Information processing device, information processing method and information processing program

Abstract

To provide an information processing, a method and a program that improve prediction accuracy of characteristics of a composite material and search accuracy of similar blending by avoiding an increase in feature quantity number and generation of a missing value.SOLUTION: A method comprises: masking mixture amounts of some of raw materials of a composite material obtained by mixing a plurality of raw materials; repeatedly solving, for some or all mixtures of the composite material, a task to predict the masked raw materials or mixture amounts of the masked raw materials through a neural network based upon the mixture amounts of unmasked raw materials or whether the unmasked raw materials are mixed; updating a weight vector of computation so as to correctly predict the masked raw materials or the mixture amounts of the masked raw materials; calculating feature amount vectors of the respective raw materials through the neural network having the weight vector; and calculating a feature amount vector of mixtures of the composite material by the use of the calculated feature amount vectors of the respective raw materials and the mixture amount of the composite material.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, and an information processing program. [Background technology]

[0002] Machine learning is being used to predict the properties and performance of composite materials, such as rubber and resin composite materials, which are made by blending multiple raw materials and ingredients. For example, Patent Document 1 discloses a technique for predicting the properties of composite materials using, as feature quantities, the blending amounts of the raw materials used, the physical properties of the raw materials alone, such as melting points, and numerical values ​​of the molecular structures of the raw materials, such as fingerprints. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2023 / 032809 Summary of the Invention [Problem to be solved by the invention]

[0004] When predicting the properties of composite materials, if there is a wide range of raw materials for the composite material to be predicted and there are many of them, problems such as an increase in the number of features and the occurrence of missing values ​​can occur.

[0005] When predicting the properties of a composite material using the blending amounts of the raw materials used as features, the number of features increases with the number of raw material types being predicted. In addition, since there is generally a limit to the number of raw materials used in a single composite material, the majority of the data will be zero, and a dataset using blending amounts as a feature will become a sparse dataset, as shown in Table 1. As the number of features increases, problems such as a decrease in prediction accuracy and an increase in calculation time will arise.

[0006] [Table 1]

[0007] On the other hand, it is also possible to predict the properties of composite materials using features such as the average values ​​of the physical properties of the individual raw materials contained in the composite material, or average values ​​weighted by the number of parts. In this case, even if the range of raw materials to be predicted is expanded, the problem of an increase in the number of features mentioned above does not occur. However, depending on the raw material, there may be cases where the physical properties of the individual raw materials have not been measured, or where it is impossible to measure the physical properties. As a result, there may be many missing physical property values, which reduces the prediction accuracy.

[0008] The present disclosure has been made in consideration of the above points, and aims to provide an information processing device, an information processing method, and an information processing program that improve the accuracy of predicting the properties of composite materials and the accuracy of searching for similar formulations by avoiding an increase in the number of features and the occurrence of missing values. [Means for solving the problem]

[0009] An information processing device according to a first aspect of the present disclosure includes: a first calculation unit that masks the amounts of some of the raw materials in a composite material containing a plurality of raw materials, and repeatedly solves a task for some or all of the compositions of the composite material to predict the masked raw materials or the amounts of the masked raw materials using a neural network based on the amounts of the unmasked raw materials or based on whether the unmasked raw materials are included; updates a calculation weight vector so that the masked raw materials or the amounts of the masked raw materials can be correctly predicted; and calculates a feature vector of each of the raw materials using the neural network having the weight vector; and a second calculation unit that calculates a feature vector of the composite material composition using the feature vectors of each of the raw materials calculated by the first calculation unit and the amount of the composite material.

[0010] An information processing device according to a second aspect of the present disclosure is the information processing device according to the first aspect, wherein the first calculation unit multiplies the amount of the masked ingredients or information on whether the masked ingredients are included by an embedding matrix of any size, applies a predetermined activation function, multiplies the amount of the masked ingredients by an embedding recovery matrix of any size, and applies the predetermined activation function to calculate the embedding matrix as a feature vector of each ingredient.

[0011] An information processing device according to a third aspect of the present disclosure is the information processing device according to the first aspect, further comprising a prediction unit that predicts the relationship of physical properties associated with the feature vector of the composition of the composite material calculated by the second calculation unit, using the feature vector of the composition of the composite material calculated by the second calculation unit.

