Prototype condition proposal system, prototype condition proposal method

Abstract

Provided is technology that enables accurate proposal of a trial production condition for favorable materials even if there is little learning data to be used in the construction of a machine learning model or even if there is a deviation in the learning data distribution. A trial production condition proposal system 1 proposes a material trial production condition to a material developer, the system comprising: a regression model construction processing unit 112; and a trial production condition proposal processing unit 113. The regression model construction processing unit 112 executes a regression model construction process for measured characteristics data representing the results of measuring characteristics of a material. The trial production condition proposal processing unit 113 uses the constructed regression model to search for the optimal trial production condition for the material, and executes a trial production condition process on the basis of the search results. The regression model construction process includes: a process for calculating weight criteria, which are criteria for weighting the measured characteristics data; and a process for weighting the measured characteristics data on the basis of the calculated weight criteria.

Description

Technical Field

[0001] The present invention relates to a system and method for proposing prototype conditions of materials to material developers.

Background Art

[0002] In the field of materials science for research and development of materials (Materials), a method called Materials Informatics (MI) that efficiently predicts physical properties, structures, etc. of materials using information technologies (Informatics) such as statistical analysis and machine learning is widely used today. Regarding the research and development of materials using this Materials Informatics, for example, the technology of Patent Document 1 is known. Patent Document 1 discloses a system for estimating production conditions of a substance having optimal physical properties and structure from a dataset including production conditions of each of a plurality of substances used as samples and substance information representing physical properties and structures of each substance.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In materials development using materials informatics, the typical process involves first measuring the properties of the material under development through various evaluation tests and obtaining data representing the measurement results (hereinafter referred to as "measured property data"). Then, the acquired measured property data is input into a computer to construct various pre-trained machine learning models, and these pre-trained machine learning models are used to estimate the conditions for prototyping the material (hereinafter referred to as "prototyping conditions"). The accuracy of the computer's estimation of the prototype conditions (hereinafter referred to as "estimation accuracy") is determined by the accuracy of the pre-trained machine learning model used in the process of estimating the prototype conditions. The accuracy of this pre-trained machine learning model is generally determined by the number of measured property data points used for training during the construction of the machine learning model (hereinafter also referred to as "sample size") and the degree of variability in the distribution.

[0005] Here, the measured characteristic data input into a computer for building a machine learning model (hereinafter also referred to as "training data") generally has the characteristic of having a very large number of explanatory variables, while conversely having a small number of samples. Furthermore, if the amount of training data is small or if the distribution of training data is biased, the computer cannot build a highly accurate machine learning model. Therefore, in such cases, even if the computer is made to perform the process of estimating prototype conditions using the constructed machine learning model, there is a risk that sufficient estimation accuracy cannot be obtained.

[0006] In view of the above problems, the present invention aims to provide a technology that can accurately propose optimal prototyping conditions for materials even when the amount of training data used to construct a machine learning model is small or when the distribution of the training data is biased. [Means for solving the problem]

[0007] The prototype condition proposal system according to the present invention proposes prototype conditions for materials to material developers and comprises a regression model construction processing unit and a prototype condition proposal processing unit. The regression model construction processing unit performs regression model construction processing on property measurement data representing the measured results of material properties. The prototype condition proposal processing unit searches for the optimal prototype conditions for the material using the constructed regression model and performs prototype condition proposal processing based on the search results. The regression model construction processing is, Based on the difference between the objective variable included in the measured characteristic data and the target characteristic representing the target value of the material's properties, This is the criterion for weighting the measured characteristic data. This corresponds to the distance between the objective variable and the target characteristic. This process includes calculating weighting criteria and weighting the measured characteristic data based on the calculated weighting criteria. Furthermore, the prototype condition proposal method according to the present invention is a method of proposing prototype conditions for a material to a material developer using a computer. This prototype condition proposal method involves having a computer perform a regression model construction process and a prototype condition proposal process. The regression model construction process represents the process of constructing a regression model for measured property data that represents the measured results of the material's properties. The prototype condition proposal process represents the process of searching for the optimal prototype conditions for the material using the constructed regression model and proposing the prototype conditions for the material based on the search results. The regression model construction process is, Based on the difference between the objective variable included in the measured characteristic data and the target characteristic representing the target value of the material's properties, This is the criterion for weighting the measured characteristic data. This corresponds to the distance between the objective variable and the target characteristic. This process includes calculating weighting criteria and weighting the measured characteristic data based on the calculated weighting criteria.

