Prediction device, prediction method, and prediction program

Abstract

The present invention provides a predictive device that predicts the absolute value of a target performance based on data of experimental conditions used in past experiments or performance data obtained from past experiments, or that searches for experimental or manufacturing conditions corresponding to the target performance. [Solution] The prediction device comprises an acquisition unit that acquires information to be predicted, and a generation unit that inputs the target to be predicted to a trained model generated by machine learning using features extracted from training data related to the information to be predicted, and obtains output from the trained model to predict the difference in the performance of the target. The predetermined operation is an operation that extracts the features from the explanatory variable column of the table in which the training data is stored.

Description

[Technical Field]

[0001] This disclosure relates to a prediction device, a prediction method, and a prediction program. [Background technology]

[0002] Material design is carried out by repeatedly creating prototypes while adjusting the composition based on the experience of the material developer. For example, Patent Document 1 discloses a material design apparatus, a material design method, and a material design program for polymer materials that simultaneously satisfy multiple desired material properties. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] International Publication No. 2021 / 045058 [Overview of the project] [Problems that the invention aims to solve]

[0004] The performance data obtained from past experiments incorporates various ideas and perspectives, and it is believed that by utilizing the performance data obtained from past experiments, the performance of the experimental subject can be effectively predicted.

[0005] This disclosure is made in view of the above points, and aims to provide a prediction device, prediction method, and prediction program that predict the absolute value of the target performance, or search for experimental or manufacturing conditions corresponding to the target performance, using a trained model generated using data of experimental conditions used in past experiments or features extracted from performance data obtained from past experiments. [Means for solving the problem]

[0006] A prediction device according to a first aspect of this disclosure includes an acquisition unit that acquires information to be predicted, and a prediction unit that inputs the information to be predicted to a trained model generated by machine learning using features extracted by a predetermined operation on training data related to the information to be predicted, and obtains an output from the trained model to predict the absolute value of the performance of the target. The predetermined operation involves duplicating a first table in which the training data is stored to generate a second table, cross-joining the first table and the second table based on an identification column for identifying the training data to generate a new third table, calculating the difference between the explanatory variable column of the first table and the explanatory variable column of the second table in the generated third table, storing it in a newly generated column in the third table, and extracting the features by executing a feature transformation process obtained during the feature extraction operation on the explanatory variable column.

[0007] A prediction device according to a second aspect of this disclosure is a prediction device according to a first aspect, wherein the trained model is modeled after performing preprocessing on the features to make them fit the prediction model.

[0008] A prediction device according to a third aspect of this disclosure is a prediction device according to a first aspect, wherein the learning data is experimental design data or experimental result data.

[0009] A prediction device according to a fourth aspect of this disclosure is a prediction device according to a first aspect, wherein the acquisition unit acquires information on experimental conditions or manufacturing conditions as information on the prediction target.

[0010] A prediction device according to a fifth aspect of this disclosure is a prediction device according to a fourth aspect, wherein when the acquisition unit acquires information on experimental conditions or manufacturing conditions as information on the target of prediction, the prediction unit performs a process to predict the absolute value of the performance of the target, and when the acquisition unit acquires the performance of an object manufactured under the predicted experimental conditions or manufacturing conditions as information on the target of prediction, the prediction unit performs a process to predict or search for information on experimental conditions or manufacturing conditions.

[0011] A prediction device according to a sixth aspect of this disclosure is a prediction device according to a first aspect, wherein the feature quantity is information related to the identification sequence used in cross-combination.

[0012] The prediction device according to the seventh aspect of this disclosure is the prediction device according to the first aspect, wherein the feature quantities are extracted by dimensionality reduction processing.

[0013] The prediction method according to the eighth aspect of this disclosure is a process in which a processor acquires information to be predicted, inputs the information to be predicted into a trained model generated by machine learning using features extracted by a predetermined operation on training data related to the information to be predicted, and obtains an output from the trained model to predict the absolute value of the performance of the target. The predetermined operation is an operation in which a first table containing the training data is duplicated to generate a second table, the first table and the second table are cross-joined based on an identification column for identifying the training data to generate a new third table, the difference between the explanatory variable column of the first table and the explanatory variable column of the second table in the generated third table is calculated and stored in a newly generated column in the third table, and the feature extraction operation obtained during the feature extraction operation is performed on the explanatory variable column to extract the features.