[0012] An information processing device according to a fourth aspect of the present disclosure is an information processing device according to the first aspect, further comprising a search unit that searches for a formulation similar to a specified formulation using the feature vector of the composite material formulation calculated by the second calculation unit.

[0013] An information processing device according to a fifth aspect of the present disclosure is an information processing device according to the fourth aspect, wherein the search unit searches for a set of multiple formulations similar to a specified set of multiple formulations using a feature vector of the set of multiple formulations obtained by calculating statistics of the feature vectors of the multiple formulations.

[0014] An information processing method according to a sixth aspect of the present disclosure includes a processor that performs a process of: masking the amounts of some of the raw materials in a composite material containing a plurality of raw materials; and using a neural network to predict the masked raw materials or the amounts of the masked raw materials based on the amounts of the unmasked raw materials or based on whether the unmasked raw materials are included; repeatedly solving this task for some or all of the compositions of the composite material; updating a weight vector of the calculation so that the masked raw materials or the amounts of the masked raw materials can be correctly predicted; calculating a feature vector of each of the raw materials using the neural network having the weight vector; and calculating a feature vector of the composite material composition using the calculated feature vectors of each of the raw materials and the amount of the composite material.

[0015] An information processing program according to a seventh aspect of the present disclosure causes a computer to execute a process of: masking the amounts of some of the raw materials in a composite material containing a plurality of raw materials; and predicting, using a neural network, the amounts of the masked raw materials or the masked raw materials based on the amounts of the unmasked raw materials or based on whether the unmasked raw materials are included; repeatedly solving this task for some or all of the compositions of the composite material; updating a weight vector of the calculation so that the masked raw materials or the amounts of the masked raw materials can be correctly predicted; calculating a feature vector of each of the raw materials using the neural network having the weight vector; and calculating a feature vector of the composite material composition using the calculated feature vector of each of the raw materials and the composition amounts of the composite material. [Effects of the Invention]

[0016] The present disclosure has been made in consideration of the above points, and can provide an information processing device, an information processing method, and an information processing program that improve the accuracy of predicting the properties of composite materials and the accuracy of searching for similar formulations by avoiding an increase in the number of features and the occurrence of missing values. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a diagram illustrating a schematic configuration of an information processing device according to an embodiment of the disclosed technology. [Figure 2] FIG. 2 is a block diagram showing a hardware configuration of the information processing device. [Figure 3] FIG. 2 is a block diagram illustrating an example of a functional configuration of an information processing device. [Figure 4] This figure shows an example of dimensionality reduction performed using t-sne on feature vectors of raw materials with an embedding dimension of 25, plotted on a two-dimensional plane. [Figure 5] This figure shows an example in which a feature vector of a composition is calculated from the feature vector of raw materials with an embedded dimension of 25, and then dimension reduction is performed using t-sne, followed by plotting on a two-dimensional plane. [Figure 6] 10 is a flowchart showing a flow of information processing by an information processing device. DETAILED DESCRIPTION OF THE INVENTION

[0018] An example of an embodiment of the present disclosure will be described below with reference to the drawings. The same reference numerals are used throughout the drawings to designate identical or equivalent components and parts. The dimensional proportions of the drawings are exaggerated for illustrative purposes and may differ from the actual proportions.

[0019] FIG. 1 is a diagram showing a schematic configuration of an information processing device according to this embodiment. The information processing device 10 according to this embodiment is a device that calculates a feature vector of a composite material formulation using data on the blend amounts of raw materials used in the composite material. The information processing device 10 according to this embodiment is a device that uses the calculated feature vector of the composite material formulation to predict the relationship between physical properties linked to the feature vector of the formulation and search for similar formulations.

[0020] The information processing device 10 according to this embodiment calculates a feature vector of the composition of a composite material using data on the composition amounts of the raw materials used in the composite material, thereby making it possible to avoid an increase in the number of feature vectors and the occurrence of missing values.The information processing device 10 according to this embodiment calculates a feature vector of the composition of a composite material using data on the composition amounts of the raw materials used in the composite material, making it possible to predict the properties of the composite material using the composition feature vector and search for similar compositions.

