differ for each unit cell, "unknown" or "known" is determined for each unit cell. For example, "unknown physical properties" means that even if the chemical formula of the inorganic compound is known, the physical properties based on the unit cell are unknown.

The second inorganic compound database 211 is a database that stores reference information 32 for second inorganic compounds whose physical properties are known. Specifically, the reference information 32 is information that associates the elemental information, composition information, and first structural information of the second inorganic compound with its physical properties. Furthermore, the second inorganic compound database 211 may also include second and third structural information, as described later.

In this embodiment, the prediction device 1 comprises a first inorganic compound database 210 and a second inorganic compound database 211, but is not limited to this configuration. The first inorganic compound database 210 and the second inorganic compound database 211 may be stored on an external server, and the prediction device 1 may access each database stored on the server. In this case, the first inorganic compound database 210 and the second inorganic compound database 211 may be stored on the same server, or they may be stored on different servers. Furthermore, the storage unit 20 may store control programs for performing various controls on the prediction device 1.

<Control Unit 10>
The control unit 10 comprises an input acquisition unit 11, a prediction unit 13, a creation unit 14, and a learning unit 15. The following describes each component of the control unit 10.

(Learning Section 15)
The learning unit 15 generates a trained model 213. The trained model 213 is generated using a dataset of reference information 32 of a second inorganic compound whose physical properties are known, i.e., elemental information, compositional information, and first structural information, and physical properties, as training data. The training data may be stored, for example, in the storage unit 20 described above. Specifically, the training data may be stored in the second inorganic compound database 211. In this case, the training data may be obtained from the second inorganic compound database 211 via the input acquisition unit 11, which will be described later. Alternatively, the training data may be obtained from an external device via the input acquisition unit 11. Details of the elemental information, compositional information, first structural information, and physical properties will be described later. The learning unit 15 may also appropriately change the hyperparameter set of the trained model 213.

(Input acquisition unit 11)
The input acquisition unit 11 acquires input information 31 of a first inorganic compound whose specific physical properties are unknown. The input acquisition unit 11 may also acquire the input information 31 from an external device other than the prediction device 1.

Furthermore, input information 31 for one inorganic compound may be entered per input, or input information 31 for multiple inorganic compounds may be entered per input. The input information 31 may be obtained from the first inorganic compound database 210.

Figure 2 shows an example of the data structure of input information 31. Input information 31 includes elemental information 310, composition information 311, and first structure information 312. Input information 31 may also include second structure information 313 and third structure information 314.

Elemental information 310 is information indicating the types of elements contained in the inorganic compound. For example, in Figure ₂ , elemental information 310 shows the inorganic compound BaTiO3, including the elements "Ba", "Ti", and "O" contained in _BaTiO3 .

The composition information 311 indicates the ratio of elements contained in the inorganic compound. For example, in Figure ₂ , the composition information 311 of the inorganic compound BaTiO3 is "Ba:Ti:O = 1:1:3", indicating that the ratio of elements Ba, Ti, and O is 1:1:3.

The first structural information 312 is information indicating the unit cell, translation, and symmetry of the crystalline structure of the inorganic compound.

The unit cell of the crystalline structure of an inorganic compound is the smallest unit contained in the crystal of the inorganic compound, and for example, known unit cells can be used. Specifically, 14 types of Bravais lattices are adopted, considering seven types of crystal systems (cubic, tetragonal, orthorhombic, hexagonal, trigonal, monoclinic, and triclinic) and four types of lattice systems (simple (P), body-centered (I), face-centered (F), and base-centered (C)).

The translational properties and symmetry of a crystal structure are information represented by a space group originating from the center of the crystal. The translational properties of a crystal structure specifically include symmetry elements such as translation, helical movement, and reflection, while symmetry specifically includes symmetry elements without translation, such as rotation, reversal, inversion, and reflection.

The first structural information 312 results in 230 three-dimensional space groups, determined by combinations of unit cells, symmetry, and translation. Specifically, 230 types of three-dimensional spaces are determined by combinations of 14 types of Bravais lattices, 32 types of point groups, and translational operations. In other words, the first structural information 312 of an inorganic compound will be assigned one of the 230 three-dimensional space groups. The first structural information 312 may be the space group symbol or corresponding number determined by international notation. In Figure 2, as an example, the space group number "221" for the inorganic compound _BaTiO3 is entered as the first structural information 312. Hereafter, the first structural information will also be referred to as the "space group".

The input information 31 may further include second structural information 313 indicating the lengths of each axis of the unit cell. The lengths of each axis of the unit cell may be expressed, for example, in Å (angstroms) or nm (nanometers).

In Figure 2, the second structural information 313 is the information "a = b = c = 3," indicating that the lengths of each axis of the unit cell are the same and that the length of each axis is 3 Å.

Thus, by further using the second structural information 313 to predict the physical properties of inorganic compounds, the prediction device 1 can predict the physical properties with high accuracy.

Furthermore, the input information may also include third structural information 314 indicating the angles between each edge of the unit cell.

In Figure 2, the third structural information 314 is the information "α = β = γ = 90°", indicating that the angles α between axis a and axis b of the unit cell, β between axis b and axis c, and γ between axis c and axis a are all equal, and each angle is 90°.

Thus, by further using the third structural information 314 to predict the physical properties of inorganic compounds, the prediction device 1 can predict the physical properties with even greater accuracy.

(Prediction unit 13)
The prediction unit 13 uses the trained model 213 to predict specific physical properties of inorganic compounds from the input information 31.