[0014] A prediction program according to the ninth aspect of this disclosure involves causing a computer to perform a process to predict the absolute value of the performance of a target by acquiring information to be predicted, inputting the information to be predicted into a trained model generated by machine learning using features extracted by a predetermined operation on training data associated with the information to be predicted, and obtaining output from the trained model. The predetermined operation involves generating a second table by duplicating a first table in which the training data is stored, generating a new third table by cross-joining the first table and the second table based on an identification column for identifying the training data, calculating the difference between the explanatory variable column of the first table and the explanatory variable column of the second table in the generated third table, storing it in a newly generated column in the third table, and extracting the features by executing a feature transformation process obtained during the feature extraction operation on the explanatory variable column. [Effects of the Invention]

[0015] According to this disclosure, it is possible to provide a prediction device, prediction method, and prediction program that predict the absolute value of a target performance, or search for experimental or manufacturing conditions corresponding to a target performance, using a trained model generated using data of experimental conditions used in past experiments or features extracted from performance data obtained from past experiments. [Brief explanation of the drawing]

[0016] [Figure 1] This figure shows a schematic configuration of a prediction device according to an embodiment of the disclosed technology. [Figure 2] This is a block diagram showing the hardware configuration of the prediction device. [Figure 3] This is a block diagram showing an example of the functional configuration of a prediction device. [Figure 4] This figure shows an example of the duplication process. [Figure 5] This figure shows an example of a merge process. [Figure 6] This figure shows an example of arithmetic processing. [Figure 7]This is a diagram showing an example of extraction processing. [Figure 8] This is a diagram explaining the effect of data processing. [Figure 9] This is a block diagram showing an example of the functional configuration of a learning device for constructing a learned model. [Figure 10] This is a diagram showing an example of experimental conditions for which experimental results are to be predicted. [Figure 11] This is a diagram showing an example of features to be predicted. [Figure 12] This is a diagram showing an example of absolute physical properties predicted by a prediction unit. [[ID=1​​​​​​​​​​​​​​​​The information to be predicted input to the prediction device 10 is experimental conditions, for example, the absolute value of the wear performance of a certain product that is to be obtained as the result of the prediction. Specifically, information such as the formulation required to achieve the wear performance of a certain product is input to the prediction device 10 as the information to be predicted. In this case, the prediction device 10 uses the trained model 20 to output the absolute value of the wear performance of that product.

[0020] The prediction device 10 according to this embodiment predicts the performance of the target to be predicted using the trained model 20, thereby enabling prediction of data that reflects features based on experimental conditions used in past experiments and performance data obtained from past experiments.

[0021] Figure 2 is a block diagram showing the hardware configuration of the prediction device 10.

[0022] As shown in Figure 2, the prediction device 10 includes a CPU (Central Processing Unit) 11, ROM (Read Only Memory) 12, RAM (Random Access Memory) 13, storage 14, input unit 15, display unit 16, and communication interface (I / F) 17. Each component is connected to the others via a bus 19 so that they can communicate with each other.

[0023] The CPU 11 is a central processing unit that executes various programs and controls various components. Specifically, the CPU 11 reads a program from the ROM 12 or storage 14 and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculations according to the program recorded in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores a prediction program that predicts the performance of the experimental target.

[0024] ROM12 stores various programs and data. RAM13 temporarily stores programs or data as a working area. Storage14 consists of 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 for various types of input.

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

[0027] The communication interface 17 is an interface for communicating with other devices, and standards such as Ethernet®, FDDI, and Wi-Fi® can be used.

[0028] When executing the above generation program, the prediction device 10 uses the above hardware resources to implement various functions.

[0029] Next, the functional configuration of the prediction device 10 will be described.

[0030] Figure 3 is a block diagram showing an example of the functional configuration of the prediction device 10.

[0031] As shown in Figure 3, the prediction device 10 has an acquisition unit 101 and a prediction unit 102 as its functional configuration. Each functional configuration is realized by the CPU 11 reading and executing a generation program stored in the ROM 12 or storage 14.

