JP7122699B2 - Material information output method, material information output device, material information output system, and program - Google Patents

Description

The present disclosure relates to technology for outputting information about materials.

In recent years, many studies have been conducted to present new material candidates to materials researchers using artificial intelligence technology. For example, Patent Literature 1 describes a technique for determining whether or not a pair of a query protein and an arbitrary compound can be synthesized.

Japanese Patent No. 5946045

However, in Patent Literature 1, there may be cases where it is determined that synthesis is not possible even though the composition formula of a pair is actually synthesizable, so further improvement is necessary.

An object of the present disclosure is to provide a technique for preventing an unknown composition formula that is actually synthesizable from being determined to be unsynthesizable.

A material information output method according to an aspect of the present disclosure is a material information output method in a material information output device that outputs information about materials,
The material information output device has a memory for pre-storing pseudo compositional formulas generated in the learning process by a pseudo compositional formula generation model generated by learning a list of known compositional formulas in association with numerical values indicating learning stages. prepared,
The processor of the material information output device,
calculating the synthesizability of the pseudo composition formula stored in the memory using the corresponding numerical values, and assigning the calculated synthesizability as a teacher label to the pseudo composition formula stored in the memory;
generating an estimation model for estimating the synthesis possibility by learning the pseudo-composition formula to which the teacher label is assigned;
calculating the synthesis possibility for an arbitrary composition formula using the learned estimation model;
Output the calculated synthesis possibility.

According to the present disclosure, it is possible to prevent an unknown composition formula that is actually synthesizable from being determined to be unsynthesizable.

1 is a block diagram showing an example of the configuration of a material information output system according to Embodiment 1; FIG. FIG. 4 is a diagram showing an example of the data structure of a list of known composition formulas; It is a figure which shows an example of the feature-value calculated according to a periodic table. FIG. 2 is a conceptual diagram for explaining a calculation model of operations performed by a neural network; FIG. 4 is a diagram explaining how a neural network learns a classification problem; FIG. 4 is a diagram explaining how a neural network learns a regression problem; FIG. 4 is a diagram explaining how a neural network learns a generation problem; 2 is a diagram showing an example of a detailed configuration of a generative model learning unit shown in FIG. 1; FIG. FIG. 4 is a diagram for explaining how a neural network that constitutes an identification unit learns; 2 is a flow chart showing an example of processing in a learning process of the material information output system shown in FIG. 1; 2 is a flow chart showing an example of processing in the process of utilizing the material information output system shown in FIG. 1; FIG. 10 is a diagram showing an example of an image for presenting the possibility of synthesis to the user; FIG. 4 is a block diagram showing an example of the configuration of a material information output system according to a modification of Embodiment 1; FIG. FIG. 10 is a block diagram showing an example of a configuration of a material information output system according to Embodiment 2 of the present disclosure; 15 is a flowchart showing an example of processing in the learning process of the material information output system shown in FIG. 14; 10 is a flowchart showing details of voting processing; It is a figure explaining the group of an element. FIG. 10 is a diagram showing an example of a data structure of a list of known composition formulas according to Embodiment 2; FIG. 11 is a block diagram showing an example of a configuration of a material information output system according to Embodiment 3 of the present disclosure; FIG. 11 is a block diagram showing an example of a configuration of a material information output system according to Embodiment 4 of the present disclosure; It is a figure which shows an example of the scatter diagram which a display part displays. FIG. 13 is a flow chart showing an example of processing in the process of utilizing the material information output system according to the fourth embodiment; FIG. It is a block diagram showing an example of hardware constitutions for realizing a material information output system. It is a figure which shows another example of the hardware configuration of the material information output system of this indication.

(Background leading up to this disclosure)
In recent years, the field of materials informatics has been in the spotlight. Materials informatics refers to Artificial Intelligence (AI) represented by machine learning for accelerating the research speed of materials research. In materials informatics, candidates for new materials are presented to materials researchers using models constructed by machine learning technology, and materials researchers create (synthesize) desired materials from the presented candidates for new materials. )do. However, since it takes an enormous amount of time to actually synthesize materials, it is inefficient to actually create a desired material from among the proposed new material candidates, as it may require rework. target. Therefore, in materials informatics, it is important to estimate the synthesis feasibility of new material candidates and present them to materials researchers together with new material candidates.

In Patent Document 1 described above, the first pair, which is a pair of a protein and a compound that are known to interact, is used as teacher data for positive examples, and the second pair, which is a combination of proteins and compounds at random, is used as teacher data for negative examples. We use machine learning techniques such as support vector machines to generate interactive learning models. Then, a technique is disclosed in which, referring to this interaction learning model, a score indicating the possibility of interaction is calculated for a plurality of pairs obtained by randomly combining query proteins and compounds.

By the way, when a pair of randomly combined proteins and compounds is created as a compositional formula, there are many cases in which synthesis is actually possible although the compositional formula is not known.

However, in Patent Document 1, the interaction learning model is learned by treating such a composition formula as negative example teacher data, which should actually be handled as positive example teacher data. Therefore, the interactive learning model after learning has a problem in that the score of an unknown composition formula that can actually be synthesized is calculated as a low value, and this composition formula is determined to be unsynthesizable.

The present disclosure provides a technique that can prevent an unknown composition formula that is actually synthesizable from being determined to be unsynthesizable.

A material information output method according to an aspect of the present disclosure is a material information output method in a material information output device that outputs information about materials,
The material information output device has a memory for pre-storing pseudo compositional formulas generated in the learning process by a pseudo compositional formula generation model generated by learning a list of known compositional formulas in association with numerical values indicating learning stages. prepared,
The processor of the material information output device,
calculating the synthesizability of the pseudo composition formula stored in the memory using the corresponding numerical values, and assigning the calculated synthesizability as a teacher label to the pseudo composition formula stored in the memory;
generating an estimation model for estimating the synthesis possibility by learning the pseudo-composition formula to which the teacher label is assigned;
calculating the synthesis possibility for an arbitrary composition formula using the learned estimation model;
Output the calculated synthesis possibility.

Generating Pseudo-Compositional Formula The possibility of synthesizing the pseudo-compositional formula generated by the model during the learning process increases as the learning stage progresses. With this configuration, a teacher label with a higher value can be assigned to a pseudo-composition formula generated later in the learning process by the generative model, and erroneous assigning of the teacher label to the pseudo-composition formula can be suppressed. An estimation model for estimating the synthesis possibility is generated by learning the pseudo-composition formula to which the teacher label is assigned in this way. Therefore, this estimation model can calculate a higher value of synthesizing possibility for an arbitrary composition formula as the actual synthesizing possibility increases. Therefore, this configuration can prevent an unknown composition formula that is actually synthesizable from being determined to be unsynthesizable.

In the above configuration, further identifying a first known composition formula of the same type as the material to be estimated from among the known composition formulas included in the known composition formula list,
voting the first known compositional formula that matches the combination for a combination of a predetermined group of elements, holding the voting result in the memory;
assigning the supervised label indicating that the synthesizing possibility is 0 to the pseudo-composition formula whose combination matches the combination for which the voting result has not yet been voted, among the pseudo-composition formulas; good too.

According to this configuration, among the pseudo-composition formulas stored in the memory, the pseudo-composition formula having a combination of element groups different from the combination of element groups possessed by the first known composition formula of the same type as the material to be estimated is given a supervised label indicating that the combinability is 0. Therefore, the estimation model can be learned by assigning a teacher label of 0 to the pseudo-composition formula having a combination of groups of elements with low relevance to the material to be estimated. As a result, an estimation model with high estimation accuracy can be generated for the material to be estimated.

In the above configuration, the known composition formula list further includes formation energies of known composition formulas,
obtaining from the memory the energy estimation model learned using the known composition formula list and the average and variance of formation energies included in the known composition formula list;
The formation energy of the pseudo composition formula is estimated using the energy estimation model, the distance of the formation energy of the pseudo composition formula to the formation energy of the known composition formula is calculated using the average and the variance, and the distance The teacher label may be corrected such that the value of the teacher label becomes smaller as .

The compositional formula becomes unstable when the formation energy is too high, and conversely becomes unstable when it is too low. According to this configuration, the formation energy is estimated using the energy estimation model for the pseudo-composition formula to which the teacher label is assigned, and the teacher label becomes larger as the distance between the estimated formation energy and the formation energy of the known composition formula increases. is corrected to be smaller. Since the estimation model is learned using the pseudo-composition formula with corrected teacher labels, the estimation model can be learned by taking into account the physical characteristic of formation energy, and the estimation accuracy of the estimation model can be improved. can be done.

In the above configuration, the memory further holds a characteristic value estimation model that outputs characteristic values for the pseudo composition formula,
get the characteristic value range entered by the user;
inputting a characteristic value within the characteristic value range into the generative model to generate the pseudo composition formula;
estimating the characteristic values of the generated pseudo composition formula using a characteristic value estimation model;
On a graph having at least two axes, a first axis indicating the synthesizability and a second axis indicating the characteristic value, the synthesizability and the estimated value of the characteristic value for the generated pseudo-composition formula are A plotted scatterplot may be shown on the display.