[0021] FIG. 2 is a block diagram showing the hardware configuration of the information processing device 10. As shown in FIG.

[0022] 2, the information processing device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0023] The CPU 11 is a central processing unit that executes various programs and controls each component. That is, the CPU 11 reads the programs from the ROM 12 or storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs recorded in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores an information processing program that calculates a feature vector of the composition of a composite material.

[0024] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured with a storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory, and stores various programs including the operating system and various data.

[0025] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to perform various inputs.

[0026] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may also function as the input unit 15 by adopting a touch panel system.

[0027] The communication interface 17 is an interface for communicating with other devices, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

[0028] When executing the above information processing program, the information processing device 10 uses the above hardware resources to realize various functions. The functional configuration realized by the information processing device 10 will be described.

[0029] FIG. 3 is a block diagram showing an example of the functional configuration of the information processing device 10. As shown in FIG.

[0030] 3, the information processing device 10 has, as functional components, a first calculation unit 101, a second calculation unit 102, a prediction unit 103, and a search unit 104. Each functional component is realized by the CPU 11 reading and executing an information processing program stored in the ROM 12 or the storage 14.

[0031] The first calculation unit 101 performs a process of masking the amounts of some of the raw materials in a composite material containing multiple raw materials, and using a neural network to predict the amounts of the masked raw materials or the amounts of the masked raw materials based on the amounts of the unmasked raw materials or based on whether or not the unmasked raw materials are included, by repeatedly solving the task for some or all of the combinations of the composite material, updating the calculation weight vectors so that the amounts of the masked raw materials or the masked raw materials can be correctly predicted, and calculating the feature vectors of each raw material using the neural network having the weight vectors. In the following description, the task of predicting the amounts of the masked raw materials or the amounts of the unmasked raw materials based on the amounts of the unmasked raw materials will be described, but other combinations can also be similarly implemented.

[0032] The processing executed by the first calculation unit 101 will be described. First, data on formulations that have actually been created in the past and whose physical properties have been measured is collected, and a formulation amount data table like Table 2 is prepared, in which each raw material type is represented as a column and the formulation is represented as a row, and the formulation amount is recorded. The formulation amount data table may be stored in any location. For example, the formulation amount data table may be stored in storage 14, or may be stored in an external storage device different from the information processing device 10. Here, in Table 2, the number of rows represents the number of formulations used in the analysis, and the number of columns represents the number of raw material types to be predicted.

[0033] [Table 2]

[0034] Although each formulation shown in the formulation data table in Table 2 needs to be a formulation that can actually be created, it is not necessary for the physical properties to be measured and linked at the feature generation stage.

[0035] The first calculation unit 101 divides the created blend amount data table by the maximum blend amount of each raw material, thereby normalizing the data in each column to a range of 0 to 1 and converting it into a data table such as Table 3. The first calculation unit 101 may also perform preprocessing such as standardization, binarization, and nonlinear conversion on the blend amount data table.

[0036] [Table 3]

[0037] The first calculation unit 101 extracts one row of data, i.e., one blend, from the standardized blend amount data table. The data for one blend is converted into a vector X in Here, we use the data for blend B. In this case, the vector X in As shown in equation (1), is a vector whose size is the number of raw material types to be predicted.

[0038]

number

[0039] The first calculation unit 101 calculates the vector X in non-zero elements of the vector X in A vector is generated by replacing some of the raw material amounts used in the formula with zero, and the vector X' in For example, the first calculation unit 101 calculates the vector X in A vector is generated by replacing one of the non-zero elements of with zero, and the vector X' is obtained as shown in equation (2). in Let's say.

[0040]

number

[0041] The first calculation unit 101 calculates the vector X' in A vector of the same size as vector X' inThe elements with the same index as the elements replaced with zeros for generation are set to 1, and the rest are set to zero. true Generate.