[0032] The acquisition unit 101 acquires information to be predicted. This information is input, for example, by a user of the prediction device 10 using the input unit 15. A specific example of the information to be predicted is, for example, "experimental conditions for a certain product." In this case, the experimental conditions include material composition, material characteristics, mixing conditions, product configuration conditions, heat treatment conditions, etc. A specific example of the performance to be predicted is, for example, "the absolute value of the performance of a certain product." In this case, the product is a tire for automobiles, and the product performance includes wear performance, fuel efficiency, durability, etc. Another specific example of the information to be predicted is, for example, information on the performance of an object manufactured under predicted experimental or manufacturing conditions.

[0033] The prediction unit 102 makes predictions using the information to be predicted acquired by the acquisition unit 101. Specifically, the prediction unit 102 inputs the information to be predicted into the trained model 20 and predicts the performance of the target by obtaining the output from the trained model 20. The trained model 20 is a model generated by machine learning using the information to be predicted acquired by the acquisition unit 101 and features extracted by predetermined operations on the training data related to that information. The predetermined operation is an operation to extract features from the explanatory variable column of the table in which the training data is stored.

[0034] Furthermore, the operation to extract features from the explanatory variable columns of the table where the data to be analyzed is the same operation as described later, which involves duplicating the first table where the data to be analyzed is stored to generate a second table, cross-joining the first table and the second table based on an identification column for data identification to generate a new third table, calculating the difference between the explanatory variable columns of the first table and the explanatory variable columns of the second table in the generated third table, storing this difference in a newly generated column in the third table, and extracting features from the generated column.

[0035] The prediction device 10, by performing prediction processing using the trained model 20, can predict the absolute value of the performance of the target product, reflecting the underlying principles of past experimental data. The features based on past experimental results are features that reflect hypotheses, thoughts, or technical knowledge accumulated to obtain the target performance. In other words, the prediction device 10 according to this embodiment can significantly improve performance, shorten product development time, and improve the accuracy of selecting development directions by making predictions using the trained model 20 that reflects hypotheses, thoughts, or technical knowledge accumulated to obtain the target performance.

[0036] The following describes the data processing for calculating the features used in the machine learning of the trained model 20. The features used in the machine learning of the trained model 20 are extracted from tabular data. Tabular data can be of any type; for example, it may be stored in a database or entered in a spreadsheet. The data to be processed may include, for example, past test data for product prototyping, and may include information such as the proportions of materials used to manufacture a certain product, material features obtained from experimental results such as analysis or simulation results (molecular weight, softening point, particle shape, etc.), material processing time, and test results of the prototype product.

[0037] First, a replication process is performed to duplicate the data to be processed. That is, the replication process performs a process to duplicate the data to be processed in a table format. The data to be processed in a table format is an example of the first table in this disclosure, and the table generated by the replication process is an example of the second table in this disclosure.

[0038] Figure 4 shows an example of the replication process. The replication process replicates Table 200A to generate Table 200B, which has exactly the same data. The columns "Test Set" and "#" are columns that store information identifying the experimental result data. Of these, the column "Test Set" is an identification column that stores identification information used in the join and calculation processes described later, while the columns "NR", "BR", and "BR Molecular Weight" are columns that store the values ​​of the experimental result data, respectively. The values ​​of the experimental result data are examples of explanatory variables in this disclosure. In the following description, Table 200A will also be referred to as "Table A", and Table 200B as "Table B". In the following description, the column "Test Set" will be referred to as column (a), the column "NR" as column (b1), the column "BR" as column (b2), and the column "BR Molecular Weight" as column (b3).

[0039] It should be noted that the structure of the table-formatted data to be processed is not limited to the example shown. In Figure 4, table 200A has only one identification column called "Test Set," but for example, the data to be processed may have multiple identification columns. In addition, any preprocessing such as feature design including ratio calculation, normalization, and nonlinear transformation may be performed before the replication process.

[0040] Next, a join process is performed to cross-join the data to be processed and the data obtained by copying the data to be processed, using the identification column as the basis. Specifically, the join process uses the identification information in the data to be processed as the key to cross-join the data to be processed and the data obtained by copying the data to be processed, using the identification column as the basis to generate new data. The table generated by the join process is an example of the third table in this disclosure.