According to this configuration, when a characteristic value range is input by a user, a pseudo composition formula having characteristic values within the characteristic value range is generated by the generative model, and the characteristic values of the pseudo composition formula are used for characteristic value estimation. estimated by the model. Then, a scatter diagram in which the synthesizability of the pseudo-composition formula and the estimated values of the characteristic values are plotted on a graph is displayed on the display. Therefore, by inputting the characteristic value range, the user can easily grasp the relationship between the synthesis possibility and the characteristic value for a large number of pseudo-composition formulas that are presumed to have characteristic values within the characteristic value range. can be done.

In the above configuration, the list of known composition formulas may be composed of a database of ICSD, materials project, or NOMAD.

According to this configuration, an existing database such as ICSD, materials project, or NOMAD is used as the list of known composition formulas, thus saving the trouble of creating the list of known composition formulas.

In the above configuration, the memory further holds an identification model that identifies whether the pseudo compositional formula generated by the generative model is genuine or fake,
The generative model and the discriminative model may consist of neural networks trained based on adversarial learning algorithms.

According to this configuration, since the discriminative model and the generative model are composed of neural networks trained based on the adversarial learning algorithm, the generative model can generate a pseudo-composition formula with higher synthesizability. .

In the above configuration, in assigning the teacher label, the teacher label is assigned only to the pseudo composition formula generated at the beginning of the learning process and the pseudo composition formula generated at the end of the learning process, and stored in the memory. death,
The pseudo-composition formula generated at the beginning of the learning stage is given a numerical value indicating that the synthesis possibility is 0,
A numerical value indicating that the synthesis possibility is 1 may be assigned to the pseudo compositional formula generated at the end of the learning.

According to this configuration, only the pseudo-composition formulas generated at the beginning of the learning process and the pseudo-composition formulas generated at the end of the learning process are given teacher labels and stored in the memory, so that the free space of the memory can be reduced. can be secured. In addition, since only numerical values 0 and 1 are used as teacher labels, the data structure can be simplified and free memory space can be secured.

In the above configuration, the teacher label may be given a ratio of a numerical value indicating the learning stage to the total number of learning times of the generative model.

According to this configuration, the teacher label can be represented by a numerical value within the range of 0-1.

(Embodiment 1)
The material information output system according to Embodiment 1 trains a generative model constructed by a neural network based on the known composition formula included in the known composition formula list, and the pseudo composition formula output from the generative model during learning has a constant and assigns the calculated synthesizing possibility as a teacher label. Then, using pairs of teacher labels and pseudo-composition formulas, an estimation model configured by a neural network is learned. Then, an unknown compositional formula is input to the learned estimation model, and the synthesizability of the compositional formula is output.

FIG. 1 is a block diagram showing an example of the configuration of a material information output system according to Embodiment 1. FIG. The material information output system includes a material information output device 100 , a known composition formula list 101 , a generative model learning section 102 , an input section 103 and a descriptor calculation section 108 . The material information output device 100 includes a pseudo composition formula storage unit 104, a teacher label assignment unit 105, an estimated model learning unit 106, an estimated model storage unit 107, a synthesis possibility calculation unit 109, and a display unit 110. . Each block of the material information output system may be composed of software implemented by a processor such as an image processor and a microprocessor executing a predetermined program. Specifically, in FIG. 1, unless otherwise specified, rectangular blocks are composed of processors, and cylindrical blocks are composed of memories such as semiconductor memories.

The known composition formula list 101 is composed of existing databases (DB) that store known composition formulas such as ICSD, materials project, and NOMAD. A known compositional formula is a compositional formula whose existence has been confirmed by materials researchers. The known composition formula list 101 may include composition formulas whose existence has been confirmed by the materials researcher's own experiments.

FIG. 2 is a diagram showing an example of the data structure of the known composition formula list 101. As shown in FIG. The known composition formula list 101 stores a "composition formula", a "characteristic value", a "characteristic quantity", and a "descriptor" in association with each of a plurality of known composition formulas. "Compositional formula" indicates a compositional formula such as Li _0-1 TiO ₂ . A "characteristic value" is a parameter that indicates a characteristic of a known composition formula, such as voltage and battery capacity. In this example, battery-related parameters such as voltage and battery capacity are used as "characteristic values", but this is just an example, and parameters related to materials other than batteries may also be used.

A "descriptor" consists of parameters representing the crystal structure of a material denoted by (a, b, c, α, β, γ). As parameters indicating the crystal structure, for example, a group of numerical values unique to each element, such as the coordinates of each atom, atomic volume, covalent bond radius, and density, can be used. Further, as parameters indicating the crystal structure, taking “CaMnO ₃ ” as an example, “Ca:Mn:O=1:1: A value of '100.8', which is a weighted average of '3', may be taken.

The "characteristic amount" is represented by a numerical value group (vector) calculated based on the ratio of elements in the periodic table, as shown in FIG. 3, for example. FIG. 3 is a diagram showing an example of feature amounts calculated according to the periodic table. Calculation of the feature quantity 304 will be described below by taking the composition formula 301 represented by LiCo(PO ₃ ) ₄ as an example. First, this compositional formula 301 is converted into a one-dimensional vector 303 by raster scanning the elements of the periodic table 302 from the upper left to the lower right. A vector 303 is composed of a plurality of vector elements to which elements are associated in the order of the periodic table. A vector 303 is created by assigning the number of elements constituting the composition formula 301 to the corresponding vector elements. LiCo(PO ₃ ) ₄ is composed of elements Li, Co, O, and P, and the number of elements is “1,” “1,” “12,” and “4,” so vector 303 is (0 , 0, 1, . . . , 12, . Next, the vector 303 is normalized so that the sum of the vector elements becomes 1, and the feature quantity 304 is calculated.

Note that these feature amounts and descriptors are values that can be easily converted into actual composition formulas, so the feature amounts and descriptors have a one-to-one correspondence with the composition formula. Therefore, if at least one of the feature quantity and the descriptor is known, the compositional formula can be uniquely specified. Therefore, in the following description, generating a pseudo-composition formula is the same as generating at least one of a feature amount and a descriptor.

Refer back to FIG. The generative model learning unit 102 learns a generative model for generating a composition formula using the known composition formulas included in the known composition formula list 101 . Composition formulas output from the generative model include composition formulas that are not included in the known composition formula list 101 . In the present embodiment, a compositional formula that is incidentally generated during learning is called a pseudo-compositional formula, and is used to build an estimation model for estimating the possibility of synthesizing a material.

A neural network that constitutes the generative model learning unit 102 will be described below. FIG. 4 is a conceptual diagram for explaining a calculation model of operations performed by the neural network. A neural network, as is well known, is a computing device that performs computation according to a computational model imitating a biological neural network.

As shown in FIG. 4, the neural network 400 is configured by arranging a plurality of units (indicated by white circles) corresponding to neurons in an input layer 401, a hidden layer 402, and an output layer 403, respectively. be.

The hidden layer 402 is composed of two hidden layers 402a and 402b in the example shown in FIG. 4, but may be composed of a single hidden layer or three or more hidden layers. A neural network having a plurality of hidden layers is sometimes called a multi-layer neural network device.

Here, a layer close to the input layer 401 is called a lower layer, and a layer close to the output layer 403 is called an upper layer. In this case, each unit combines the calculation results received from the units arranged in the lower layer according to the weight value (for example, a weighted sum operation), and transmits the result of the combination to the unit arranged in the upper layer. It is a computational element.

The functions of the neural network 400 include configuration information representing the number of layers and the number of units arranged in each layer of the neural network 400, and weight W=[w1, w2, . . . ].

As shown in FIG. 4, the neural network 400 receives input data X=[x1, x2, . As a result, a weighted sum operation using weights W=[w1, w2, . Output data Y=[y1, y2, . . . ] is output from the output values.

Neural network learning refers to adjusting the value of the weight W of the neural network 400 so that the desired output data Y is output when the input data X is input to the neural network 400 . When learning the neural network 400, for example, a loss function representing the error between the input data X and the output data Y is defined, and the weight W is updated along the gradient that decreases the loss function by the gradient descent method.

Problems in machine learning including neural networks are divided into three problems, identification problems, regression problems, and generation problems, depending on the type of output data. First, the identification problem will be explained. FIG. 5 is a diagram explaining how the neural network learns the classification problem. The identification problem refers to the problem of obtaining the value of the weight W of the neural network so as to determine whether a given image is a "dog" or a "cat".

More specifically, in neural network 400, each output unit is associated with a different correct label for classifying input data X. FIG. Also, the load W is such that when each of a plurality of input data X is input, the output value of the output unit corresponding to the correct correct label of the input data X approaches 1, and the output value of the other output units becomes 0. adjusted to come closer.

That is, in the example shown in FIG. 5, in the neural network 400, for example, each output unit has a correct label of "1, 0" for "dog" and a correct label of "0, 1" for "cat". , one different correct label out of the correct labels indicating "dog" and "cat" is associated. In addition, the load W outputs output data [1, 0] when learning data with the correct label of "dog" is input, and outputs data [1, 0] when learning data with the correct label of "cat" is input. Adjusted to output [0,1].