[0042]

number

[0043] The first calculation unit 101 calculates the vector X' in is multiplied by an embedding matrix We having a size of (embedding dimension) × (number of raw material types) as shown in formula (4) to obtain an embedding vector h of the combination as shown in formula (5). Here, the initial value of the embedding matrix We can be determined by providing a random number. The larger the embedding dimension, the better the expressive power, but setting it to an excessively large value will lead to an increase in calculation costs. In this embodiment, the embedding dimension is set to 25 or 50. Note that the embedding dimension may be changed to an appropriate value depending on the raw material types, the number of data, etc.

[0044]

number

[0045]

number

[0046] The first calculation unit 101 applies an activation function to the blended embedding vector h to obtain a vector h'. In this embodiment, the ReLU function is used as the activation function, but other activation functions such as a Leaky Relu function or a tanh function may also be used. The first calculation unit 101 multiplies the vector h' by an embedding recovery matrix Ws and applies a softmax function to obtain a vector y as shown in Equation (6). pred We obtain the vector y pred is the vector y true The input vector X' is calculated by making the loss function calculated from inIt is a vector that expresses the probability of predicting the ingredients replaced with 0 in

[0000] . In this embodiment, the softmax function is applied because there is one masked ingredient, but if there are two or more masked ingredients, it is desirable to apply the sigmoid function.

[0047]

number

[0048] The first calculation unit 101 calculates the vector y true and vector y pred A loss function is calculated from the difference between the two, and the values ​​of the embedding matrix We, bias b, and embedding recovery matrix Ws are updated so as to minimize the loss function. In this embodiment, binary cross entropy or cross entropy is used as the loss function. In this embodiment, stochastic gradient descent is used to minimize the loss function.

[0049] The first calculation unit 101 calculates a vector X in The process from generating the embedding matrix We to updating the values ​​of the embedding matrix We, bias b, and embedding recovery matrix Ws is repeated. Equation (7) shows an example of the embedding matrix We generated by the first calculation unit 101. After repeating the process to make the loss function sufficiently small, each row of the embedding matrix We becomes a distributed representation for each raw material type and can be used as a feature.

[0050]

number

[0051] The series of calculations performed by the first calculation unit 101 involves masking one of the ingredients used in the formula data table and solving the task of predicting the masked formula amount from the unmasked formula amount, thereby obtaining the feature values ​​of the ingredients. The series of calculations performed by the first calculation unit 101 is similar to the CBOW (Continuous Bag of Words) algorithm in natural language processing. CBOW is a type of model used in natural language processing and was proposed within Word2Vec. CBOW aims to predict a specific word based on the "context" (surrounding words) that surrounds it. This makes it possible to create a distributed representation of the word, i.e., to represent the meaning of the word as a dense vector. CBOW is based on the idea that the meaning of a word is determined by the context in which it is used.

[0052] Furthermore, while the series of computational processes performed by the first computation unit 101 was a task of predicting one masked ingredient, it is also possible to obtain ingredient features by solving the task of predicting other ingredients used at the same time from one type of ingredient. This is similar to the skip-gram algorithm in natural language processing. Skip-gram is one of the models that make up Word2Vec. While CBOW predicts a target word from context, Skip-gram is an algorithm that predicts surrounding context words from a specific word. Skip-gram generates dense vector representations of words, making it possible to capture semantic relationships between words.

[0053] In the above description, the series of calculation processes performed by the first calculation unit 101 involves masking one of the raw materials used from the blending amount data table and predicting the masked blend amount from the unmasked blend amount. However, the first calculation unit 101 may also mask one of the raw materials used from the blending amount data table and predict whether or not the masked raw material has been used from the unmasked blend amount. The first calculation unit 101 may also mask one of the raw materials used and predict the masked blend amount from whether or not the unmasked raw material has been used. The first calculation unit 101 may also mask one of the raw materials used and predict the masked raw material from whether or not the unmasked raw material has been used.

[0054] The embedding matrix We, which is a set of feature vectors for each ingredient obtained by the first calculation unit 101 as shown in Equation (7), represents the features of the ingredients. Figure 4 shows the feature vectors of ingredients with an embedding dimension of 25 plotted on a two-dimensional plane after dimensionality reduction using t-sne (t-distributed stochastic neighbor embedding). t-sne is a machine learning algorithm widely used for visualizing high-dimensional data, and is particularly suitable for mapping and visualizing patterns and clusters in multidimensional datasets into two- or three-dimensional space.