Specific physical properties of inorganic compounds include, for example, magnetic moment, formation energy per molecule, formation energy per atom, band gap, density, stability from the thermodynamic convex hull, and volume of the unit structure.

The trained model 213 is a model that has been trained to understand the relationship between the reference information 32 (elemental information, composition information, and first structural information) of a second inorganic compound whose physical properties are known, and the physical properties themselves. Details of the trained model 213 will be described later.

The prediction unit 13 may predict a first contribution, which indicates the degree to which each element contained in the first inorganic compound influences the physical properties of the first inorganic compound, based on the input information 31.

The elements contained in inorganic compounds influence their physical properties. Furthermore, the influence of multiple elements in an inorganic compound is not the same; the degree of influence varies depending on the type of element. The first contribution indicates the degree of influence each element has on the physical properties of the inorganic compound, i.e., the magnitude of its contribution to those properties.

The first contribution may be a value calculated for each element contained in the first inorganic compound, or it may be the sum of the contributions of all elements contained in the first inorganic compound. The first contribution can be expressed numerically, for example. A larger numerical value for the first contribution indicates a greater influence that element has on the physical properties of the first inorganic compound.

Because the prediction unit 13 can predict the first contribution, the user can find out which element in the first inorganic compound has a significant influence on the physical properties of that first inorganic compound.

The prediction unit 13 may also predict a second contribution, which indicates the degree to which the structure indicated by the structural information influences the physical properties of the first inorganic compound, based on the input information 31.

The structural information of inorganic compounds influences their physical properties. For example, even if the elements in an inorganic compound are the same, different structural information can lead to different physical properties. The second contribution indicates the degree to which structural information influences the physical properties of an inorganic compound, i.e., the magnitude of its contribution to those properties.

The second contribution is represented, for example, by a numerical value. The larger the numerical value of the second contribution, the greater the influence of the structural information indicating that second contribution has on the physical properties of the inorganic compound. The second contribution and the first contribution mentioned above can be compared with each other. Therefore, for example, if the influence of the inorganic compound's structure on physical properties is greater than the influence of a certain element on physical properties, then "second contribution > first contribution" is acceptable. In this case, for example, when a user attempts to control the physical properties of this inorganic compound (for example, to increase them), they can gain guidance that they should focus on designing the inorganic compound's structure.

Figure 3 shows an example of output information 41 output by the prediction unit 13 when input information ₃₁ for the inorganic compound BaTiO3 shown in Figure 2 is input. The output information 41 includes a first contribution 411, a second contribution 412, and physical property values 410. The first contribution 411 indicates the elemental contribution of the inorganic compound _BaTiO3 . The first contribution 411, the second contribution 412, and the physical property values may each be expressed as numerical values.

The prediction unit 13 may predict the physical properties based on a first contribution, which indicates the degree of influence of each element contained in the first inorganic compound on the physical properties of the first inorganic compound, and a second contribution, which indicates the degree of influence of the structure indicated by the structural information on the physical properties of the first inorganic compound. Specifically, the prediction unit 13 may calculate the physical properties of the first inorganic compound by calculating the average of the first and second contributions of each element. According to this method, the prediction device 1 can predict the physical properties of the first inorganic compound with high accuracy.

The prediction unit 13 may calculate a third contribution from the first and second contributions of at least one element included in a predetermined period among the elements contained in the first inorganic compound. The third contribution indicates the degree of influence of at least one element included in the predetermined period on the physical properties of the first inorganic compound. The third contribution is an index representing the extent to which at least one element contained in the first inorganic compound contributes to the physical properties of the inorganic compound when the first inorganic compound takes on a predetermined structure.

The first contribution is the degree of influence each element has on the physical properties of the inorganic compound, and does not take into account the structural information of the inorganic compound. However, the degree of influence an element has on the physical properties of an inorganic compound can change depending on how that element is arranged in the structure. Therefore, by calculating the third contribution, the user can understand the degree of influence an element has on the physical properties of an inorganic compound when that element is arranged in a predetermined structure.

The predetermined structure is the structure indicated by the structural information contained in the input information 31. In this case, the structural information may consist only of the first structural information 312, or it may include the first structural information 312, the second structural information 313, and the third structural information 314.

The specified period is not particularly limited, but for example, it could be the elements of the fourth or fifth period.

[Pre-trained model]
As described above, the function of the prediction unit 13 to predict physical properties is realized when the prediction unit 13 inputs input information 31 to the learned model 213.

The pre-trained model 213 will be described below. Figure 4 shows an example of a functional block diagram of the prediction unit 13, which performs prediction processing based on the pre-trained model 213. Here, the information after predetermined processing of elemental information 310 and compositional information 311 will be simply referred to as "elemental information" and "compositional information," respectively. Similarly, the information after predetermined processing of the first structural information 312, second structural information 313, and third structural information 314 will be simply referred to as "structural information."

The trained model 213 comprises a first transformation unit 2130, a first NN unit 2131, a second transformation unit 2132, an addition unit 2133, a vector transformation unit 2134, a scaling unit 2135, a structural information integration unit 2136, a second NN unit 2137, an integration unit 2138, a relationship learning unit 2139, a third NN unit 2140, and a calculation unit 2141.

The first conversion unit 2130 converts the input elemental information 310 into a unique sequence. Specifically, the first conversion unit 2130 converts the atomic numbers of the elements constituting the inorganic compound into an embedded representation of the elemental information. As an example, the first conversion unit 2130 converts the information of the atomic number "1" of a hydrogen atom into a unique sequence such as [0.0571, -0.2255, ...]. For the elemental information 310, the first conversion unit 2130 uses, as an example, Mat2vec embedding processing.