In addition, when the neural network 400 is a multi-layer neural network, the weight values are individually adjusted for each hidden layer by a learning method called “layer-wise pre-training” especially before performing the supervised learning. Also good. As a result, a load W with high generalization performance is obtained in subsequent learning.

In addition to the gradient descent method described above, for example, a known algorithm such as “Adam” may be used to adjust the weight values of the neural network 400 . Also, if the neural network 400 is a multilayer neural network, each layer may be trained separately.

Next, the regression problem will be explained. FIG. 6 is a diagram explaining how a neural network learns a regression problem. The regression problem is different from the correct labels such as "dog" and "cat", and is the problem of obtaining the value of the weight W of the neural network so that the real value such as the "characteristic value" of the material is estimated. Point.

More specifically, in the neural network 400, real values y1 and y2 forming the output data Y are associated with each of the output units, and real values x1 forming the input data X are associated with each of the input units. , x2 are associated. The load W is adjusted so that when each of the plurality of input data X is input, a correct real value corresponding to the input data X is output from each output unit.

In the example of FIG. 6, in the neural network 400, the output data Y is composed of material characteristic values, such as the voltage for the real value y1 and the capacitance for the real value y2. Also, the input data X is composed of "feature amounts and descriptors" of the material such that x1 is the parameter α, x2 is the parameter β, and x3 is the feature amount. Here, "feature quantity and descriptor" are described, but at least one of these may be used. A descriptor is a parameter that indicates the crystal structure of the material described in FIG. The feature quantity is a one-dimensional vector calculated from the ratio of elements as shown in FIG. The weight W is adjusted so that the characteristic value output (estimated) from the neural network 400 approaches the correct characteristic value.

Next, the generation problem will be explained. FIG. 7 is a diagram explaining how the neural network learns the generation problem. A generation problem is a problem that generates some data that is not included in the known DB, unlike the identification problem or regression problem described above. For example, a generation problem corresponds to an inverse problem of a regression problem, such as generating multiple composition formulas with a certain characteristic value. In addition, while identification problems and regression problems require a one-to-one correspondence between inputs and outputs, a plurality of output data Y may be output for one input data X in a generation problem. This is because there can be a plurality of composition formulas that satisfy certain characteristic values. The generative model learning unit 102 shown in FIG. 1 learns a neural network that solves this generative problem.

More specifically, in the neural network 400 shown in FIG. 7, real numbers y1 and y2 forming the output data Y are associated with each of the output units, and real numbers y1 and y2 forming the input data X are associated with each of the input units. Numerical values x1, x2, . . . are associated with each other. Also, the load W is adjusted so that a correct real number corresponding to each of the plurality of input data X is output when each of the plurality of input data X is input. For example, when voltage “9.43” and capacity “1214”, which are characteristic values of the composition formula “Li _0-1 TiO ₂ ”, are input as input data X[x1, x2], the composition formula “Li _0-1 The weight W is adjusted so that the "feature quantity and descriptor" of " _TiO2 " is output.

At this time, the input data X corresponds to the output data Y in the identification problem or regression problem, and the output data Y corresponds to the input data X in the identification problem or regression problem.

That is, in the neural network 400 shown in FIG. 7, the output data Y is such that y1 is the parameter α of the composition formula “Li _0-1 TiO ₂ ”, and y2 is the composition formula “Li _0-1 TiO ₂ ”. The parameter β of the descriptor is composed of the feature amount of the material and the descriptor. The input data X is composed of material characteristic values such that x1 is, for example, the voltage of the composition formula “Li _0-1 TiO ₂ ” and x2 is, for example, the capacity of the composition formula “Li _0-1 TiO ₂ ”. there is

Then, the weight W is adjusted so that the "feature quantity and descriptor" output (estimated) from the neural network 400 approaches the correct answer.

As described above, in the generative model learning unit 102, the neural network 400 solves the generative problem of generating the output data Y from the input data X. FIG. This generation problem is derived, for example, by an adversarial learning algorithm. Adversarial learning algorithms are also called “generative adversarial networks (GANs)” and have been extensively studied in recent years.

In the following, conditional-GAN (C-GAN), which is a type of GAN suitable for material application, will be described with reference to the generation of compositional formulas. C-GAN is an algorithm that trains the neural network of the generator by using two types of neural networks that constitute the generator and the discriminator. FIG. 8 is a diagram showing an example of a detailed configuration of the generative model learning unit 102 shown in FIG. 1. As shown in FIG.

Generative model learning unit 102 includes random number generating unit 801 , generating unit 802 , and identifying unit 803 . A random number generator 801 generates a desired number of random numbers. The number of random numbers generated is the number of pseudo composition formulas generated by the generator 802 . For example, if the number of random numbers generated is 1000, 1000 pseudo composition formulas are generated.

Next, the generation unit 802 inputs pairs of each random number and each characteristic value of the known composition formula included in the known composition formula list 101 to the neural network NN1. That is, the input data X of the neural network NN1 is X=[random number, characteristic value]. When this input data X is given, the neural network NN1 calculates output data Y according to the value of the load W. FIG. This output data Y corresponds to the pseudo composition formula. Here, the pseudo compositional formula refers to a quantity that corresponds to the descriptor shown in FIG. 2, the feature quantity shown in FIG. 3, or a combination thereof, and that can be easily converted to the original compositional formula.

The identification unit 803 determines whether the “feature quantity and descriptor” of the pseudo composition formula generated by the generation unit 802 can generate values close to the “feature quantity and descriptor” of the actual composition formula. In order to make this determination, the neural network NN2 that constitutes the identification unit 803 solves both the classification problem and the regression problem, as shown in FIG.

FIG. 9 is a diagram for explaining how the neural network NN2 forming the identification unit 803 learns. In FIG. 9, neural network 400 corresponds to neural network NN2. The “characteristic quantity and descriptor” of the pseudo-composition formula generated by the generator 802 is input as input data X to the neural network 400 . In addition, “characteristic amounts and descriptors” of known composition formulas included in the known composition formula list 101 are input as input data X to the neural network 400 .

In the classification problem, the neural network 400 determines the "feature amounts and descriptors" of the known composition formulas included in the known composition formula list 101 as "genuine", and the "feature amounts and descriptors" of the pseudo-composition formulas as " Binary discrimination to discriminate as "fake" is performed. That is, the neural network 400 outputs output data Y=[1, 0] when "features and descriptors" with genuine labels are input, and "features and descriptors" with fake labels are input. If so, the value of the load W is adjusted so as to output the output data Y=[0, 1]. Here, [1, 0] represents the output data Y corresponding to the genuine label, and [0, 1] represents the output data Y corresponding to the fake label.

On the other hand, in the regression problem, the neural network 400 adjusts the value of the weight W so that the characteristic value can be estimated correctly when the "feature quantity and descriptor" of the known composition formula are input. That is, the neural network NN2 that constitutes the identification unit 803 adjusts the weight W so that both "binary discrimination + characteristic value estimation" can be performed correctly.

Refer back to FIG. Output data (Y [genuine, counterfeit], Y [characteristic value]) from the identification unit 803 is input to the generation unit 802 and used to adjust the weight W of the neural network NN1 that constitutes the generation unit 802 . The generation unit 802 adjusts the weight W of the neural network NN1 so that the identification unit 803 may mistake the pseudo composition formula for "genuine".

When the output data (Y [genuine, fake], Y [characteristic value]) are input, the generation unit 802 generates the output data Y [genuine, fake] and the output data Y [1, 0] corresponding to the genuine label. , the distance D is calculated. For example, when Y=[0.8, 0.2] is input as the output data Y [genuine, fake], the generating unit 802 generates D=sqrt((1−0.8)×(1 −0.8)+(0−0.2)×(0−0.2))=0.3. On the other hand, when Y=[0.2, 0.8] is input as the output data Y [genuine, fake], the generation unit 802 generates the distance D as D=sqrt((1−0.2)×( 1−0.2)+(0−0.8)×(0−0.8))=1.1. In this case, since the output data Y[0.8, 0.2] has a smaller distance D than the output data Y[0.2, 0.8], it is found to be closer to the real thing. Therefore, the pseudo-composition formula corresponding to the output data Y[0.8, 0.2] is higher than the pseudo-composition formula corresponding to the output data Y[0.2, 0.8]. It is preferable as a result. Therefore, the generator 802 adjusts the weight W of the neural network NN1 so that the distance D becomes smaller.

In addition, the output data Y [characteristic value] from the identification unit 803 is also input to the generation unit 802 . This output data Y [characteristic value] represents the estimation result of the characteristic value of the identification unit 803 for the pseudo composition formula generated by the generation unit 802 . Whether or not the output data Y [characteristic value] from the identifying unit 803 is close to the characteristic value (the characteristic value of the known composition formula) used to generate the corresponding pseudo composition formula is also important information. That is, the closer the estimation result of the characteristic value of the identification unit 803 for a certain pseudo composition formula to the characteristic value used to generate the pseudo composition formula, the closer the generation unit 802 is to the real (high possibility of existing) pseudo It can be said that the composition formula has been generated.