[0055] As shown in Figure 4, although information on the raw material classification is not used in calculating the feature vector, it can be seen that points of the same raw material classification are plotted at close positions in the space compressed by t-SNE. This indicates that the feature vector of the raw material calculated by the first calculation unit 101 represents the characteristics of the material.

[0056] The second calculation unit 102 calculates a feature vector of the composition of the composite material using the feature vectors of each raw material calculated by the first calculation unit 101, that is, the embedding matrix We, and the composition amount of the composite material.

[0057] The second calculation unit 102 calculates a vector X in and the embedding matrix We calculated by the first calculation unit 101, to calculate the feature quantity vector of the blend of the composite material as shown in Equation (8).

[0058]

number

[0059] Even if the number of raw material types used increases, the number of dimensions of the feature vector of the formulation becomes the embedding dimension, so the number of dimensions does not increase too much and no missing values ​​occur. Therefore, by having the second calculation unit 102 calculate the feature vector of the composite material formulation, the information processing device 10 can avoid an increase in the number of feature vectors and the occurrence of missing values.

[0060] The results of calculations performed by the first calculation unit 101 and the second calculation unit 102 on compounding amount data of rubber material for tires are shown as an example.

[0061] This shows an example of calculating a feature vector of a compounding material from the feature vectors of the raw materials, performing dimensionality reduction using t-SNE, and plotting it on a two-dimensional plane. Figure 5 shows an example of calculating a feature vector of a compounding material from the feature vectors of raw materials with an embedded dimension of 25, performing dimensionality reduction using t-SNE, and plotting it on a two-dimensional plane. The two-dimensional plane in Figure 5 shows the components that use rubber separately. As shown in the two-dimensional plane in Figure 5, it can be seen that points are concentrated for combinations of the same components, forming clusters. Generally, similar rubber compounds are used for the same components, and this can be seen to be reflected in the feature vector of the compounding material.

[0062] The calculation results of the first calculation unit 101 and the second calculation unit 102 may be stored in any location. For example, the feature vectors of the raw materials and the feature vectors of the composition may be stored in the storage 14, or may be stored in an external storage different from the information processing device 10.

[0063] The prediction unit 103 predicts the physical properties of the blend from the blend amount data, using the feature vector of the composite material blend calculated in advance by the first calculation unit 101 and the second calculation unit 102. For example, the prediction unit 103 predicts the physical properties of the blend amount data input to the information processing device 10 by the user of the information processing device 10, using the feature vector of the blend calculated by the first calculation unit 101 and the second calculation unit 102 for the blend amount data input to the information processing device 10 by the user of the information processing device 10, and the feature vector of the blend calculated in advance by the first calculation unit 101 and the second calculation unit 102.

[0064] The search unit 104 searches for blending amount data similar to blending amount data specified by the user of the information processing device 10, using the feature quantity vector of the blend of the composite material calculated in advance by the first calculation unit 101 and the second calculation unit 102. For example, the search unit 104 searches for blending amount data similar to the blending amount data input to the information processing device 10 by the user of the information processing device 10, using the feature quantity vector of the blend calculated by the first calculation unit 101 and the second calculation unit 102 for the blending amount data input to the information processing device 10 by the user of the information processing device 10, and the feature quantity vector of the blend calculated in advance by the first calculation unit 101 and the second calculation unit 102.

[0065] Before creating and testing a new formula, it is very important in development to check data on similar past formulas. However, in conventional formula searches, data was narrowed down using the amount of a specific ingredient or the amount of an ingredient group as a key. When narrowing down using the amount of an ingredient as a key, there is a possibility that ingredients with different names may be omitted even though they have similar functions. On the other hand, when using the amount of an ingredient group as a key, ingredients with different functions from the same ingredient group may be included in the search results, reducing search efficiency.

[0066] The search unit 104 can search for data of similar compositions using the feature vector of the composition of the composite material. This allows the information processing device 10 to output similar past compositions in the composition vector space when the user inputs a composition.