[0008] Furthermore, the problems disclosed in this application and their solutions will be made clear from the description in the section on embodiments for carrying out the invention and from the drawings. [Effects of the Invention]

[0009] According to the present invention, even when the amount of training data used to construct a machine learning model is small or when the distribution of the training data is biased, it is possible to accurately propose the optimal prototyping conditions for materials. [Brief explanation of the drawing]

[0010] [Figure 1] A schematic diagram showing an overview of a prototype condition suggestion system according to one embodiment of the present invention. [Figure 2] A diagram showing the functional blocks of a prototype condition suggestion system according to one embodiment of the present invention. [Figure 3] A flowchart showing the overall processing flow of a prototype condition proposal system according to one embodiment of the present invention. [Figure 4] A flowchart showing the details of preprocessing of measured characteristic data. [Figure 5] A flowchart illustrating the details of the regression model construction process. [Figure 6] This figure shows a comparison of estimation accuracy before and after weighting. [Figure 7] A flowchart showing the details of the prototype condition proposal process. [Modes for carrying out the invention]

[0011] This embodiment will be described in detail below. Figure 1 is a schematic diagram showing an overview of a prototype condition proposal system according to one embodiment of the present invention. The prototype condition proposal system 1 shown in Figure 1 optimizes various prototype conditions that should be considered when prototyping a material, such as material composition and firing conditions, and proposes the results to the material developer, who is the user of this system. This optimization of prototype conditions is performed using various machine learning algorithms based on measured property data that represents the measured properties of the material. In this case, various regression models such as Gaussian Process Regression (GPR), Linear Regression, Regression Tree (including the case of ensemble method), Neural Network Regression, Support Vector Regression (SVR), Logistic Regression, and LASSO Regression (Least Absolute Shrinkage and Selection Operator Regression) are used as prediction models.

[0012] As shown in Figure 1, when a user of the prototype condition suggestion system 1 prototypes a material using the prototype conditions suggested by the system and measures the properties of the resulting material, new measured property data representing the evaluation values ​​of the properties of this material is generated. When this measured property data is used as new training data to train the prototype condition suggestion system 1, the system proposes more optimized prototype conditions to the user. In this embodiment, the prototype condition suggestion system 1 can propose prototype conditions with more favorable predicted property values ​​each time the user repeats this cycle. The prototype condition suggestion system may include a function to prototype a material and a function to measure the properties of the material to be prototyped, or these functions may be integrated into the system.

[0013] The prototype condition suggestion system 1 of this embodiment is implemented by a single general-purpose computer device, as shown in Figure 1. The following description assumes that the prototype condition suggestion system 1 is implemented by a single general-purpose computer device equipped with one or more processor devices, one or more storage devices, one or more input / output devices, and wired or wireless communication lines (none of which are shown) connecting them.

[0014] This computer device is installed, for example, as a terminal inside a laboratory and is connected to various other terminals installed inside and outside the laboratory, as well as to various terminals owned by each user, such as laptop PCs, tablets, and smartphones (hereinafter referred to as "user terminals"), and other equipment such as server devices, via a communication network such as the Internet 400 or a dedicated line. The computer device and the Internet 400 are connected by wire via well-known communication equipment (not shown), but they may also be connected wirelessly.

[0015] Next, various functions of the prototype condition proposal system 1 will be described with reference to FIG. 2. FIG. 2 is a diagram showing the functional blocks of the prototype condition proposal system according to an embodiment of the present invention. Note that each block described below represents a functional unit block, not a hardware unit configuration. As shown in FIG. 2, the prototype condition proposal system 1 of the present embodiment includes a control unit 11, a storage unit 12, a user interface unit 13, and a communication unit 14.

[0016] The control unit 11 executes various data processes based on the user's operation input detected by the user interface unit 13, the data acquired by the communication unit 14, and the programs and data stored in the storage unit 12. The control unit 11 also functions as an interface for the user interface unit 13, the communication unit 14, and the storage unit 12.