Therefore, the generation unit 802 determines that the difference (distance) between the output data Y [characteristic value] from the identification unit 803 and the characteristic value used when generating the pseudo composition formula corresponding to the output data Y [characteristic value] is The weight W of the neural network NN1 is adjusted so that it becomes smaller. Thus, the generator 802 adjusts the weight W of the neural network NN1 using the two types of distances.

As a result, the neural network NN1 that constitutes the generator 802 can generate a pseudo-composition formula having the input characteristic values.

At this time, the neural network NN2 constituting the identification unit 803 simultaneously solves two problems (tasks) of binary discrimination and characteristic value estimation. This is a technique called multitask learning. Multi-task learning solves different tasks at the same time, so that knowledge between tasks can be shared and the accuracy of each task can be improved.

Since the accuracy of the pseudo-composition formula generated by the generation unit 802 gradually improves as the number of times of learning increases, the difficulty level of the problem to be solved by the identification unit 803 gradually increases. Curriculum learning is a learning method in which a simple problem is solved by a neural network and the difficulty level of the problem is gradually increased. Therefore, in this embodiment, the identification unit 803 may be caused to carry out curriculum learning. In curriculum learning, easy problems act as constraints to difficult problems. Therefore, the identification unit 803 can generate a neural network NN2 with high discrimination accuracy and identification accuracy.

By implementing multitask learning and curriculum learning in this way, the identification unit 803 improves generalization performance for data outside the known DB. When the generalization performance of the identification unit 803 is improved, the generalization performance of the generation unit 802 can also be improved. As a result, the generation unit 802 can generate a pseudo-composition formula that causes the neural network NN2 of the identification unit 803 to erroneously determine that it is "genuine" even if a characteristic value not included in the known DB is input.

Referring back to FIG. 1, the input unit 103 will be described. The input unit 103 receives an input from the user specifying the number of random numbers generated by the random number generator 801, that is, the number of pseudo composition formulas to be generated. The input unit 103 also receives an input from the user who designates a composition formula to be evaluated for synthesis possibility. The input unit 103 is composed of input devices such as a keyboard and a mouse.

Next, the material information output device 100 will be described. The pseudo composition formula holding unit 104 holds the pseudo composition formula generated by the generative model learning unit 102 . The pseudo-composition formula refers to the output data Y of the neural network NN1 of the generating unit 802 provided in the generative model learning unit 102, and is composed of "feature amounts and descriptors." Normally, the pseudo-composition formula holding unit 104 holds the pseudo-composition formula generated by the generation unit 802 after the learning is completed. do.

In the learning process, first, initial values are set for the weights W of the neural networks NN1 and NN2 of the generation unit 802 and the identification unit 803 . After that, the neural networks NN1 and NN2 to which the initial values have been set proceed with learning by the above-described hostile learning algorithm. At this time, the generation unit 802 causes the pseudo composition formula storage unit 104 to store the pseudo composition formula generated in each learning stage by the neural network NN1 in association with the numerical value indicating the learning stage. The weights W of the neural networks NN1 and NN2 are adjusted by, for example, the gradient descent method.

For example, if the number of random numbers generated is 1000, the neural network NN1 generates 1000 pseudo-composition formulas each time learning is performed. Therefore, assuming that the total number of times of learning is T (T is an integer of 2 or more), the generation unit 802 generates 1000 pseudo-composition formulas generated in the 0th learning and 1000 generated in the 1st learning. Pseudo-composition formulas, . , T-th time) and stored in the pseudo-composition-formula storage unit 104 . In the 0th learning, 1000 pairs of random numbers and characteristic values of known composition formulas are input to the neural network NN1 in which the initial value is set as the weight W, and 1000 pairs corresponding to each pair are input. A pseudo-formula is generated.

The teacher label assigning unit 105 calculates the synthesizability of each pseudo composition formula held in the pseudo composition formula holding unit 104 by using the numerical value indicating the learning stage associated with each pseudo composition formula. The synthesis possibility is given as a teaching label of the pseudo-composition formula held by the pseudo-composition formula holding unit 104 .

Here, the teacher label assigning unit 105 assigns the ratio of the numerical value t indicating the learning stage to the total number of learning times T as the teacher label. In this case, the teacher label is a one-dimensional real number with a value between 0 and 1. For example, the teacher label L of the pseudo-composition formula obtained in the third learning is L=[3/T]. However, in this teaching label assigning method, when the same pseudo-composition formula is generated for different numerical values t, a plurality of teaching labels L are assigned to the same pseudo-composition formula.

Therefore, in this embodiment, when a plurality of teacher labels L are assigned to the same pseudo-composition formula, the teacher-label assigning unit 105 assigns the last generated teacher label to the pseudo-composition formula. A label L, that is, the maximum teacher label L is given. For example, if a certain pseudo-composition formula is generated in each of the 5th, 20th, and 80th learning stages, the teacher label L assigned to the pseudo-composition formula is L=80/T. As a result, the teacher label assigning unit 105 can associate the pseudo composition formula with the teacher label L on a one-to-one basis.

The estimation model learning unit 106 generates an estimation model for estimating the possibility of synthesizing an arbitrary composition formula by learning the pseudo-composition formula to which the teacher label is assigned by the teacher label assignment unit 105 . Here, the estimation model learning unit 106 generates an estimation model by solving a regression problem of estimating a real value of 0 to 1 that indicates how many times the pseudo composition formula is obtained by learning. . That is, in the estimation model learning unit 106, the input data X is the descriptor and feature amount of the pseudo-composition formula, and the output data Y is the real value indicating the teacher label L. FIG.

The reason why estimating the number of learning stages in which the pseudo-composition formula was obtained coincides with judging the possibility of synthesis will be described below.

The aforementioned C-GAN is trained using the known composition list 101 . The known composition formula list 101 is a list of known composition formulas whose existence has already been confirmed. That is, when the range of synthesis possibility is set to a value of 0 (low synthesis possibility) to 1 (high synthesis possibility), the synthesis possibility of the known composition formula included in the known composition formula list 101 is 1. On the other hand, the neural network NN1 that constitutes the generation unit 802 generates a pseudo-composition formula that is significantly different from the known composition formula at the initial stage of learning. However, by gradually repeating learning, the neural network NN1 generates a pseudo compositional formula similar to the known compositional formula.

From this, it can be said that the pseudo-composition formulas obtained at the initial stage of learning have a low synthesizability, but the pseudo-composition formulas obtained just before the end of the learning have a high synthesizability. Therefore, the pseudo-composition formula obtained in the initial learning, that is, the pseudo-composition formula with L=[0/T]=[0] assigned as the teacher label L represents the lowest possibility of synthesis. be. On the other hand, the pseudo-composition formula obtained in the T-th learning, that is, the pseudo-composition formula with L=[T/T]=[1] assigned as the teacher label L represents the highest possibility of synthesis. be.

The assignment of the teacher label L described above assumes that the relationship between the possibility of synthesis and the numerical value t indicating the learning stage is linear. However, this is just an example, and in the present embodiment, the relationship between the possibility of synthesis and the numerical value t may be non-linear instead of linear. For example, assuming that the synthesizing possibility Z of the pseudo-composition formula obtained in the t-th learning, the synthesizing possibility may be determined by Z=sigmoid(t), for example. where sigmoid(t) represents the sigmoid function. In this case, the teacher label assigning unit 105 may assign the synthesizing possibility Z as the teacher label L to the pseudo composition formula.

In the present embodiment, the estimation model learning unit 106 estimates the synthesis possibility as a regression problem instead of binary discrimination of positive and negative cases. In addition, the teacher label L is given based on the numerical value t indicating the learning stage of the neural network NN1. By repeating learning, the neural network NN1 that solves the generation problem generates pseudo-compositional formulas that make it indistinguishable from the real compositional formulas by the neural network NN2 that constitutes the identification unit 803 . That is, the pseudo-composition formula generated by the generation unit 802 has a stepwise increase in synthesizability.

Furthermore, in the present embodiment, when a plurality of identical pseudo-compositional formulas are generated in the learning process, the last-generated teacher label L is assigned to the pseudo-compositional formula. For example, assume that a pseudo-composition formula with a high possibility of synthesis is generated in the early stages of the learning process. Since this pseudo-composition formula has a high possibility of being synthesized, it is generated by the generation unit 802 with a high probability even in the middle to late stages of learning and just before the end of learning. In addition, as the number of random numbers generated increases, the probability of generation by the pseudo composition formula generator 802 increases. As described above, by assigning the maximum teacher label L to the pseudo-composition formula, it is possible to suppress a decrease in the learning efficiency due to misassignment of the teacher label indicating synthesis possibility in the learning process.

Furthermore, in the present embodiment, estimation model learning section 106 uses the pseudo composition formula generated by generation section 802 for learning the estimation model. The generation unit 802 generates pseudo composition formulas as many as the set number of random numbers generated. Therefore, a large amount of data necessary for learning the estimation model can be prepared. In addition, the pseudo composition formula generated by the generating unit 802 includes many pseudo composition formulas that are not included in the known composition formula list 101, that is, do not exist in the known DB. With these, the estimation model learning unit 106 can suppress over-learning of the estimation model.

Estimation model holding unit 107 holds the estimation model generated by estimation model learning unit 106 . The information held by the estimated model holding unit 107 includes configuration information indicating the number of layers of the neural network that constitutes the estimated model and the number of units arranged in each layer, and the weights used for weight sum calculation in each unit. A weight W representing a value is included.