The first NN unit 2131 converts the dimensionality of the elemental information having a unique sequence, converted in the first conversion unit 2130, to a predetermined number of dimensions through the neural network. The neural network used in the first NN unit 2131 may be, for example, a neural network without an intermediate layer between the input layer and the output layer, and the input layer and output layer may be weighted by a predetermined weight parameter W1. In other words, the weight W1 is a parameter learned during model training.

The second conversion unit 2132 converts the input composition information 311 into an array of a specific length. As an example, positional encoding is used.

The addition unit 2133 adds the elemental information and the compositional information together to generate information that incorporates the compositional information into the elemental information. For example, the second conversion unit 2132 converts the elemental information Mn (manganese) into information called " _Mn² " which includes the compositional ratio of the inorganic compound.

The vector conversion unit 2134 vectorizes the input first structural information 312 and further converts it into a fixed-length vector. Specifically, it converts the space group number, which is the first structural information 312, into a fixed-length vector. The space group number is a number assigned to a predetermined space group, and the space group number itself has no meaning. However, the numbers representing this space group number are converted into similar sequences for information representing similar space groups, and into completely different sequences for information representing dissimilar space groups.

The scaling unit 2135 performs a standardization process on the input second structure information 313 and third structure information 314. While the first structure information 312, which indicates the spatial group number, is simply a numerical value, the lengths of each axis of the unit cell (second structure information 313) and the angles between each edge of the unit cell (third structure information 314) are one-dimensional arrays. For example, the angles between each edge of the unit cell in the third structure information 314 are represented by a one-dimensional array such as [90, 90, 90]. Therefore, unlike the first structure information 312, the scaling unit 2135 performs a standardization process on the second structure information 313 and third structure information 314.

The structural information integration unit 2136 integrates the first structural information after vector transformation processing with the standardized second structural information and the standardized third structural information.

The second NN unit 2137 converts the dimensionality of the structural information integrated in the structural information integration unit 2136 to a predetermined dimensionality through the neural network. The neural network used in the second NN unit 2137 may be, for example, a neural network without an intermediate layer between the input layer and the output layer, and the input layer and output layer may be weighted by a predetermined weight parameter W2. In other words, the weight W2 is a parameter learned during model training.

Figure 5 schematically illustrates the processing by the second NN unit 2137 from the input of first structural information 312, second structural information 313, and third structural information 314. In Figure 5, as an example, the case in which information of the inorganic compound " _MnO₂ " is input is explained. For example, the numerical value of space group number "166" input as first structural information 312 is converted by the vector conversion unit 2134 into a predetermined vector (first structural information m1), and further converted into a fixed-length vector (first structural information m2). In addition, the lengths of each axis of the unit cell input as second structural information 313 [2.93, 2.93, 2.93] and the angles between each edge of the unit cell input as third structural information 314 [77.0, 77.0, 60.0] are standardized by the scaling unit 2135. The standardized second structural information and the third structural information may be combined to form information indicating a predetermined arrangement (structural information m3). The first structural information m2 after vector transformation and the standardized structural information m3 are integrated in the structural information integration unit 2136. The integrated structural information m4 is then processed in the second NN unit 2137 to become information with a predetermined number of dimensions (structural information m5).

Returning to Figure 4, in the second NN section 2137, the integrated structural information is further integrated in the integration section 2138 into information that adds up elemental information and compositional information.

Figure 6 schematically illustrates the process by which elemental information m6 converted in the first NN unit 2131, compositional information m7 converted in the second conversion unit 2132, and structural information m5 integrated in the structural information integration unit 2136 are integrated. For example, the converted elemental information m6 and the converted compositional information m7 are added in the addition unit 2133 to obtain elemental information m8. In the integration unit 2138, elemental information m8 is further integrated with structural information m5, which is formed by integrating the first to third structural information, to obtain integrated information m9. Integrated information m9 is then input to the relationship learning unit 2139, which will be described next.

Returning to Figure 4, the relationship learning unit 2139 learns the interrelationships between each element and the structural information by taking the dot product between each element and the structural information of the integrated information m9, which is a combination of elemental information, compositional information, and structural information. The processing performed by the relationship learning unit 2139 reveals which elements influence the physical properties of the inorganic compound. The relationship learning unit 2139 can use known attention mechanisms. For example, Multi-head Attention can be used.

The third NN unit 2140 converts the information processed by the relational learning unit 2139 into information of a predetermined number of dimensions through the neural network.

The calculation unit 2141 calculates the contribution of each element by applying a sigmoid function to the information output by the third NN unit 2140. Furthermore, the calculation unit 2141 calculates the physical property value by finding the average contribution of all elements and outputs the output information 41.

(Creation section 14)
The creation unit 14 creates display information 42 based on the output information 41 predicted and output by the prediction unit 13. Here, the display information 42 is not particularly limited, but may be, for example, an image, graph, numerical value, list, etc., based on the results predicted by the prediction unit 13. The display information 42 may also be a GUI (Graphical User Interface) screen for presentation to the user. The creation unit 14 may also obtain the data necessary to create a predetermined GUI from, for example, the second inorganic compound database 211. The following describes a specific example of the display information 42 created by the creation unit 14.

The creation unit 14 may create a heatmap showing the distribution of the third contribution as display information 42, using the first structural information 312 and the number of electrons that affect the physical properties of the first inorganic compound as two axes. The creation unit 14 may also use structural information that includes the second structural information 313 and the third structural information 314 as input information, in addition to the first structural information 312. The number of electrons that affect the physical properties of the first inorganic compound is not particularly limited, but for example, it is the number of outermost shell electrons of at least one element included in a predetermined period within the first inorganic compound.