[0017] The control unit 11 has functional blocks of a characteristic measurement data preprocessing unit 111, a regression model construction processing unit 112, and a prototype condition proposal processing unit 113. The control unit 11 is configured using a processor device such as a CPU (Central Processing Unit) and various coprocessors (hereinafter also simply referred to as "processor"), and these functional blocks can be realized by executing a predetermined program. Note that instead of the processor, the control unit 11 may be configured using a logic circuit such as an FPGA (Field Programmable Gate Array). Further, the control unit 11 may be configured by a combination of a processor and a logic circuit.

[0018] The program executed by the control unit 11 may be installed from a program source. The program source may be, for example, a program distribution computer or a computer-readable recording medium, etc. Also, the program executed by the control unit 11 may be composed of a device driver, an operating system, various application programs located in the upper layers thereof, and a library that provides common functions to these programs. Furthermore, two or more programs may be realized as one program, or one program may be realized as two or more programs.

[0019] The characteristic measurement data preprocessing unit 111 performs preprocessing on the characteristic measurement data in the state of so-called raw data immediately after recording. This processing performed by the characteristic measurement data preprocessing unit 111 is referred to as characteristic measurement data preprocessing.

[0020] The regression model construction processing unit 112 executes a process of constructing a regression model for the characteristic measurement data subjected to preprocessing. This processing executed by the regression model construction processing unit 112 is referred to as regression model construction processing.

[0021] The prototype condition proposal processing unit 113 searches for the optimal prototype conditions of the material using the regression model constructed by the regression model construction processing unit 112 for the characteristic measurement data subjected to preprocessing, and executes a process of proposing the prototype conditions of the material to the user based on the search results. This processing executed by the prototype condition proposal processing unit 113 is referred to as prototype condition proposal processing.

[0022] Note that the specific contents of these processes will be described later.

[0023] The memory unit 12 is configured using a storage device such as RAM or flash memory, and stores programs that supply various processing instructions to the control unit 11, as well as data representing various information used in the processing executed by the control unit 11. For example, the memory unit 12 stores characteristic measurement data that has been preprocessed by the characteristic measurement data preprocessing unit 111 (hereinafter referred to as "preprocessed data"), and data representing the regression model constructed by the regression model construction processing unit 112. The control unit 11 can realize the aforementioned functional blocks of the characteristic measurement data preprocessing unit 111, the regression model construction processing unit 112, and the prototype condition proposal processing unit 113 by reading and writing this information to the memory unit 12.

[0024] The user interface unit 13 accepts input operations from the user and is responsible for processing related to the user interface, such as displaying images and outputting sound. The user interface unit 13 has functional blocks for input unit 131 and output unit 132. The input unit 131 detects various operations from the user. The input unit 131 is configured using, for example, a keyboard, pointing device, or touch panel. The output unit 132 performs actions such as displaying images on the screen or outputting sound to the user. The output unit 132 is configured using, for example, a liquid crystal display or touchscreen.

[0025] The communications unit 14 is responsible for processing communication between user terminals owned by each user and other devices such as server equipment, which takes place via the Internet 400. The communications unit 14 is configured using, for example, a NIC (Network Interface Card) or an HBA (Host Bus Adapter).

[0026] In this embodiment, each function of the prototype condition suggestion system 1 has been described as being integrated and implemented by a single computer device. However, these functions may be implemented by multiple interconnected computer devices or server devices. Furthermore, the prototype condition suggestion system 1 may consist of a general-purpose computer device such as a laptop PC and a web browser installed thereon, or it may consist of a web server and various portable devices.

[0027] Furthermore, the descriptions of each function are merely examples, and multiple functions may be combined into one function, or one function may be divided into multiple functions.

[0028] Next, the overall processing flow of the prototype condition suggestion system 1 will be explained with reference to Figure 3. Figure 3 is a flowchart showing the overall processing flow of the prototype condition suggestion system according to one embodiment of the present invention. In the following explanation, the processing may be described using each of the aforementioned functions or programs as the subject, but the processing described using a function or program as the subject may also be processing performed by a processor or a device having such a processor.

[0029] In step S310, the control unit 11 performs preprocessing of the measured characteristics data using the measured characteristics data preprocessing unit 111. This preprocesses the measured characteristics data, making it preprocessed data, which allows subsequent processes to be executed normally. Details of the measured characteristics data preprocessing performed in step S310 will be explained later with reference to the flowchart in Figure 4. Once the measured characteristics data preprocessing is complete, the control unit 11 proceeds to step S320.