In the utilization process, the user inputs a composition formula to be evaluated for synthesis possibility into the input unit 103 . The descriptor calculation unit 108 converts the composition formula input to the input unit 103 into feature amounts and descriptors, and inputs them to the synthesis possibility calculation unit 109 . Here, the descriptor calculation unit 108 may calculate the descriptor by inputting the composition formula to be evaluated into a predetermined formula for calculating the parameters representing the crystal structure of the composition formula. Further, the descriptor calculation unit 108 may calculate the feature amount of the composition formula to be evaluated by calculating the one-dimensional vector 303 shown in FIG.

The synthesizing possibility calculation unit 109 reads out the estimation model represented by the configuration information and the weight W held in the estimation model holding unit 107 . Then, the synthesizing possibility calculation unit 109 inputs the feature amount and the descriptor calculated by the descriptor calculation unit 108 as the input data X to the estimation model. The synthesizing possibility calculation unit 109 then calculates the weighted sum in each unit constituting the estimation model when the input data X is given to the input unit, and outputs the output data Y representing the synthesizing possibility.

The display unit 110 includes, for example, a display device such as a liquid crystal display and a processor, and generates and displays an image showing the combining possibility calculated by the combining possibility calculating unit 109 .

FIG. 10 is a flow chart showing an example of processing in the learning process of the material information output system shown in FIG.

In S<b>101 , the generative model learning unit 102 reads known composition formulas included in the known composition formula list 101 . In S102, the generative model learning unit 102 inputs a pair of the random number generated by the random number generation unit 801 and the characteristic value of the known composition formula read in S101 to the neural network NN1 of the generation unit 802 to obtain a pseudo composition formula. . For example, if the number of random numbers generated is 1000, 1000 pairs are input to the neural network NN1 to generate 1000 pseudo composition formulas.

In S103, the generating unit 802 of the generative model learning unit 102 associates the pseudo compositional formula generated by the neural network NN1 with the numerical value t indicating the learning stage and stores it in the pseudo compositional formula holding unit 104. FIG. For example, in the above example of 1000, since 1000 pseudo-composition formulas are generated for one learning, 1000 pseudo-composition formulas are stored in association with one numerical value t.

In S104, the identification unit 803 of the generative model learning unit 102 inputs the pseudo composition formula and the known composition formula to the neural network NN2. For example, the identification unit 803 may alternately input the 1000 pseudo composition formulas generated by the generation unit 802 in the t-th learning and the known composition formulas included in the known composition formula list 101 to the neural network NN2. .

After that, the processes from S105 to S107 are performed in parallel.

In S105, the identification unit 803 estimates the characteristic value of the known composition formula input to the neural network NN2, and calculates the error from the actual characteristic value. The actual characteristic value refers to the characteristic value held in the known composition formula list 101 corresponding to the known composition formula for which the characteristic value is estimated.

In S106, the identification unit 803 performs binary discrimination on each of the known composition formula and the pseudo composition formula input to the neural network NN2, and calculates the percentage of correct answers.

In S107, the identification unit 803 estimates the characteristic value of the pseudo-composition formula input to the neural network NN2, and calculates the error between the estimated characteristic value and the characteristic value input in S102 as a pair with the random number. Here, the characteristic value input in a pair with the random number refers to the characteristic value of the known composition formula input to the generator 802 when generating the pseudo composition formula in which the characteristic value is estimated.

In S108, the identification unit 803 updates the weight W of the neural network NN2 so that the error calculated in S105 becomes smaller and the correct answer rate calculated in S106 increases.

In S109, the generation unit 802 updates the weight W of the neural network NN1 so that the percentage of correct answers calculated in S106 decreases and the error calculated in S107 decreases. Here, the identification unit 803 outputs the binary discrimination result of the pseudo composition formula to the generation unit 802 , but does not output the binary discrimination result of the known composition formula to the generation unit 802 . Therefore, output data Y [genuine, counterfeit] indicating the binary determination result of the pseudo composition formula of the identification unit 803 is input to the generation unit 802 .

If the identification unit 803 determines that the pseudo composition formula generated by the generation unit 802 is genuine, that is, if the identification unit 803 makes an erroneous determination, the generation unit 802 can generate a pseudo composition formula that is closer to the real thing. It can be said that Therefore, it can be said that the smaller the distance D between the output data Y [genuine, fake] for the pseudo composition formula and the correct label [1, 0], the more realistic the pseudo composition formula the generation unit 802 can generate. Therefore, the generating unit 802 adjusts the weight W of the neural network NN1 so that the distance D becomes smaller, thereby lowering the percentage of correct answers calculated in S106.

In S110, the generative model learning unit 102 shifts the process to S111 when the number of learning times of the neural networks NN1 and NN2 reaches the total number of learning times T, and shifts the process to S102 if the total number of learning times T has not been reached. Let

In S<b>111 , the teacher label assigning unit 105 assigns a teacher label L to the pseudo composition formula held in the pseudo composition formula holding unit 104 .

In S112, the estimation model learning unit 106 learns a neural network that solves the regression problem of estimating the value of the teacher label L from the pseudo-composition formula. In S113, the estimation model learning unit 106 causes the estimation model holding unit 107 to hold the neural network learned in S112 as an estimation model. The learning process is completed by the above.

FIG. 11 is a flow chart showing an example of processing in the utilization process of the material information output system shown in FIG. In S201, the input unit 103 receives a composition formula to be evaluated for synthesis possibility. In S<b>202 , the synthesis possibility calculation unit 109 reads the estimated model from the estimated model storage unit 107 . In S203, the descriptor calculation unit 108 calculates the feature amount and descriptor of the composition formula input in S201.

In S204, the synthesis possibility calculation unit 109 inputs the feature amount and the descriptor calculated in S203 to the estimation model, and calculates the synthesis possibility. In S205, the display unit 110 displays an image indicating the calculated combination possibility. The utilization process is completed by the above.

FIG. 12 is a diagram showing an example of an image G1 for presenting synthesis possibilities to the user. The image G1 displays the numerical value of the synthesis possibility for the composition formula input by the user. This allows the user to easily recognize the possibility of synthesizing the composition formula to be evaluated.

As described above, according to the material information output system according to Embodiment 1, it is possible to assign a teacher label L indicating an appropriate synthesis possibility to a pseudo composition formula. can be suppressed. Then, since the synthesis possibility is calculated using the estimated model learned using the pseudo-composition formula to which the appropriate teacher label L is assigned, the appropriate value of the synthesis possibility is presented to the materials researcher. can be done. As a result, it is possible to prevent an unknown composition formula that is actually synthesizable from being determined to be unsynthesizable.

Further, the material information output device according to the present embodiment does not explicitly use the known composition formula list 101 for constructing the estimated model. That is, a pseudo composition formula including an unknown composition formula to which no teacher label L is assigned is given as an input, and an estimation model for estimating synthesis possibility is learned. Therefore, it is not necessary to hold the known composition formula list 101 when learning the estimation model. As a result, the company holding the known composition formula list 101 can leave the known composition formula list 101 as confidential information such as experimental data in the company, while the material information output device 100 for learning the estimation model can be provided as an external service. You will be able to build on the computer of the provider company. That is, the company holding the known composition formula list 101 can allow the service providing company to provide a material information service using the material information output device 100 without providing the known composition formula list 101 .

In order to realize this aspect, the remaining blocks not included in the material information output device 100 in FIG. Then, the material information output device 100 and the first computer may be communicably connected via a network such as the Internet. Alternatively, the input unit 103 and the display unit 110 in FIG. 1 may be configured by another second computer. In this case, the second computer is a computer held by a user to whom services are provided by the service provider company. As a result, the service providing company can provide services to users other than the company holding the known composition list 101 .

(Modification 1 of Embodiment 1)
13 is a block diagram showing an example of a configuration of a material information output system according to a modification of Embodiment 1. FIG. In this modification, the teacher label assigning unit 105 reads known composition formulas included in the known composition formula list 101 . The teacher label assigning unit 105 assigns a teacher label L of “1”, which is the maximum synthesizable value, to the known composition formulas included in the known composition formula list 101 . On the other hand, the teacher label assigning unit 105 assigns the teacher label L to the pseudo composition formula by setting the upper limit of the possibility of synthesis to "1-ε".

As a result, the estimation model learning unit 106 can learn the estimation model using not only the pseudo-composition formula to which the teacher label L is assigned but also the known composition formula to which the teacher label L is assigned as learning data. . As the value of ε, for example, a value such as ε=0.001 is set, but this is an example and can be changed as appropriate. Specifically, the teacher label assigning unit 105 may correct the teacher label L by multiplying the teacher label L calculated for the pseudo composition formula by 1-ε.

As a result, the estimation model can be learned by taking into account the known composition formula that can be reliably synthesized, and the estimation accuracy of the estimation model can be improved.