Figure 7 is an example of a heatmap created by the creation unit 14. In the heatmap 141 of Figure 7, the horizontal axis shows the space group (space group number), and the vertical axis shows the number of valence electrons. There are 230 types of space groups, and the number of valence electrons is shown in the range of 0 to 12. The third contribution corresponding to each space group and each number of valence electrons is shown by a gradient of color. The third contribution is shown in the range of 0 to 200, and in the heatmap 141, areas with relatively lighter colors have a high third contribution. In the heatmap 141, it can be seen that the elements of inorganic compounds classified as space group number 160 and space group number 166 have relatively high third contributions. In this way, by creating a heatmap that shows the distribution of third contributions corresponding to space groups and the number of valence electrons, the creation unit 14 can, for example, allow the user to visually understand the magnitude of the third contributions via a display device 2. Specifically, users can visually identify the number of outermost electrons that result in a high third contribution to an element, and which space group the element belongs to that results in a high third contribution.

Furthermore, the creation unit 14 may further create a diagram as display information 42, showing the correlation between the number of electrons in the first structural information 312 that affect the physical properties of the inorganic compound and the third contribution, for regions with a high third contribution in the heatmap created by the creation unit 14. Specifically, the diagram showing the correlation between the number of electrons and the third contribution may be a graph.

Figure 8 is an example of a graph showing the correlation between the number of electrons and the third contribution. Graph 142 in Figure 8, as an example, shows the correlation between the number of electrons and the third contribution for data belonging to space group number 166, a region with a high third contribution, as seen in heatmap 141 of Figure 7. Graph 142 shows the correlation between the third contribution on the vertical axis and the number of electrons in space group number 166 on the horizontal axis. Furthermore, in addition to the correlation between the third contribution and the number of electrons, graph 142 may also show the number of data points extracted from the second inorganic compound database 211.

The number of data points represents the number of data points extracted from the second inorganic compound database 211. More specifically, the number of data points represents the number of inorganic compounds in the second inorganic compound database 211 that have a structure with space group number 166 and contain elements with 0 to 10 outermost electrons. That is, according to Graph 142, the user can find out how many compounds containing elements with a high third contribution are included in the second inorganic compound database 211.

Furthermore, the generation unit 14 may add an approximation curve based on the plot of the third contribution to the correlation diagram between the number of electrons and the third contribution. In Figure 8, the dotted curve is the approximation curve. This allows, for example, the user to easily understand the number of electrons that maximizes the third contribution from the approximation curve. The generation unit 14 may also extract compounds from the second inorganic compound database 211 with the number of electrons corresponding to the point showing a maximum value on the approximation curve, i.e., the number of electrons estimated to maximize the third contribution.

Furthermore, the creation unit 14 may further create a list of inorganic compounds as display information 42, which includes first structural information 312 indicating the third contribution and the number of electrons that affect the physical properties of the inorganic compound.

Figure 9 is an example of a list of inorganic compounds that possess the first structural information (space group) 312 indicating the third contribution, and the number of electrons (number of outermost electrons) that influences the physical properties of the inorganic compound. List 143 in Figure 9 is a list extracted and displayed from the second inorganic compound database 211, which contains data for multiple inorganic compounds, showing inorganic compounds containing an element with space group number 160 and a number of outermost electrons of 5. The ID is the identification number assigned to each extracted inorganic compound.

List 143 includes, as an example, inorganic compounds with the same ID and identical composition (No. 3 and No. 4). These indicate that while the inorganic compounds have the same composition, the elements targeted are different. Specifically, No. 3 shows the third contribution of Fe (iron), and No. 4 shows the third contribution of Co (cobalt). This allows users to visually identify inorganic compounds with a high third contribution of 413, and the elements contained within those inorganic compounds.

The display information 42 created in the creation unit 14 may, for example, be output to an external display device 2 and displayed on the display device 2, or, if the prediction device 1 is equipped with a display unit (not shown), it may be output to the display unit of the prediction device 1. This allows the user to visually identify the prediction results made by the prediction device 1 by looking at the display information 42 created by the creation unit 14.

[Flow of prediction processing in prediction device 1]
Next, the flow of the prediction process in the prediction device 1 will be explained using Figure 10. Figure 10 is a flowchart showing an example of the prediction process flow.

First, the input acquisition unit 11 acquires input information 31 of a first inorganic compound whose specific physical properties are unknown (Step S1: Input Step). The input information 31 includes elemental information indicating the elements of the inorganic compound, compositional information indicating the ratio of elements, and first structural information 312 indicating the unit cell, translation, and symmetry of the inorganic compound's crystal structure.

The prediction unit 13 uses a trained model to predict specific physical properties of the first inorganic compound from the input information 31. Specifically, the prediction unit 13 predicts a first contribution, indicating the degree of influence of each element contained in the inorganic compound on the physical properties of the first inorganic compound, from the converted input information (step S2). Then, the prediction unit 13 predicts a second contribution, indicating the degree of influence of the structure indicated by the structural information on the physical properties of the first inorganic compound (step S3). Based on the first and second contributions, the prediction unit 13 predicts the physical properties of the first inorganic compound (step S4). The trained model has been trained to learn the relationship between the elemental information, composition information, and first structural information 312 of a second inorganic compound whose physical properties are known, and the aforementioned physical properties.