[0067] In addition, in formulation development, when creating and testing a new formulation, multiple formulations may be created and tested together. Generally, a single group contains similar formulations, and the formulation factors to be examined are varied within the group. Furthermore, since the creation and testing of each formulation within a single group are carried out at the same time, there are no seasonal differences in the test results within a single group, making it easy to handle. The search unit 104 may search for similar groups by using a vector obtained by calculating the statistics of the above-mentioned formulation vector. Examples of formulation vector statistics include the arithmetic mean, variance, and standard deviation. If the variance or standard deviation of a formulation vector is close to that of a specified set of formulations, this means that the formulation factors varied within the set of formulations are similar.

[0068] Next, the operation of the information processing device 10 will be described.

[0069] 6 is a flowchart showing the flow of information processing by the information processing device 10. The CPU 11 reads out an information processing program from the ROM 12 or the storage 14, loads it into the RAM 13, and executes it, thereby performing information processing.

[0070] In step S101, CPU 11 divides the compounding amount data table created as shown in Table 2 by the maximum compounding amount for each raw material, thereby normalizing the data in each column to the range of 0 to 1 and converting it into a data table such as Table 3.

[0071] Following step S101, in step S102, the CPU 11 extracts one row of data, i.e., one blend, from the standardized blend amount data table. The data for one blend is represented by a vector X in Let's say.

[0072] Following step S102, the CPU 11 calculates the vector X in non-zero elements of the vector X in A vector is generated by replacing some of the raw material amounts used in the formula with zero, and the vector X' in For example, CPU 11 calculates the vector X in A vector is generated by replacing one of the non-zero elements of with zero, and the vector X' is obtained as shown in equation (2). in Then, the CPU 11 calculates the vector X' in A vector of the same size as vector X' in The elements with the same index as the elements replaced with zeros for generation are set to 1, and the rest are set to zero. true Generate.

[0073] Following step S103, the CPU 11 calculates the vector X' in step S104. in This is multiplied by an embedding matrix We, which has a size of (number of embedding dimensions) × (number of raw material types) as shown in equation (4), to obtain an embedding vector h of the combination as shown in equation (5).

[0074] Following step S104, in step S105, the CPU 11 applies an activation function to the blended embedding vector h to obtain a vector h'. As described above, in this embodiment, the ReLU function is used as the activation function, but other activation functions such as a Leaky Relu function or a tanh function may also be used. In step S106, the CPU 11 multiplies the vector h' by the embedding recovery matrix Ws and applies a softmax function to obtain a vector y as shown in Equation (6). pred As described above, in this embodiment, the softmax function is applied because there is one masked raw material, but if there are two or more masked raw materials, it is desirable to apply the sigmoid function.

[0075] Following step S106, in step S107, the CPU 11 calculates the vector ytrue and vector y pred A loss function is calculated from the difference between the two, and the values ​​of the embedding matrix We, bias b, and embedding recovery matrix Ws are updated so as to minimize the loss function. In this embodiment, binary cross entropy or cross entropy is used as the loss function. In this embodiment, stochastic gradient descent is used to minimize the loss function.

[0076] Following step S107, in step S108, CPU 11 determines whether processing has been completed for all rows in the blending amount data table. If processing has not been completed for all rows in the blending amount data table (step S108; No), CPU 11 returns to step S102 and starts processing the next row. On the other hand, if processing has been completed for all rows in the blending amount data table (step S108; Yes), CPU 11 determines in step S109 whether the loss function has become sufficiently small. Whether the loss function has become sufficiently small can be determined by whether the feature vectors of the raw materials no longer change. If the loss function has not become sufficiently small (step S109; No), CPU 11 returns to step S102 and starts calculations on the blending amount data table again. On the other hand, if the loss function has become sufficiently small (step S109; Yes), CPU 11 ends the series of processes.

[0077] The information processing device 10 can obtain a feature vector of a raw material that indicates the characteristics of the material by executing a series of processes.

[0078] Although the embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings, the technical scope of the present disclosure is not limited to such examples. It is clear that a person skilled in the art of the present disclosure can conceive of various modifications or alterations within the scope of the technical idea described in the claims, and it is understood that these modifications or alterations also naturally fall within the technical scope of the present disclosure.