[0041] Figure 5 shows an example of a join process. The join process performs a cross join of tables 200A and 200B for all combinations of rows with the same column (a), thereby generating a new table 200C. In the example shown, the join process performs a join of the data from the first row of table 200A with the data from four rows of table 200B where column (a) is "A0001". Similarly, the join process performs a join of the data from rows 2 to 4 of table 200A with the data from four rows of table 200B where column (a) is "A0001". In other words, in the example shown, the join process performs a process that generates 16 rows of data where column (a) is "A0001".

[0042] Furthermore, the merge process may be performed on all data in the data to be processed, or it may not be performed on all data in the data to be processed.

[0043] Next, an operation is performed on the data generated by the join process, applying calculations to arbitrary columns. Specifically, the operation involves subtracting the same column from the original two tables onto the data generated by the join process, and storing the result of the subtraction in a new column. The column where the result of the subtraction is stored is also called the difference column.

[0044] Figure 6 shows an example of the calculation process. The "ΔNR" column is a difference column that stores the value obtained by subtracting the value of the "NR" column from Table B from the value of the "NR" column from Table A. The "ΔBR" column is a difference column that stores the value obtained by subtracting the value of the "BR" column from Table B from the value of the "BR" column from Table A. The "ΔMolecular Weight" column is a difference column that stores the value obtained by subtracting the value of the "BR Molecular Weight" column from Table B from the value of the "BR Molecular Weight" column from Table A.

[0045] The calculation process subtracts the value from Table B from the value from Table A for the same column in Table 200C, and stores the result in a new column. In the example in Figure 6, for the first row, the values ​​from Table A and Table B are the same, so the calculation process stores 0 in the "ΔNR", "ΔBR", and "ΔMolecular Weight" columns. For the second row and beyond, there is a difference between the values ​​from Table A and Table B, so the calculation process stores the values ​​obtained by subtracting the values ​​from Table B from the values ​​from Table A in the "ΔNR", "ΔBR", and "ΔMolecular Weight" columns, respectively.

[0046] Furthermore, before subtraction, any preprocessing such as feature design (including ratio calculation), normalization, or nonlinear transformation may be performed.

[0047] The extraction process extracts features from each difference column where the results of the calculation process are stored. Specifically, the extraction process extracts information related to the discriminant column used by the join process for the cross-join of the two tables as features for each difference column. In other words, the extraction process extracts information representing the differences in the discriminant column used by the join process for the cross-join of the two tables as features for each difference column. In other words, the extraction process extracts information common to each element of the discriminant column used by the join process for the cross-join of the two tables as features for each difference column. The extraction process can extract features using any method, but examples include dimensionality reduction (ICA (Independent Component Analysis), PCA (Principal Component Analysis), LSI (Latent Semantic Indexing), LDA (Linear Discriminant Analysis), IA, MDS, NMF, etc.), unsupervised learning to generate latent variables (t-SNE), and supervised learning (NN). The extracted features can be used for data analysis of the processed data.

[0048] Here, we show an example where data to be processed is subjected to merging and arithmetic operations before being extracted.

[0049] Figure 7 shows an example of the extraction process. In table 200C in Figure 7, the "IC component 1", "IC component 2", and "IC component 3" columns are columns that store feature quantities corresponding to the values ​​stored in the "ΔNR", "ΔBR", and "Δ molecular weight" columns. In this embodiment, the extraction unit 104 uses ICA to extract feature quantities from the values ​​stored in the "ΔNR", "ΔBR", and "Δ molecular weight" columns, and stores the extracted feature quantities in the "IC component 1", "IC component 2", and "IC component 3" columns, respectively.

[0050] If we let X be the matrix containing the values ​​before the subtraction operation, ΔX be the matrix containing the values ​​after the subtraction operation, and ΔS be the matrix containing the features, then the feature calculation on the matrix after the subtraction operation is: ΔS=W DI ·ΔX (1) It can be expressed by the following equation. ΔS is the explanatory variable. In other words, ΔX is a matrix that stores the differences in experimental conditions and manufacturing conditions (composition, temperature, etc.), and W DI This matrix represents the weights that are optimized to maximize or minimize the parameters used to find the features.

[0051] Furthermore, during the extraction process, any preprocessing, such as normalization or nonlinear transformation, may be performed on each difference sequence when extracting features.