Furthermore, the prediction unit 13 calculates a third contribution, which indicates the degree of influence of at least one element on the physical properties of the first inorganic compound, from the first and second contributions of at least one element included in a predetermined period among the elements contained in the first inorganic compound (Step S5: Calculation Step).

The creation unit 14 creates an image or list as display information 42 related to the third contribution calculated by the prediction unit 13 (step S6: creation step).

The creation unit 14 outputs the created display information 42 to the display device 2, which is an external device of the prediction device 1 (step S7: output step), and the process ends.

Thus, the prediction process related to the prediction device 1 includes an input acquisition step of acquiring input information 31, and a prediction step of predicting specific physical properties of the first inorganic compound from the input information using a trained model.

By employing this prediction method, the prediction device 1 can accurately predict specific physical properties of the first inorganic compound. Furthermore, by inputting first structural information 312, which indicates the unit cell, translation, and symmetry of the crystal structure of the first inorganic compound, as input information 31, the prediction device 1 can accurately predict physical properties even for the first inorganic compound where the crystal structure is dominant over the physical properties.

Furthermore, by adopting information indicating the unit cell, translation, and symmetry of the crystal structure of the first inorganic compound as the first structural information 312, users can easily input structural information without having to meticulously set the spatial arrangement of each atom of the first inorganic compound when inputting the information. This also has the effect of making it easier for users to prepare a dataset of structural information when predicting the physical properties of not only known compounds but also unknown compounds.

In the prediction process of this embodiment, the prediction unit 13 predicts the physical property values, then further calculates the third contribution, and the creation unit 14 creates the display information. However, the system is not limited to this configuration; the prediction unit 13 may predict only the physical property values and output the results to the display device 2.

Furthermore, the prediction results generated by the prediction unit 13 may be transmitted to an external device via the communication unit 30 instead of being sent to the creation unit 14. In this case, the display information 42 described above may be created in the external device.

<Different example>
Modifications of the present disclosure are described below. For the sake of convenience, components having the same function as those described in the above embodiments are denoted by the same reference numerals, and their descriptions are not repeated.

In Embodiment 1, an example was described in which the input acquisition unit 11 acquires elemental information 310 indicating the elements of the inorganic compound, compositional information 311 indicating the ratio of elements, and first structural information 312 from the first inorganic compound database 210. This modified example describes an example that includes an optimal condition prediction unit 17 for predicting the optimal first inorganic compound. The optimal condition prediction unit 17 can predict at least one optimal condition for the first inorganic compound. In this modified example, the first inorganic compound may be represented, for example, by an expression including one or more variables, and the optimal condition prediction unit 17 predicts the optimal condition for these variables. The predicted optimal condition may also be structural information. The input acquisition unit 11 acquires the predicted optimal condition as input information 31 from the optimal condition prediction unit 17. This modified example describes, as a specific example, the case where the variables are elemental information 310 and compositional information 311.

[Configuration of prediction device 1A]
Figure 11 is a block diagram showing an example configuration of the prediction device 1A according to this modified example.

As shown in Figure 11, the prediction device 1A includes a control unit 10A that comprehensively controls each part of the prediction device 1A, a storage unit 20A that stores various data used by the control unit 10A, and a communication unit 30 that communicates with external devices.

The memory unit 20A stores the trained model 213, the first inorganic compound database 210, the second inorganic compound database 211, as well as the statistical model 214 and the search space database 220.

The statistical model 214 represents the probability distribution of what values will be output when the prediction unit 13 predicts physical properties using the trained model 213. The statistical model 214 may be updated by the prediction unit 13.

The search space database 220 stores multiple combinations of the type of additive metal and the amount of metal added, which the search space determination unit 16 (described later) uses to determine the search space.

In this modified example, a configuration in which the prediction device 1A comprises a first inorganic compound database 210, a second inorganic compound database 211, a statistical model 214, and a search space database 220 is described, but the device is not limited thereto. For example, the first inorganic compound database 210, the second inorganic compound database 211, the statistical model 214, and the search space database 220 may be stored on an external server, and the prediction device 1A may access each database stored on the server. In this case, the first inorganic compound database 210, the second inorganic compound database 211, the statistical model 214, and the search space database 220 may be stored on the same server, or each database may be stored on a different server. Furthermore, the storage unit 20A may store control programs for performing various controls on the prediction device 1A.

<Control Unit 10A>
The control unit 10A comprises a search space determination unit 16, an optimal condition prediction unit 17, an input acquisition unit 11, a prediction unit 13A, a creation unit 14, and a learning unit 15. The following describes each component of the control unit 10A.

(Search space determination unit 16)
The search space determination unit 16 determines the search space for the optimal condition prediction unit 17 to predict the optimal first inorganic compound. More specifically, the search space determination unit 16 determines the search space for the optimal condition prediction unit 17 to predict the optimal conditions for the variables in the equation representing the first inorganic compound. As a specific example, the search space determination unit 16 determines the search space for the optimal condition prediction unit 17 to predict the optimal conditions for elemental information 310 and compositional information 311. That is, the search space determination unit 16 identifies candidate element types and composition ratio ranges for defining the search space for the optimal condition prediction unit 17 to predict elemental information 310 and compositional information 311. As an example, if the elemental information 310 includes "FeCo," which is known as a promising magnetic material among magnetic materials, the first inorganic compound can be expressed by the equation M_xFe_(1-x/2)Co_(1-x/2), where the type of metal (M) and the amount of metal added (x) to FeCO are variables. The search space determination unit 16 determines candidate types of metal (M) to be added to FeCO, and ranges for the amount of metal to be added (x). The search space determination unit 16 may extract candidate types of added metals and ranges for the amount of metal to be added from the search space database 220. As an example, the search space determination unit 16 determines that the search space is defined as follows: added metal M: [Sc-Au transition metal element], amount x = [0-1] (M_xFe_(1-x/2)Co_(1-x/2)). The amount x is a factor that determines the composition ratio of the composition formula of the inorganic compound.