(Modification 2 of Embodiment 1)
The teacher label assigning unit 105 divides the pseudo-composition formula generated at the beginning of the learning process, that is, the pseudo-composition formula in which the numerical value t indicating the learning stage is 0, and the pseudo-composition formula generated at the end of the learning, that is, the learning stage. A teacher label L may be given only to the pseudo composition formulas whose indicated numerical value t is T, and only these pseudo composition formulas may be held in the pseudo composition formula holding unit 104 . This makes it possible to secure a free space in the memory of the pseudo composition formula holding unit 104 .

(Embodiment 2)
The material information output system according to Embodiment 2 generates an estimation model in consideration of the type of material whose synthesizability is to be estimated. FIG. 14 is a block diagram showing an example configuration of a material information output system according to Embodiment 2 of the present disclosure. In addition, in Embodiment 2, the same code|symbol is attached|subjected to the component same as Embodiment 1, and detailed description is abbreviate|omitted.

14 differs from FIG. 1 in that a combination calculation unit 1301 and a combination storage unit 1302 are newly added. The combination calculation unit 1301 extracts from the known composition formula list 101 the known composition formula of the same type as the type indicated by the type information of the material to be evaluated input by the input unit 103 as the first known composition formula. Then, the combination calculating unit 1301 executes voting processing, which will be described later, and causes the combination holding unit 1302 to store the obtained voting results.

The combination holding unit 1302 is configured by, for example, a semiconductor memory, and holds the voting results generated by the combination calculation unit 1301 .

The teacher label assigning unit 105 compares the pseudo composition formula held by the pseudo composition formula holding unit 104 with the voting results held by the combination holding unit 1302, and selects a pseudo composition formula from among the pseudo composition formulas held by the pseudo composition formula holding unit 104. A first pseudo-composition formula is extracted. Then, the teacher label assigning unit 105 assigns a teacher label to each of the first pseudo compositional formulas using the same method as in the first embodiment.

15 is a flowchart showing an example of processing in the learning process of the material information output system shown in FIG. 14. FIG. In the flow of FIG. 15, the same step numbers are assigned to the same processes as in FIG.

In S301, the combination calculation unit 1301 acquires the known composition formula included in the known composition formula list 101 and the type information indicating the type of material to be evaluated input by the user. Here, the type information is, for example, information for classifying materials such as lithium ion batteries and magnesium secondary batteries. Information such as a solid battery, a solar battery, and a thermoelectric material may be used as the type information.

Thereafter, the processing from S102 to S110 and the processing of S302 described in Embodiment 1 are executed in parallel. In S302, the combination calculation unit 1301 executes voting processing. FIG. 16 is a flowchart showing the details of voting processing.

In S1601, the combination calculation unit 1301 extracts from the known composition formula list 101 the known composition formula of the same type as the type indicated by the type information acquired in S301 as the first known composition formula.

FIG. 18 is a diagram showing an example of the data configuration of the known composition formula list 101 according to the second embodiment. In FIG. 18, "type" is further provided with respect to FIG. "Type" is information about the type of material indicated by each composition formula, such as "lithium ion battery" or "magnesium battery." A compositional formula is extracted as a first known compositional formula.

Refer back to FIG. In S1602, the combination calculation unit 1301 identifies a group combination of elements for each of the first known compositional formulas.

A group of elements refers to a cluster corresponding to one vertical column of the periodic table of elements. In this embodiment, the group of elements classifies the periodic table into seven clusters as shown in FIG. It consists of FIG. 17 is a diagram illustrating groups of elements. As shown in FIG. 17, in this embodiment, the groups of elements are alkali metals 1401, alkaline earth metals 1402, halogens 1403, rare gases 1404, transition elements 1405, and other metal elements 1406. , and other nonmetallic elements 1407 .

The combination of element groups corresponds to a combination of arbitrary two element groups among the seven element groups shown in FIG. 16 for a binary system, and a combination shown in FIG. Combinations of arbitrary three groups of elements among the seven groups of elements shown apply. Therefore, in the binary system, there are ₇ C ₂ =21 combinations of element groups, and in the ternary system, there are ₇ C ₃ =70 combinations of element groups.

For example, both Al ₂ O ₃ and MgO have the same element group combination because the element group combination is “other metal elements 1406” and “other non-metal elements 1407”. On the other hand, MgCl ₂ has a combination of element groups of “other metal element 1406” and “halogen 1403”, and therefore has a different combination of element groups from Al ₂ O ₃ .

Refer back to FIG. In S<b>1603 , the combination calculation unit 1301 votes for the combination of the group of elements of the first known compositional formula with respect to the predetermined combination of the group of elements obtained from the periodic table.

For example, if Al ₂ O ₃ is included in the first known compositional formula, one vote is cast for the combination of elements “other metallic element 1406” and “other non-metallic element 1407”. Assuming MgCl2 is included in the _first known compositional formula, one vote is cast for the combination of elements "Other metal elements 1406" and "Halogen 1403". The combination of predetermined element groups obtained from the periodic table corresponds to the above-mentioned 21 combinations of element groups in the case of a binary system, and the above-mentioned 70 combinations in the case of a ternary system. Combinations of groups of elements apply. Therefore, the voting result is, for example, a vote for the combination of the element group of "other metal element 1406" and "halogen 1403", the element of "other metal element 1406" and "other non-metal element 1407" Thus, there is a data structure in which predetermined element group combinations and the number of votes for each element group combination are associated with each other, such as b votes for a group combination.

In S1604, the combination calculation unit 1301 causes the combination holding unit 1302 to hold the voting results. Voting processing is completed as described above.

Refer back to FIG. In S<b>303 , the teacher label assigning unit 105 selects, among the pseudo-composition formulas held in the pseudo-composition formula holding unit 104 , the combination of the group of elements of the elements for which the voting result held in the combination holding unit 1302 is not 0. A pseudo-composition formula having a combination of groups is extracted as a first pseudo-composition formula. Then, the teacher label assigning unit 105 assigns the teacher label L to the first pseudo compositional formula using the same method as in the first embodiment.

After that, the processes of S112 and S113 are performed, and an estimated model is generated using the pseudo-composition formula to which the teacher label L is assigned in S303.

Different groups of elements have different characteristics of the elements. In addition, in machine learning, when the ratio of the number of data related to the material to be evaluated is small compared to other materials, the accuracy of estimating the possibility of synthesizing the material to be evaluated decreases. Therefore, in the present embodiment, a first known composition formula having the same type as the material to be evaluated is extracted, and a first pseudo composition formula having the same combination of the first known composition formula and the group of elements is used. An estimation model has been generated. Therefore, in the present embodiment, it is possible to improve the accuracy of estimating the possibility of synthesizing the material (composition formula) to be evaluated.

On the other hand, in the learning of the generating unit 802 in the generative model learning unit 102, the use of various types of data increases the possibility of generating pseudo-composition formulas that have not been conceived through previous experience. should not be limited. Therefore, in the present embodiment, as in the first embodiment, the generative model learning unit 102 uses all the known compositional formulas included in the known compositional formula list 101 to generate a generative model.

As described above, according to the material information output system according to the second embodiment, by allowing the user to select the type of the material to be evaluated, the possibility of synthesizing the material (composition formula) to be evaluated is determined. Estimation accuracy can be improved.

(Embodiment 3)
The material information output system according to Embodiment 3 estimates the formation energy of the pseudo composition formula, and corrects the teacher label L given to the pseudo composition formula according to the estimation result.

Formation energy is a characteristic value related to whether a substance can exist stably. If the energy of formation is too low, the material becomes less stable. On the other hand, if the energy of formation is too high, the stability of the material will also decrease. Therefore, substances can exist stably by having appropriate formation energy. The ability of a substance to exist stably means ease of synthesis. Therefore, in the present embodiment, the formation energy of the pseudo-composition formula is estimated, and the teacher label L is corrected using the estimation result.

FIG. 19 is a block diagram showing an example configuration of a material information output system according to Embodiment 3 of the present disclosure. In addition, in this embodiment, the same reference numerals are assigned to the same constituent elements as in the first and second embodiments, and the description thereof is omitted.

19 differs from FIG. 1 in that an energy estimation model learning unit 1601 and an energy estimation model holding unit 1602 are newly added. In this embodiment, it is assumed that the characteristic values of the known composition formulas included in the known composition formula list 101 include formation energy. That is, in the present embodiment, the formation energy of the known compositional formula is pre-stored in the memory.

The energy estimation model learning unit 1601 reads the pair information of the feature quantity, the descriptor, and the formation energy for each known composition formula from the known composition formula list 101, and adjusts the weight W of the neural network for estimating the formation energy. to generate an energy estimation model.

At this time, since the formation energy becomes a real value, this neural network solves the regression problem. That is, the energy estimation model is a neural network in which input data X is the feature quantity and the descriptor, and output data Y is the formation energy.

Then, the energy estimation model learning unit 1601 inputs the feature values and descriptors of each known composition formula included in the known composition formula list 101 to the energy estimation model obtained as a result of learning, Obtain formation energy. Then, the energy estimation model learning unit 1601 calculates the average and variance of the formation energies of the obtained known composition formulas. Here, the formation energy estimated by the energy estimation model is used to calculate the average and variance of the known energies, but the formation energies stored in the known composition formula list 101 may be used.