[0052] Figure 8 illustrates the effects of the series of data processing described above. Figure 8 shows graphs of the relationship between IC component 1 column and IC component 3, the relationship between IC component 2 column and IC component 3, and the relationship between IC component 1 column and IC component 2, as shown in Figure 7.

[0053] As shown in the graph in Figure 8, it can be seen that there is a predetermined relationship between the features extracted as a result of the series of data processing described above. In other words, when combining and arithmetic operations are performed on the data to be processed and then extraction is carried out, it becomes easy to understand what kind of characteristics the extracted features have.

[0054] The feature extraction process described above extracted features from the difference column using dimensionality reduction processing, unsupervised learning to generate latent variables, supervised learning, etc. The trained model 20 of this embodiment is characterized by being trained using features obtained by performing feature transformation processing on the columns that are subject to subtraction processing of the data generated in the above-described join process, using dimensionality reduction processing after subtraction processing, unsupervised learning to generate latent variables, supervised learning, etc. In other words, the trained model 20 of this embodiment is characterized by being trained using features extracted by performing feature transformation processing on the values ​​stored in the "NR", "BR", and "Molecular Weight" columns of table 200A, using dimensionality reduction processing after subtraction processing, unsupervised learning to generate latent variables, supervised learning, etc., rather than features extracted from the values ​​stored in the "ΔNR", "BR", and "Molecular Weight" columns as shown in Figure 7.

[0055] Furthermore, depending on the feature extraction method, features may be extracted using operations other than matrix multiplication.

[0056] As described above, if we let X be the matrix containing the values ​​before the subtraction operation, ΔX be the matrix containing the values ​​after the subtraction operation, and S be the matrix containing the features, then the feature calculation on the matrix after the subtraction operation is: S=W DI ·ΔX (1) This can be expressed by the following equation. The matrix W obtained in this operation DI Using S'=W DI ·X (2) The operation represented by equation (2) allows us to obtain a matrix S' containing feature quantities for the values ​​before the subtraction operation. That is, if the values ​​stored in the "NR", "BR", and "Molecular Weight" columns of table 200A are matrix X, then the matrix S' obtained by multiplying matrix S by matrix W will be the feature quantities for the values ​​stored in the "NR", "BR", and "Molecular Weight" columns of table 200A.

[0057] The trained model 20 is then obtained by calculating the features extracted as a result of the series of data processing described above, and the resulting matrix W DI This is a model built using [the specified method / technology] to predict absolute values.

[0058] Figure 9 is a block diagram showing an example of the functional configuration of a learning device 30 for constructing a trained model 20. As shown in Figure 9, the learning device 30 includes an acquisition unit 301, a calculation unit 302, and a recording unit 303. The learning device 30 may be configured as the same device as the prediction device 10, or as a separate device.

[0059] The acquisition unit 301 acquires training data to be used for learning. The acquisition unit 301 acquires past experimental data X as training data, which is, for example, an example of values ​​before subtraction processing. Past experimental data is data that records experimental conditions and experimental results for said experimental conditions, such as Table 200A shown in Figure 4.

[0060] The calculation unit 302 performs the above-described replication, joining, and calculation processes to extract a matrix W for the feature quantities S of the training data. DI The calculation unit 302 also creates a preprocessing formula for the data to be predicted, which is represented by the following formula (3). In the following formula (3), S is the explanatory variable. S=W DI ·X (3)

[0061] The arithmetic unit 302 then uses the parameters obtained from equation (3) to generate a trained model 20 modeled by the following equation (4). Y=F(S) (4) In equation (4), Y is the dependent variable, and is, for example, a physical property value corresponding to experimental data X.

[0062] The computation unit 302 may perform existing dimensionality reduction processing or other feature extraction processing. For example, when using a multiple regression model to generate the trained model 20, the explanatory variables S may have multicollinearity, so the computation unit 302 may perform processing to resolve confounding. In this case, the trained model 20 is modeled by the following equation (5). Y=F(P) (5) In equation (5), P is an explanatory variable of the trained model 20, and is, for example, the result of principal component analysis of the feature quantity S extracted by the above calculation.

[0063] The recording unit 303 records the calculation formula for feature extraction shown in formula (1) and the trained model 20 modeled using the formulas represented by formula (4) or formula (5). The trained model 20 recorded by the recording unit 303 is used for prediction by the prediction unit 102.