The search space determination unit 16 may have the first structural information 312 pre-inputted into it. That is, the search space determination unit 16 may determine the search space with the first structural information 312 already considered. Furthermore, after the optimal condition prediction unit 17 (described later) predicts the optimal conditions for the elemental information 310 and compositional information 311, the first structural information 312 may be input to the input acquisition unit 11 from the first inorganic compound database 210.

(Optimal condition prediction unit 17)
The optimal condition prediction unit 17 predicts the optimal first inorganic compound. More specifically, the optimal condition prediction unit 17 predicts the optimal conditions for the variables in the formula representing the first inorganic compound from the search space defined in the search space determination unit 16. As a specific example, the optimal condition prediction unit 17 predicts the optimal conditions for elemental information 310 and compositional information 311. The search space determined in the search space determination unit 16 is input to the optimal condition prediction unit 17. Based on the statistical model 214 stored in the storage unit 20A, the optimal condition prediction unit 17 predicts the optimal conditions for elemental information 310 and compositional information 311 from the search space. Specifically, the optimal condition prediction unit 17 predicts the optimal conditions for the type M of the added metal and the amount x of the added metal.

The optimal condition prediction unit 17 uses Bayesian optimization to predict the optimal conditions for elemental information 310 and compositional information 311 based on a portion of the input elemental information 310 and a portion of the compositional information 311.

Bayesian optimization is an algorithm that uses a statistical model to represent the relationship between an input and its corresponding output, and then uses the acquisition function, an evaluation metric derived from the statistical model, to determine the optimal input conditions to consider next. In this modified example, the statistical model 214, as described above, shows the probability distribution of what values will be output when the prediction unit 13 predicts the physical properties using the trained model 213.

Thus, the optimal condition prediction unit 17 uses Bayesian optimization to predict the optimal conditions for the type M of additive metal and the amount x of the metal added. The elemental information 310 and the composition information 311 are uniquely determined.

(Input acquisition unit 11)
The input acquisition unit 11 acquires elemental information 310 and compositional information 311 as input information 31, based on the conditions predicted by the optimal condition prediction unit 17. As described above, the first structure information 312 may also be acquired from the optimal condition prediction unit 17 along with the elemental information 310 and compositional information 311, or the first structure information 312 may be acquired separately from the first inorganic compound database 210.

(Prediction unit 13A)
The prediction unit 13A, similar to the prediction unit 13 in Embodiment 1, uses the trained model 213 to predict specific physical properties of the inorganic compound from the input information 31. The prediction unit 13A may also update the statistical model 214 based on the predicted physical properties. After this, the optimal condition prediction unit 17 may use the updated statistical model 214 to determine the type M of the added metal and the amount x of the added metal, which are predicted to be the optimal conditions.

The series of processes performed by the optimal condition prediction unit 17, the input acquisition unit 11, and the prediction unit 13A may be repeated a predetermined number of times. The prediction unit 13A may determine the optimal conditions by comparing multiple physical property values predicted through these repetitions. Each time the optimization process by the optimal condition prediction unit 17, the input processing by the input acquisition unit 11, and the prediction process by the prediction unit 13A are repeated, the statistical model 214 is updated by the prediction unit 13A. The predetermined number of repetitions may be determined in advance.

(Creation section 14)
The creation unit 14 creates display information based on the output information predicted and output by the prediction unit 13A. In this modified example, the creation unit 14 may create display information each time the prediction unit 13A outputs a physical property value, or it may create display information after a predetermined number of processing steps from optimization to prediction have been performed.

Figure 12 is a graph 144 showing the correlation between the type of metal (M) added and the amount of metal added (x) created by the creation unit 14, along with the magnetic value. The magnetic value may be represented, for example, by the difference in color of each plot, or by the intensity of the color. In graph 144, for example, the darker the intensity of the plot, the lower the magnetic value, and the lighter the intensity of the plot, the higher the magnetic value. In this way, by showing the correlation between the type of metal (M) and the amount of metal added (x) along with the magnetic value, the creation unit 14 allows the user to visually recognize the predicted result of which metal to add and in what amount to obtain a highly magnetic material. Alternatively, the creation unit 14 may create display information 42 based on the output information 41 output based on the optimal conditions determined by the prediction unit 13A.

Thus, by equipping the prediction device 1A with an optimal condition prediction unit 17, the input acquisition unit 11 receives optimized elemental information 310 and compositional information 311. The prediction unit 13A then predicts physical properties based on the optimized elemental information 310, optimized compositional information 311, and first structural information 312. This reduces the number of prediction processes required for the prediction unit 13A to output optimal physical properties compared to when unoptimized elemental information 310 and compositional information 311 are input. In other words, the prediction time can be shortened.

In this modified example, an example in which the prediction device 1A includes an optimal condition prediction unit 17 has been described. However, the optimal condition prediction device that predicts the optimal conditions for elemental information 310 and compositional information 311 may be located outside the prediction device 1 of Embodiment 1. In this case, the prediction device 1 may acquire the elemental information 310 and compositional information 311 from the optimal condition prediction device. Furthermore, the storage unit of the optimal condition prediction device may include a statistical model and a search space database, and the statistical model may be updated by acquiring the results predicted by the prediction device 1.