The neural network whose weight W has been adjusted is held in the energy estimation model holding unit 1602 as an energy estimation model. The information held in the energy estimation model holding unit 1602 includes configuration information indicating the number of layers and the number of units arranged in each layer of the neural network that constitutes the energy estimation model, as in the estimation model holding unit 107. and a load W representing the load value used in the load sum calculation in the unit. In addition, the energy estimation model storage unit 1602 also stores the average and variance of the formation energy of the known composition formula calculated by the energy estimation model learning unit 1601 .

The teacher label assigning unit 105 assigns the teacher label L to the pseudo compositional formula using the method of the first embodiment, and corrects the teacher label L using the formation energy of the pseudo compositional formula estimated by the energy estimation model. .

Specifically, first, the teacher label assigning unit 105 assigns the teacher label L to each pseudo compositional formula held in the pseudo compositional formula holding unit 104 by the method of the first embodiment. Next, the teacher label assigning unit 105 acquires the energy formation model from the energy estimation model holding unit 1602 . Next, the teacher label assigning unit 105 obtains the formation energy of each pseudo composition formula by inputting the feature amount and descriptor of each pseudo composition formula held in the pseudo composition formula holding unit 104 into the energy formation model. Next, the teacher label assigning unit 105 reads the average and variance of the formation energy of the known composition formula held in the energy estimation model holding unit 1602 . Next, the teacher label assigning unit 105 obtains the Mahalanobis distance M of each pseudo composition formula using the read mean and variance and the formation energy of each pseudo composition formula.

Next, the teacher label assigning unit 105 corrects the teacher label L by calculating L′=exp(−M)×L on the teacher label L assigned to the pseudo composition formula. L' represents the teacher label after correction. This means that the teacher label L obtained based on the number of times of learning is weighted according to the formation energy. As a result, the teacher label L is corrected so that its value decreases as the Mahalanobis distance M increases.

As mentioned above, if the energy of formation is too high or too low, the material will not be stable. Therefore, it can be said that the possibility of synthesizing a pseudo-composition formula having a formation energy far away from that of a known composition formula is low. Therefore, in the present embodiment, the teacher label L is corrected using the Mahalanobis distance M, which indicates how far the formation energy of the pseudo-composition formula is separated from the group of formation energies of the known composition formula. Thus, in the present embodiment, a teacher label L' based on physics knowledge can be assigned to a known composition formula. As a result, the present embodiment can generate an estimation model that can estimate the synthesis possibility with high accuracy.

In the present embodiment, the Mahalanobis distance M is used as the distance, but this is an example, and the distance to the group of formation energies of the known composition formula, such as the Euclidean distance to the average of the formation energies of the known composition formula, can be measured. Any distance may be adopted as long as it is the distance.

(Embodiment 4)
The material information output system according to Embodiment 4 presents to the user, when the user inputs the range of characteristic values, pseudo-composition formulas having characteristic values within that range together with the possibility of synthesis.

FIG. 20 is a block diagram showing an example of a configuration of a material information output system according to Embodiment 4 of the present disclosure. In this embodiment, the same components as in Embodiments 1 to 3 are denoted by the same reference numerals, and descriptions thereof are omitted.

20, the difference from FIG. 1 is that a characteristic value estimation model holding unit 1701, a generative model holding unit 1702, and a characteristic value calculation unit 1703 are further provided in the material information output device 100. .

As shown in Embodiment 1, the learning of the generative model learning unit 102 uses an adversarial learning algorithm by the generating unit 802 and the identifying unit 803 . When the characteristic value and the number of random numbers generated are set, the neural network NN1 of the generation unit 802 obtained as a result of learning generates a pseudo-composition formula that is estimated to have the set characteristic value. Only the number of occurrences can be generated. Further, when the feature quantity and the descriptor of the pseudo composition formula are input, the neural network NN2 of the identification unit 803 can output the estimation result of the characteristic value of the pseudo composition formula.

Therefore, in the present embodiment, the trained neural network NN1 and the trained neural network NN2 are used in the utilization process.

That is, the characteristic value estimation model holding unit 1701 holds the trained neural network NN2 as a characteristic value estimation model. The generative model holding unit 1702 holds the trained neural network NN1 as a generative model.

In the utilization process, the input unit 103 is input by the user not with the compositional formula but with the range of characteristic values. The characteristic value is, for example, the charge/discharge rate or capacity of the battery. Alternatively, the characteristic value may be Young's modulus, ionization potential, or the like. A material developer who is a user may want to search for a composition formula that satisfies a desired range of characteristic values. Therefore, in the present embodiment, the user is made to input the range of characteristic values using the input unit 103 . In addition, the user is made to input the number N of pseudo composition formulas to be generated using the input unit 103 . Here, the generated number N is a number that can be easily visually recognized on the display unit 110, and is 10 or 100, for example.

The descriptor calculation unit 108 acquires a generative model from the generative model holding unit 1702, inputs characteristic values to the generative model, and calculates feature amounts and descriptors.

Specifically, the descriptor calculation unit 108 generates N pieces of input data X=[random number, characteristic value] consisting of a pair of a characteristic value and a random number within the characteristic value range input by the user. Here, the N pieces of input data X are expressed as X1=[random number A1, characteristic value B1], X2=[random number A2, characteristic value B2], . . . , XN=[random number AN, characteristic value BN]. be.

Then, the descriptor calculation unit 108 sequentially inputs N pieces of input data X to the generative model, and obtains N pieces of output data Y representing feature amounts and descriptors, that is, N pieces of pseudo compositional formulas. Here, the descriptor calculation unit 108 may generate N characteristic values by randomly sampling N values from within the range of characteristic values input by the user to the input unit 103 .

A characteristic value calculation unit 1703 acquires the characteristic value estimation model held in the characteristic value estimation model holding unit 1701 . Then, the characteristic value calculation unit 1703 inputs the N feature values and descriptors obtained by the descriptor calculation unit 108 to the characteristic value estimation model as input data X, and estimates characteristic values for the N pseudo composition formulas. Calculate N output data Y representing the result.

Synthesis possibility calculation unit 109 acquires an estimated model from estimated model holding unit 107, inputs N feature amounts and descriptors calculated by descriptor calculation unit 108 to the estimated model, and calculates N pseudo composition formulas. Calculate the synthesizability.

The display unit 110 is a scatter diagram in which the synthesizing possibilities and the characteristic values of the N pseudo-composition formulas are plotted on a graph having a first axis indicating the synthesizing possibilities and a second axis indicating the characteristic values. is generated and displayed.

FIG. 21 is a diagram showing an example of a scatter diagram G2 displayed by display unit 110. As shown in FIG. The scatter diagram G2 has a two-axis graph 1801 with characteristic values on the vertical axis and synthesis possibilities on the horizontal axis. Below the graph 1801, the range of characteristic values input by the user using the input unit 103 and the number of generated pseudo-composition formulas are displayed. Here, 50 to 500 are input as characteristic values and 6 as the number of generations.

Graph 1801 plots six points PP1 representing the six pseudo-composition formulas generated. Also, near each point PP1, a composition formula calculated from the feature amount and the descriptor is displayed. As shown in the first embodiment, it is easy to convert the feature values and descriptors into the composition formula. This display allows material developers to easily recognize what kind of compositional formula can be generated within the range of targeted characteristic values.

FIG. 22 is a flow chart showing an example of processing in the utilization process of the material information output system according to the fourth embodiment.

In S401, the user inputs the range of characteristic values and the number N of pseudo compositional formulas to be generated to the input unit 103 . In S402, the descriptor calculation unit 108 generates N pieces of input data X consisting of pairs of characteristic values and random numbers from the input characteristic value range. In S403, the descriptor calculation unit 108 acquires the generative model from the generative model holding unit 1702, inputs N pieces of input data X=[random numbers, characteristic values] to the generative model, and generates N pseudo composition formulas. do.

In S404, the characteristic value calculation unit 1703 acquires the characteristic value estimation model from the characteristic value estimation model holding unit 1701, inputs N pseudo composition formulas to the characteristic value estimation model, and calculates N characteristic values.

Simultaneously with S404, in S405, the synthesizing possibility calculation unit 109 acquires an estimated model from the estimated model holding unit 107, inputs N pseudo composition formulas to the estimated model, and performs synthesis for each of the N pseudo composition formulas. Calculate probabilities.

In S406, the display unit 110 calculates the corresponding composition formula from the feature amount and descriptor of the pseudo composition formula calculated in S403, plots the N pseudo composition formulas on the graph 1801, and plots the plotted points PP1. A scatter diagram G2 in which composition formulas corresponding to the neighborhood are displayed on the graph 1801 is generated and displayed.

As described above, according to the material information output system according to Embodiment 4, when the user inputs the target characteristic value range, a pseudo composition formula having characteristic values within the range is generated, and the pseudo composition formula is generated. A scatter diagram G2 is generated in which the compositional formula is plotted on the graph 1801 . Therefore, the present embodiment can present useful judgment materials for assisting the material developer's analysis.

In the fourth embodiment, the graph 1801 is composed of two axes, but this is just an example. If there are parameters to be displayed other than characteristic values and synthesizing possibilities, the graph may be composed of three or more axes. good.