[0064] This section will explain specific examples of predictions made by the prediction device 10. First, we will explain a specific example of a forward analysis in which the prediction device 10 predicts experimental results (e.g., physical properties) from experimental conditions. In the following explanation, we will describe a forward analysis that predicts hardness as a physical property.

[0065] The acquisition unit 101 acquires the experimental conditions X' for which the experimental results are to be predicted. Figure 10 shows an example of experimental conditions X' for which the experimental results are to be predicted. Figure 10 shows three examples of experimental conditions.

[0066] Next, the prediction unit 102 retrieves the feature calculation formula shown in equation (1) which is recorded in the recording unit 303. Subsequently, the prediction unit 102 calculates the feature quantity S' of the target to be predicted by applying the experimental condition X' to the feature calculation formula shown in equation (1). Figure 11 shows an example of the feature quantity S' of the target to be predicted obtained by applying the experimental condition X' to the feature calculation formula shown in equation (1).

[0067] Subsequently, the prediction unit 102 obtains the learned model 20 modeled by the mathematical formula represented by formula (3) or formula (4) recorded in the recording unit 303. Subsequently, the prediction unit 102 applies the feature amount of the prediction target to the learned model 20, and predicts the physical property Y' of the experimental result from the obtained experimental conditions. FIG. 12 is a diagram showing an example of the physical property Y' predicted by the prediction unit 102. In FIG. 12, as the physical property Y', the result predicted by the prediction unit 102 for the hardness from the experimental conditions is shown.

[0068] Next, a specific example of inverse analysis in which the prediction device 10 predicts experimental conditions from the absolute value of performance will be described. The acquisition unit 101 acquires information Y' on the performance (for example, hardness) for which experimental conditions are to be predicted. Subsequently, the prediction unit 102 acquires the learned model 20 recorded in the recording unit 303. Subsequently, the prediction unit 102 obtains the experimental conditions X' by applying an optimization method for solving a known inverse problem to the learned model 20. Examples of the optimization method for solving the inverse problem include the Gradient Descent method, Newton's method, Quasi-Newton methods, Stochastic Gradient Descent method, Steepest Descent method, Simulated Annealing, Genetic Algorithms, and the like.

[0069] When the learned model 20 extracts feature amounts by dimensionality reduction processing after subtraction processing for the column before subtraction processing, the prediction device 10 can predict the absolute value of performance using the learned model 20 by learning the experimental result before subtraction processing, that is, the absolute value of performance, using the feature amounts obtained by the operation using the matrix W. DI And since the learned model 20 is learned using the feature amounts extracted as a result of the above-described series of data processing, it becomes possible to predict data reflecting the idea incorporated in the analysis target data.

[0070] And since the learned model 20 is learned using the feature amounts extracted as a result of the above-described series of data processing, it becomes possible to predict data reflecting the idea incorporated in the analysis target data.

[0071] When extracting features used in constructing the trained model 20 from past experimental data, using experimental data on specific themes, categories, etc., makes it possible to construct a trained model 20 that embodies the ideas behind the experimental data for those themes, categories, etc.

[0072] For example, if a trained model 20 is constructed from features extracted from experimental data for a specific year, the prediction device 10 can construct an experimental concept by predicting the absolute value of the target's performance based on technologies or know-how discovered or developed in that year, or by predicting experimental conditions to which technologies developed in that year are applied.

[0073] For example, if a trained model 20 is constructed from features extracted from the experimental data of a specific experimental designer, the prediction device 10 can predict the absolute value of the performance of the target based on the experience, knowledge, experimental philosophy, and hypotheses of that experimental designer, or it can predict the experimental conditions based on the experience, knowledge, experimental philosophy, and hypotheses of that experimental designer, thereby enabling the construction of an experimental philosophy.

[0074] For example, if a trained model 20 is constructed from features extracted from experimental data for a specific product application, the prediction device 10 can predict the absolute value of the target's performance based on the technology, experience, experimental philosophy, and hypotheses in the domain of that product application, or it can predict experimental conditions that apply the technology, experience, experimental philosophy, and hypotheses suitable for improving the target performance of that product application or for multi-objective optimization, thereby enabling the construction of an experimental philosophy.