[0030] In step S320, the control unit 11 executes the regression model construction process using the regression model construction processing unit 112. This constructs a regression model for the pre-processed data. Details of the regression model construction process performed in step S320 will be explained later with reference to the flowchart in Figure 5. Once the regression model construction process is complete, the control unit 11 proceeds to step S330.

[0031] In step S330, the control unit 11 executes a regression model evaluation process. In this regression model evaluation process, the generalization performance, which is an indicator representing the prediction accuracy of the regression model, is evaluated for each of the multiple regression models constructed as a result of the processes up to step S320. This evaluation is performed, for example, by performing cross-validation with other regression models. The evaluation results are visualized using graphs such as scatter plots and box plots. This allows the user to receive suggestions for prototype conditions based on regression models with good generalization performance. Once the regression model evaluation process is complete, the control unit 11 proceeds to step S340.

[0032] In step S340, the control unit 11 executes a prototype condition proposal process using the prototype condition proposal processing unit 113. In this prototype condition proposal process, the user of the prototype condition proposal system 1 can modify the prototype conditions for the material proposed by the prototype condition proposal system 1 as needed to make them even more favorable. The control unit 11 obtains predicted values ​​for the material properties when prototyped using the modified prototype conditions by the user by applying them to a selected regression model and presents them to the user. In other words, the user can interactively modify the prototype conditions while confirming the predicted values. Thus, the prototype condition proposal system 1 is a system that can incorporate the knowledge of the user, who is the developer of the material, into the prototype conditions for the material proposed to the user. Details of the prototype condition proposal process performed in step S340 will be explained later with reference to the flowchart in Figure 7. Once the prototype condition proposal process is complete, the control unit 11 terminates the process shown in the flowchart in Figure 3.

[0033] The prototype condition suggestion system 1 of this embodiment performs the processes in steps S310 to S340 of Figure 3 to suggest good prototype conditions to the user. Specifically, in step S310 of the prototype condition suggestion system 1 of this embodiment, the necessary preprocessing is automatically applied to the measured characteristic data in its raw state. Therefore, the processes in steps S320 to S340 can be executed without the need for manual, complex preprocessing of the measured characteristic data. Furthermore, in the prototype condition suggestion system 1 of this embodiment, the user can select a regression model with good generalization performance. Therefore, the prototype condition suggestion system 1 can suggest prototype conditions to the user that result in good predicted characteristic values.

[0034] As mentioned above in relation to Figure 1, the user of the prototype condition proposal system 1 can perform a prototype of a material using the prototype conditions proposed by the prototype condition proposal system 1, measure the properties of the resulting product, and then train the prototype condition proposal system 1 with the data representing the measurement results as new property measurement data. After that, the user can have the prototype condition proposal system 1 perform each of the steps S310 to S340 in Figure 3 again. In this case, the prototype condition proposal system 1 can propose more optimized prototype conditions to the user. That is, the prototype condition proposal system 1 of this embodiment can propose prototype conditions with better predicted property values ​​each time the steps S310 to S340 in Figure 3 are repeated for the same prototype target.

[0035] Figure 4 is a flowchart showing the details of the preprocessing of measured characteristic data.

[0036] In step S410, the control unit 11 receives the input of the measured characteristics data from the user via the input unit 131 or the communication unit 14 using the measured characteristics data preprocessing unit 111. The measured characteristics data to be input to the prototype condition proposal system 1 may be, for example, categorical data, continuous data, or discrete data. The specific data format of the measured characteristics data to be input to the prototype condition proposal system 1 can be determined as appropriate. Once the processing in step S410 is complete, the control unit 11 proceeds to step S420.

[0037] In step S420, the control unit 11 sets the type of variable for the measured characteristic data input from the user in step S410 using the characteristic measurement data preprocessing unit 111. Here, either an explanatory variable or a target variable is set. An explanatory variable is a variable that forms the basis for determining the predicted value of the characteristic. In this embodiment, the composition of the material constituting the prototype conditions and the firing conditions are examples of explanatory variables. A target variable is a variable that represents the characteristic value of the prototype material that is the subject of prediction. As an example of the specific processing in step S420, the explanatory variable may be set by default, and the control unit 11 may accept a setting operation from the user who wishes to change it to the target variable. Once the processing in step S420 is complete, the control unit 11 proceeds to step S430.