[Examples of implementation using software]
The function of the prediction device 1 (hereinafter referred to as "the device") is a program that causes the device to function as a computer, and can be realized by a program that causes the computer to function as each control block of the device (particularly each part included in the control unit 10).

In this case, the device comprises a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the program using this control device and storage device, the functions described in each of the above embodiments are realized.

The above program may be recorded on one or more computer-readable recording media, rather than on a temporary storage medium. This recording media may or may not be present in the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.

Furthermore, some or all of the functions of each of the above control blocks can also be implemented by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of this disclosure. In addition, it is also possible to implement the functions of each of the above control blocks using, for example, a quantum computer.

〔summary〕
The prediction device 1 according to Embodiment 1 of the present disclosure comprises an input acquisition unit 11 that acquires input information 31 of a first inorganic compound whose specific physical properties are unknown, and a prediction unit 13 that uses a trained model 213 to predict specific physical properties 410 of the first inorganic compound from the input information 31. The trained model 213 has been trained to learn the relationship between reference information 32 of a second inorganic compound whose physical properties are known and the physical properties. The input information 31 and the reference information 32 each include elemental information indicating the elements of the inorganic compound, compositional information indicating the ratio of the elements, and first structural information 312 indicating the unit cell, translation, and symmetry of the crystal structure of the inorganic compound.

In the prediction device 1 according to aspect 2 of this disclosure, the prediction unit 13 may be capable of predicting a first contribution 411, which indicates the degree of influence of each element contained in the first inorganic compound on the physical properties of the first inorganic compound, from the input information 31.

In the prediction device 1 according to aspect 3 of this disclosure, the prediction unit 13 may be capable of predicting a second contribution 412, which indicates the degree of influence of the structure indicated by the first structural information 312 on the physical properties of the first inorganic compound, based on the input information 31.

In the prediction device 1 according to aspect 4 of this disclosure, in aspect 3, the prediction unit 13 may calculate the physical property value 410 of the first inorganic compound based on the first contribution 411 and the second contribution 412.

In any of the embodiments 1 to 4, the prediction device 1 according to embodiment 5 of this disclosure may have input information 31 that includes second structural information 313 indicating the lengths of each axis of the unit cell.

In any of the embodiments 1 to 5, the prediction device 1 according to embodiment 6 of this disclosure may have a third structural information 314 in which the input information 31 indicates the angle between each edge of the unit cell.

In the prediction device 1 according to embodiment 7 of this disclosure, in embodiment 3 or 4, the prediction unit 13 may calculate a third contribution 413, which indicates the degree of influence of at least one element included in a predetermined period on the physical properties of the first inorganic compound, from the first contribution 411 and second contribution 412 of at least one element included in a predetermined period among the elements included in the first inorganic compound.

The prediction device 1A according to embodiment 8 of this disclosure may further include an optimal condition prediction unit 17 that predicts the optimal first inorganic compound in any of embodiments 1 to 7.

The prediction device 1 according to aspect 9 of this disclosure may further include a creation unit 14 that creates a heat map 141 showing the distribution of the third contribution, using the first structural information 312 and the number of electrons affecting the physical property value 410 of the first inorganic compound as two axes, as described in aspect 7.

In the prediction device 1 according to embodiment 10 of this disclosure, in embodiment 9, the creation unit 14 may further create a diagram showing the correlation between the number of electrons in the first structural information 312 and the third contribution 413 in the region where the third contribution 413 is high in the heat map 141.

In the prediction device 1 according to aspect 11 of this disclosure, in aspect 9 or 10, the creation unit 14 may further create a list 143 of inorganic compounds having first structural information 312 indicating a predetermined third contribution 413 and the number of electrons.

The prediction method according to aspect 12 of this disclosure includes an input acquisition step (S1) of acquiring input information 31 of a first inorganic compound whose specific physical properties are unknown, and a prediction step (S4) of predicting specific physical properties 410 of the first inorganic compound from the input information 31 using a trained model 213. The trained model has been trained to learn the relationship between the physical properties and reference information 32 of a second inorganic compound whose physical properties are known. The input information 31 and reference information 32 each include elemental information indicating the elements of the inorganic compound, compositional information indicating the ratio of the elements, and first structural information 312 indicating the unit cell, translation, and symmetry of the crystal structure of the inorganic compound.

An embodiment of this disclosure is described below. In this embodiment, a prediction device 1 was used, and the input information consisted of (1) elemental information, (2) compositional information, and (3) first structural information indicating the unit cell, translation, and symmetry. The property to be predicted was the magnetic moment. Furthermore, the dataset used for training, validation, and testing of the trained model 213 consisted of data containing (1) elemental information, (2) compositional information, (3) information indicating the unit cell, translation, and symmetry, and (4) magnetic moment. Figures 13 and 14 are graphs of this embodiment.

Furthermore, as a comparative example, only (1) elemental information and (2) compositional information were used as input information, while the dataset used for training, validation, and testing of the trained model consisted of (1) elemental information, (2) compositional information, and (4) magnetic moment data. Figures 15 and 16 are graphs of the examples.

The example differs from the comparative example in that it uses a unit cell and first structural information indicating translation and symmetry as input information.

In both the examples and comparative examples, the prediction device 1 used 35,000 datasets each. These 35,000 datasets were randomly divided into training data, validation data, and test data, with 70% used for training, 15% for evaluation, and 15% for testing. The training data was used to instruct the trained model 213 to learn, the validation data was used to verify the accuracy of the trained model 213 during the learning process, and the test data was used to verify the accuracy of the final trained model 213. The trained model 213 was created using the training data. The data shown as "Training data" in Figures 13 and 15 is the training data, and the data shown as "Test data" in Figures 14 and 16 is the test data. Figures 13 to 16 are graphs plotting the predicted values against the measured values of the magnetic moment in the range of 1 to 250 _μB / ^nm³ .