[0075] For example, if a trained model 20 is constructed from features extracted from experimental data for a specific component application, the prediction device 10 can predict the absolute value of the target's performance based on the technology, experience, experimental philosophy, and hypotheses in the domain of that component application, or it can predict experimental conditions that apply the technology, experience, experimental philosophy, and hypotheses suitable for improving the target performance of that component application and for multi-objective optimization, thereby enabling the construction of an experimental philosophy.

[0076] For example, if a trained model 20 is constructed from features extracted from experimental data of a specific development stage, the prediction device 10 can construct an experimental plan by predicting the absolute value of the experimental level or performance to be evaluated that should be carefully evaluated at that development stage, or by predicting experimental conditions suitable for improving or optimizing the target performance that is important at that development stage.

[0077] For example, if a trained model 20 is constructed from features extracted from experimental data of a specific business function or organizational unit, the prediction device 10 can predict the absolute value of the target performance based on the technical expertise of that business function or organizational unit, or it can construct an experimental concept by predicting experimental conditions to which the technical expertise, experimental concept, and hypothesis of that business function or organizational unit are applied.

[0078] For example, if a trained model 20 is constructed from features extracted from experimental data of a specific element under specific experimental conditions, the prediction device 10 can construct an experimental concept by predicting the absolute value of the performance of the target brought about by that specific element, or by predicting experimental conditions that utilize the effect of that specific element.

[0079] For example, if a trained model 20 is constructed from features extracted from experimental data from a specific sample provider, the prediction device 10 can predict the absolute value of the target's performance based on the technical expertise of that sample provider, or it can construct an experimental concept by predicting experimental conditions to which the technical expertise, experimental concept, and hypothesis of that sample provider are applied.

[0080] For example, if a trained model 20 is constructed from features extracted from experimental data for a specific experimental purpose, the prediction device 10 can construct an experimental concept by predicting the absolute value of the performance of the target based on a technology or hypothesis that is particularly effective in achieving that experimental purpose, or by predicting experimental conditions based on a technology or hypothesis that is particularly effective in achieving that experimental purpose.

[0081] For example, if a trained model 20 is constructed from features extracted from experimental data of a specific test evaluation method, the prediction device 10 can construct an experimental concept by predicting the absolute value of the target's performance based on techniques, hypotheses, or experimental conditions suitable for improving or optimizing the experimental results of that test evaluation method.

[0082] Next, the operation of the prediction device 10 will be explained.

[0083] Figure 13 is a flowchart showing the flow of the prediction process by the prediction device 10. The CPU 11 reads the prediction program from the ROM 12 or storage 14, loads it into the RAM 13, and executes it, thereby performing the prediction process.

[0084] In step S101, the CPU 11 acquires the experimental information X' to be predicted. The experimental information to be predicted is input, for example, by the user of the prediction device 10 using the input unit 15.

[0085] Following step S101, in step S102, the CPU 11 obtains a feature calculation formula for calculating features from the experimental information X'. The feature calculation formula is the formula represented by equation (3).

[0086] Following step S102, in step S103, the CPU 11 uses the feature calculation formula to calculate the feature S' of the experimental information X' obtained in step S101.

[0087] Following step S103, in step S104, the CPU 11 obtains a trained model 20 for predicting experimental results from the feature vector S'. The trained model 20 is modeled by the equations represented by equation (4) or equation (5) as described above.

[0088] Following step S104, in step S105, the CPU 11 inputs the feature vector S' into the trained model 20 to predict the experimental result Y' of the experimental information X'.

[0089] Following step S105, in step S106, the CPU 11 outputs the prediction result (experimental result Y'). The CPU 11 may output the prediction result by recording it in the storage 14, or by displaying it on the display unit 16.

[0090] The prediction device 10 predicts the test results using the trained model 20 through the series of processes described above, thereby predicting the absolute value of the target performance, which reflects the principles learned from past test results.

[0091] While embodiments of the present disclosure have been described in detail above with reference to the attached drawings, the technical scope of the present disclosure is not limited to these examples. It is clear that a person with ordinary skill in the art of the present disclosure may conceive of various modifications or alterations within the scope of the technical idea set forth in the claims, and these modifications or alterations are also understood to fall within the technical scope of the present disclosure.