[0079] Furthermore, the effects described in the above embodiments are explanatory or exemplary and are not limited to those described in the above embodiments. In other words, the technology according to the present disclosure may achieve other effects that are obvious to a person skilled in the art of the present disclosure from the description in the above embodiments, in addition to or instead of the effects described in the above embodiments.

[0080] In the above embodiments, the information processing performed by the CPU after reading the software (program) may be performed by various processors other than the CPU. Examples of such processors include programmable logic devices (PLDs) whose circuit configuration can be changed after fabrication, such as field-programmable gate arrays (FPGAs), and dedicated electrical circuits, such as application-specific integrated circuits (ASICs), which are processors with circuit configurations specifically designed to perform specific processing. The information processing may be performed by one of these processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). The hardware structure of these processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor devices.

[0081] In addition, in each of the above embodiments, the information processing program is described as being pre-stored (installed) in a ROM or storage, but this is not limiting. The program may be provided in a form recorded on a non-transitory recording medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network. [Explanation of symbols]

[0082] 10. Information processing equipment 101 1st calculation section 102 2nd calculation section 103 Prediction Department 104 Search Section

Claims

1. a first calculation unit that masks the compounding amounts of some raw materials in a composite material containing a plurality of raw materials, and repeatedly solves a task for a part or all of the compoundings of the composite material to predict the masked raw materials or the compounding amounts of the masked raw materials using a neural network based on the compounding amounts of the unmasked raw materials or based on whether the unmasked raw materials are included, updates a calculation weight vector so that the masked raw materials or the compounding amounts of the masked raw materials can be correctly predicted, and calculates a feature vector of each of the raw materials using the neural network having the weight vector; a second calculation unit that calculates a feature vector of the blend of the composite material using the feature vectors of each of the raw materials calculated by the first calculation unit and the blend amount of the composite material; An information processing device comprising:

2. 2. The information processing device according to claim 1, wherein the first calculation unit calculates the embedding matrix as a feature vector of each of the ingredients by multiplying the blending amount of the masked ingredient or information on whether the masked ingredient is blended by an embedding matrix of any size, applying a predetermined activation function, multiplying the blending amount by an embedding recovery matrix of any size, and applying the predetermined activation function.

3. The information processing device according to claim 1 , further comprising a prediction unit that predicts a relationship between physical properties associated with the feature vector of the composition of the composite material calculated by the second calculation unit, using the feature vector of the composition of the composite material calculated by the second calculation unit.

4. The information processing device according to claim 1 , further comprising a search unit that searches for a composition similar to a specified composition using the feature vector of the composition of the composite material calculated by the second calculation unit.

5. The information processing device described in claim 4, wherein the search unit searches for a set of multiple formulations similar to a specified set of multiple formulations using a feature vector of the set of multiple formulations obtained by calculating statistics of the feature vectors of the multiple formulations.

6. The processor: a task of masking the compounding amounts of some of the raw materials in a composite material containing a plurality of raw materials, and predicting the masked raw materials or the compounding amounts of the masked raw materials using a neural network based on the compounding amounts of the unmasked raw materials or based on whether the unmasked raw materials are included, is repeatedly solved for a part or all of the compounding amounts of the composite material, updating a weight vector of the operation so that the masked raw materials or the compounding amounts of the masked raw materials can be correctly predicted, and calculating a feature vector of each of the raw materials using a neural network having the weight vector; The calculated feature vectors of the raw materials and the blending amounts of the composite material are used to calculate a feature vector of the blending of the composite material. A method for processing information.

7. On the computer, a task of masking the compounding amounts of some of the raw materials in a composite material containing a plurality of raw materials, and predicting the masked raw materials or the compounding amounts of the masked raw materials using a neural network based on the compounding amounts of the unmasked raw materials or based on whether the unmasked raw materials are included, is repeatedly solved for a part or all of the compounding amounts of the composite material, updating a weight vector of the operation so that the masked raw materials or the compounding amounts of the masked raw materials can be correctly predicted, and calculating a feature vector of each of the raw materials using a neural network having the weight vector; The calculated feature vectors of the raw materials and the blending amounts of the composite material are used to calculate a feature vector of the blending of the composite material. An information processing program that executes processing.

Images