The solid lines in the graphs in Figures 13 to 16 represent the true regression line obtained when the predicted magnetic moment matches the calculated magnetic moment. The coefficient of determination ^R² indicates the correlation with the true regression line. The closer the coefficient of determination ^R² is to 1, the closer it is to the true regression line, meaning the prediction accuracy of the prediction device is high. The value shown as mse represents the Mean Squared Error (mse), which is calculated by squaring the difference between the predicted and measured values for each data point and dividing the sum of these values by the number of data points. The value shown as mae represents the Mean Absolute Error (mae), which is calculated by squaring the difference between the predicted and measured values for each data point and dividing the sum of these values by the number of data points. Both MSE and MAE indicate that the smaller the value, the smaller the error between the predicted value and the actual value. In other words, the smaller the values of MSE and MAE, the higher the prediction accuracy of the prediction device.

<Results>
(1) Example In the training data of the example, the coefficient of determination ^R² was 0.85, mse was 146.78, and mae was 7.54 (Figure 13). In the test data of the example, the coefficient of determination ^R² was 0.75, mse was 258.93, and mae was 9.26 (Figure 14).

(2) Comparative Example In the comparative example, the coefficient of determination ^R² was 0.61, the mse was 383.49, and the mae was 10.99 in the training data (Figure 15). In addition, in the comparative example, the coefficient of determination ^R² was 0.53, the mse was 454.49, and the mae was 12.37 (Figure 16).

From the ^R² coefficient of determination, it was found that the predicted values predicted by the prediction device 1 in the example showed a good correlation with the measured values. Furthermore, the physical properties of test data not included in the training data were predicted with the same accuracy as the training data, demonstrating that the prediction device 1 has good performance.

The inventions described in this disclosure have been explained based on the drawings and embodiments. However, the inventions described in this disclosure are not limited to the embodiments described above. That is, the inventions described in this disclosure can be modified in various ways within the scope shown in this disclosure, and embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included within the technical scope of the inventions described in this disclosure. In other words, it should be noted that it will be easy for those skilled in the art to make various modifications or alterations based on this disclosure. Furthermore, it should be noted that these modifications or alterations are included within the scope of this disclosure.

1 Prediction device 11 Input acquisition unit 13 Prediction unit 14 Creation unit 16 Search space determination unit 17 Optimal condition prediction unit 31 Input information 32 Reference information 141 Heat map 143 List 213 Trained model 310 Elemental information 311 Composition information 312 First structural information 313 Second structural information 314 Third structural information 410 Physical property values 411 First contribution 412 Second contribution

Claims (12)

An input acquisition unit that acquires input information for a first inorganic compound whose specific physical properties are unknown,
The system includes a prediction unit that uses a trained model to predict the physical properties of the first inorganic compound from the input information,
The aforementioned trained model has been trained to learn the relationship between the reference information of the second inorganic compound whose physical properties are known and the physical properties themselves.
A prediction device wherein the input information and the reference information each include elemental information indicating the elements of an inorganic compound, compositional information indicating the ratio of the elements, and first structural information indicating the unit cell, translational properties, and symmetry of the crystal structure of the inorganic compound. The prediction device according to claim 1, wherein the prediction unit can predict a first contribution, which indicates the degree to which each element contained in the first inorganic compound influences the physical properties of the first inorganic compound, based on the input information. The prediction device according to claim 2, wherein the prediction unit can predict, from the input information, a second contribution indicating the degree to which the structure indicated by the first structural information influences the physical properties of the first inorganic compound. The prediction device according to claim 3, wherein the prediction unit calculates the physical property values of the first inorganic compound based on the first contribution and the second contribution. The prediction device according to claim 1, wherein the input information includes second structural information indicating the lengths of each axis of the unit cell. The prediction device according to claim 1, wherein the input information includes third structural information indicating the angles between each edge of the unit cell. The prediction device according to claim 3, wherein the prediction unit calculates a third contribution, indicating the degree of influence of at least one element included in a predetermined period on the physical properties of the first inorganic compound, from the first and second contributions of at least one element included in a predetermined period among the elements included in the first inorganic compound. The prediction apparatus according to claim 1, further comprising an optimal condition prediction unit for predicting the optimal first inorganic compound. The prediction apparatus according to claim 7, further comprising a creation unit that creates a heat map showing the distribution of the third contribution, using the first structural information and the number of electrons affecting the physical properties of the first inorganic compound as two axes. The prediction apparatus according to claim 9, wherein the creation unit further creates a diagram showing the correlation between the number of electrons in the first structural information and the third contribution in the region of the heat map where the third contribution is high. The prediction apparatus according to claim 9, wherein the creation unit further creates a list of inorganic compounds having the first structural information indicating a predetermined third contribution and the number of electrons. An input acquisition step to obtain input information for a first inorganic compound whose specific physical properties are unknown,
The prediction step includes predicting the physical properties of the first inorganic compound from the input information using a trained model,
The aforementioned trained model has been trained to learn the relationship between the reference information of the second inorganic compound whose physical properties are known and the physical properties themselves.
A prediction method wherein the input information and the reference information each include elemental information indicating the elements of an inorganic compound, compositional information indicating the ratio of the elements, and first structural information indicating the unit cell, translational properties, and symmetry of the crystal structure of the inorganic compound.