[0092] Furthermore, the effects described in the above embodiments are descriptive or illustrative, and are not limited to those described in the above embodiments. In other words, the technology relating to this disclosure may produce other effects that would be obvious to a person of ordinary skill in the art of this disclosure from the descriptions in the above embodiments, in addition to or in lieu of the effects described in the above embodiments.

[0093] In addition, the prediction processing that the CPU reads and executes in each of the above embodiments may be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits that are processors with circuit configurations specifically designed to execute specific processing, such as ASICs (Application Specific Integrated Circuits). Furthermore, the prediction processing may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0094] Furthermore, while the above embodiments describe a configuration in which the prediction processing program is pre-stored (installed) in ROM or storage, the invention is not limited to this. 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), DVD-ROM (Digital Versatile Disk Read Only Memory), or USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that can be downloaded from an external device via a network. [Explanation of Symbols]

[0095] 10 Prediction device 20 Pre-trained Models

Claims

1. An acquisition unit that acquires information to be predicted, The system includes a prediction unit that inputs the information to be predicted to a trained model generated by machine learning using features extracted by predetermined calculations on the training data related to the information to be predicted, and obtains the output from the trained model to predict the absolute value of the performance of the target. The aforementioned predetermined calculation is, The first table containing the aforementioned training data is duplicated to generate a second table. The first table and the second table are cross-joined based on the identification column for identifying the learning data to generate a new third table. The difference between the explanatory variable column in the first table and the explanatory variable column in the second table in the generated third table is calculated and stored in a newly generated column in the third table. This is an operation to extract features by applying the feature transformation process obtained during the operation to extract features from the generated column to the column of explanatory variables. Prediction device.

2. The prediction device according to claim 1, wherein the trained model is modeled after performing operations to eliminate confounding on the explanatory variables.

3. The prediction device according to claim 1, wherein the learning data is data from an experimental design or data from an experimental result.

4. The prediction device according to claim 1, wherein the acquisition unit acquires information on experimental conditions or manufacturing conditions, or the performance of an object manufactured under predicted experimental conditions or manufacturing conditions, as information on the object to be predicted.

5. The prediction device according to claim 4, wherein when the acquisition unit acquires information on experimental conditions or manufacturing conditions as information on the target to be predicted, the prediction unit performs a process to predict the absolute value of the performance of the target, and when the acquisition unit acquires the performance of the target manufactured under the predicted experimental conditions or manufacturing conditions as information on the target to be predicted, the prediction unit performs a process to predict or search for information on experimental conditions or manufacturing conditions.

6. The prediction device according to claim 1, wherein the feature quantities are information related to the identification sequence used in the cross-combination.

7. The prediction device according to claim 1, wherein the aforementioned feature quantities are extracted by dimensionality reduction processing.

8. The processor, Obtain information to be predicted, The information to be predicted is input to a trained model generated by machine learning using features extracted through predetermined calculations on the training data associated with the information to be predicted, and the output from the trained model is obtained to predict the absolute value of the performance of the target. Execute the process, The aforementioned predetermined calculation is, The first table containing the aforementioned training data is duplicated to generate a second table. The first table and the second table are cross-joined based on the identification column for identifying the learning data to generate a new third table. The difference between the explanatory variable column in the first table and the explanatory variable column in the second table in the generated third table is calculated and stored in a newly generated column in the third table. This is an operation to extract features by applying the feature transformation process obtained during the operation to extract features from the generated column to the column of explanatory variables. Prediction method.

9. On the computer, Obtain information to be predicted, The information to be predicted is input to a trained model generated by machine learning using features extracted through predetermined calculations on the training data associated with the information to be predicted, and the output from the trained model is obtained to predict the absolute value of the performance of the target. Execute the process, The aforementioned predetermined calculation is, The first table containing the aforementioned training data is duplicated to generate a second table. The first table and the second table are cross-joined based on the identification column for identifying the learning data to generate a new third table. The difference between the explanatory variable column in the first table and the explanatory variable column in the second table in the generated third table is calculated and stored in a newly generated column in the third table. This is an operation to extract features by applying the feature transformation process obtained during the operation to extract features from the generated column to the column of explanatory variables. Prediction program.

Images