(Hardware configuration)
A material information output system is realized using a computer. FIG. 23 is a block diagram showing an example of hardware configuration for realizing the material information output system. The material information output system comprises a computer 2000, a keyboard 2011 and a mouse 2012 for giving instructions to the computer 2000, a display 2010 for presenting information such as calculation results of the computer 2000, and a program executed by the computer 2000. ODD (Optical Disk Drive) 2008 for reading.

A program executed by the material information output system is stored in a computer-readable optical storage medium 2009 and read by the ODD 2008 . Or read by NIC 2006 through a computer network.

A computer 2000 includes a CPU (Central Processing Unit) 2001, a ROM (Read Only Memory) 2004, a RAM (Random Access Memory) 2003, a HDD (Hard Disk Drive) 2005, and a NIC (Network Interface 6 and bus controller) 2000. 2007.

Furthermore, the computer 2000 may include a GPU (Graphical Processing Unit) 2002 for performing high-speed calculations.

CPU2001 and GPU2002 run the program read via ODD2008 or NIC2006. ROM 2004 stores programs and data necessary for the operation of computer 2000 . A RAM 2003 stores data such as parameters for program execution. The HDD 2005 stores programs, data, and the like. NIC 2006 communicates with other computers via a computer network. A bus 2007 connects the CPU 2001, ROM 2004, RAM 2003, HDD 2005, NIC 2006, display 2010, keyboard 2011, mouse 2012 and ODD 2008 to each other.

Note that the keyboard 2111, mouse 2012, and ODD 2008 connected to the computer 2000 may be removed if the display 2010 is a touch panel or if the NIC 2006 is used.

The computer 2000 shown in FIG. 23 may constitute each of the material information output device 100 shown in FIG. 1, the above-described first computer, and the second computer.

Furthermore, some or all of the constituent elements of the material information output system that constitute each of the devices described above may be configured from one system LSI (Large Scale Integration). A system LSI is an ultra-multifunctional LSI manufactured by accumulating a plurality of components on a single chip. Specifically, it is a computer system that includes a microprocessor, ROM, RAM, etc. . A computer program is stored in the RAM. The system LSI achieves its functions by the microprocessor operating according to the computer program.

Moreover, some or all of the components constituting each device described above may be configured from an IC card or a single module that can be attached to and detached from each device. An IC card or module is a computer system that consists of a microprocessor, ROM, RAM, and so on. The IC card or module may include the super-multifunctional LSI described above. The IC card or module achieves its function by the microprocessor operating according to the computer program. This IC card or this module may have tamper resistance.

The present disclosure may also be a method for executing the processing of the material information output system described above. Moreover, a computer program for implementing these methods by a computer may be included, or a digital signal composed of the computer program may be included.

Furthermore, the present disclosure provides a computer-readable non-transitory storage medium for the computer program or the digital signal, such as a flexible disk, hard disk, CD-ROM, MO, DVD, DVD-ROM, DVD-RAM, BD ( Blu-ray (registered trademark) Disc), semiconductor memory, or the like may be included. Moreover, the digital signal recorded in these non-temporary storage media may also be included.

Further, according to the present invention, the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication circuit, a network represented by the Internet, data broadcasting, or the like.

In addition, by recording the program or the digital signal in the non-temporary storage medium and transferring it, or by transferring the program or the digital signal via the network or the like, another independent computer It may be implemented by the system.

FIG. 24 is a diagram showing another example of the hardware configuration of the material information output system of the present disclosure; The example of FIG. 24 includes a computer 2000 and a data server 2101 connected to the computer 2000 via a network. The data server 2101 has memory for storing data. This data is read out to the computer 2000 via the network or the like. Also, the number of computers 2000 that read information from the data server 2101 does not have to be one, and may be plural. At that time, each computer 200 may implement a part of the components of the material information output system.

When adopting the hardware configuration of FIG. 24, the estimated model storage unit 107, the pseudo composition formula storage unit 104, and the known composition formula list 101 may be provided in the data server 2101. FIG. Moreover, when adopting the hardware configuration of FIG. 24, each block shown in FIG.

Furthermore, the above embodiments and modifications may be combined.

It should be considered that the embodiments disclosed this time are illustrative in all respects and not restrictive. The scope of the present disclosure is indicated by the scope of claims rather than the above description, and is intended to include all changes within the scope and meaning equivalent to the scope of the claims.

INDUSTRIAL APPLICABILITY According to the present invention, suitable synthesizing possibilities for unknown composition formulas can be presented, which is useful in the field of new material development.

100: Material information output device 101: Known composition formula list 102: Generative model learning unit 103: Input unit 104: Pseudo composition formula holding unit 105: Teacher label assigning unit 106: Estimation model learning unit 107: Estimation model holding unit 108: Description child calculation unit 109 : synthesis possibility calculation unit 110 : display unit 801 : random number generation unit 802 : generation unit 803 : identification unit 1301 : combination calculation unit 1302 : combination storage unit 1601 : energy estimation model learning unit 1602 : energy estimation model storage Unit 1701: Characteristic value estimation model holding unit 1702: Generative model holding unit 1703: Characteristic value calculation unit

Claims (11)

A material information output method in a material information output device that outputs information about materials,
The material information output device has a memory for pre-storing pseudo compositional formulas generated in the learning process by a pseudo compositional formula generation model generated by learning a list of known compositional formulas in association with numerical values indicating learning stages. prepared,
The processor of the material information output device,
calculating the synthesizability of the pseudo composition formula stored in the memory using the corresponding numerical values, and assigning the calculated synthesizability as a teacher label to the pseudo composition formula stored in the memory;
generating an estimation model for estimating the synthesis possibility by learning the pseudo-composition formula to which the teacher label is assigned;
calculating the synthesis possibility for an arbitrary composition formula using the learned estimation model;
outputting the calculated combinability;
Material information output method. Further, identifying a first known composition formula of the same type as the material to be estimated from among the known composition formulas included in the known composition formula list,
voting the first known compositional formula that matches the combination for a combination of a predetermined group of elements, holding the voting result in the memory;
assigning the supervised label indicating that the synthesizing possibility is 0 to the pseudo-composition formula whose combination matches the combination for which the voting result has not yet been voted, among the pseudo-composition formulas;
The material information output method according to claim 1. The known composition formula list further includes formation energies of known composition formulas,
obtaining from the memory the energy estimation model learned using the known composition formula list and the average and variance of formation energies included in the known composition formula list;
The formation energy of the pseudo composition formula is estimated using the energy estimation model, the distance of the formation energy of the pseudo composition formula to the formation energy of the known composition formula is calculated using the average and the variance, and the distance correcting the teacher label so that the value of the teacher label decreases as the
The material information output method according to claim 1 or 2. The memory further holds a characteristic value estimation model that outputs characteristic values for the pseudo composition formula,
get the characteristic value range entered by the user;
inputting a characteristic value within the characteristic value range into the generative model to generate the pseudo composition formula;
estimating the characteristic values of the generated pseudo composition formula using a characteristic value estimation model;
On a graph having at least two axes, a first axis indicating the synthesizability and a second axis indicating the characteristic value, the synthesizability and the estimated value of the characteristic value for the generated pseudo-composition formula are show the plotted scatterplot on the display,
The material information output method according to any one of claims 1 to 3. The known composition list is composed of ICSD, materials project, or NOMAD databases,
The material information output method according to any one of claims 1 to 4. The memory further holds an identification model that identifies whether the pseudo compositional formula generated by the generative model is genuine or fake,
The generative model and the discriminative model are composed of a neural network trained based on an adversarial learning algorithm,
The material information output method according to any one of claims 1 to 5. In assigning the teacher label, assigning a teacher label only to the pseudo composition formula generated at the beginning of the learning process and the pseudo composition formula generated at the end of the learning process, and holding the teacher label in the memory;
The pseudo-composition formula generated at the beginning of the learning stage is given a numerical value indicating that the synthesis possibility is 0,
The pseudo-composition formula generated at the end of the learning is given a numerical value indicating that the synthesis possibility is 1;
The material information output method according to any one of claims 1 to 6. The teacher label is assigned a numerical ratio indicating the learning stage with respect to the total number of learning times of the generative model.
The material information output method according to any one of claims 1 to 6. A material information output device that outputs information about materials,
a first memory for pre-storing pseudo composition formulas generated by learning a list of known composition formulas generated by the model in the learning process in association with numerical values indicating learning stages;
calculating the synthesis possibility of the pseudo composition formula stored in the first memory using the corresponding numerical values, and using the calculated synthesis possibility as a teacher label of the pseudo composition formula stored in the first memory; a teacher label assigning unit to assign
an estimation model learning unit that generates an estimation model for estimating the possibility of synthesizing the pseudo composition formula by learning the pseudo composition formula to which the teacher label is assigned;
a synthesizing possibility calculation unit that calculates the synthesizing possibility for an arbitrary composition formula using the estimated model;
and an output unit that outputs the synthesis possibility. a material information output device according to claim 9;
a second memory for pre-storing the known composition formula;
a generative model learning unit that generates the generative model by learning the known composition formula stored in the second memory;
and an input unit that receives input of the arbitrary composition formula. A program that causes a computer to function as the material information output device according to claim 9 .

Images