[0038] In step S430, the control unit 11 uses the characteristic measurement data preprocessing unit 111 to determine whether there are any outliers in the characteristic measurement data for which one of the explanatory variables and the objective variable was set in step S420. This determination of the presence or absence of outliers is performed, for example, by first determining whether there are outliers outside the range of mean ± 2σ by representing the characteristic measurement data as a histogram, and then determining whether there were any data input errors or malfunctions in the evaluation test machine when generating the characteristic measurement data for which outliers were determined to be present. If the control unit 11 determines that the characteristic measurement data contains outliers, it deletes the outliers and makes them missing values, and proceeds to step S440. Furthermore, if the type of outlier included in the characteristic measurement data is one of the objective variables in a case where the explanatory variables completely overlap but there are levels with different characteristics, the control unit 11 deletes the sample related to this outlier and proceeds to step S440. However, when treating levels with different characteristics where all explanatory variables overlap as outliers, it is necessary to delete the level itself, unlike when only one explanatory variable is outlier due to an input error or the like. On the other hand, if the control unit 11 determines that the measured characteristic data does not contain outliers, it proceeds to step S440.

[0039] Furthermore, in cases where all of the explanatory variables overlap and there are levels with different characteristics, it may be desirable to retain both explanatory variables, assuming that the overlap is meaningful. In such cases, the prototype condition proposal system 1 of this embodiment can omit the processing in step S430.

[0040] In step S440, the control unit 11 determines whether or not there are missing values ​​in the measured characteristic data using the pre-processing unit 111. This is because the measured characteristic data may already contain missing values. If the control unit 11 determines that the pre-processed data contains missing values, it imputes the missing values ​​and proceeds to step S450. The imputation of missing values ​​is performed, for example, by using the mean, median, minimum, or maximum value of the measured characteristic data (excluding outliers) as the value to impute the missing values. Alternatively, the imputation of missing values ​​may be performed by linear interpolation. The pre-processing unit 111 may, for example, display the imputed value in red when it imputes missing values ​​in step S440, to make it easier to identify. Alternatively, the pre-processing unit 111 may, for example, delete the level itself without imputing the missing value in step S440. Furthermore, in the prototype condition proposal system 1 of this embodiment, the characteristic measurement data preprocessing unit 111 can also delete the explanatory variables themselves if there is a large percentage of missing data in the explanatory variables, for example, if more than 50% of the data is missing. On the other hand, if it is determined that the characteristic measurement data does not contain any missing values, the control unit 11 proceeds to step S450.

[0041] In step S450, the control unit 11, using the characteristic measurement data preprocessing unit 111, performs encoding processing on the explanatory variable if it is a categorical value rather than a continuous value, converting it into numerical data. The characteristic measurement data preprocessing unit 111 performs this encoding processing by, for example, referring to a record in a table stored in the storage unit 12 that represents the correspondence between categorical data and numerical data. Once the processing in step S450 is complete, the control unit 11 proceeds to step S460.

[0042] In step S460, the control unit 11 determines, using the characteristic measurement data preprocessing unit 111, whether or not the characteristic measurement data contains redundant explanatory variables. This determination is based on whether or not it is possible to extract combinations of explanatory variables whose correlation coefficient is greater than or equal to a predetermined number, for example, 0.8 or more. If it is determined that the characteristic measurement data contains redundant explanatory variables, the control unit 11 deletes one of the redundant explanatory variables and proceeds to step S470. In this embodiment of the prototype condition proposal system 1, combinations of explanatory variables whose correlation coefficient is 0.8 or more are visualized to the user through the output unit 132, and the user can select the explanatory variable to delete through the input unit 131. On the other hand, if it is determined that the characteristic measurement data does not contain redundant explanatory variables, the control unit 11 proceeds directly to step S470.

[0043] In step S470, the control unit 11 uses the characteristic measurement data preprocessing unit 111 to perform standardization processing on the characteristic measurement data as necessary. This standardization processing transforms the scale of the characteristic measurement data so that the mean = 0 and the standard deviation (variance) = 1. Once the processing in step S470 is complete, the control unit 11 stores the preprocessed characteristic measurement data, which has been preprocessed in steps S410 to S470 of Figure 4, into the storage unit 12, and terminates the characteristic measurement data preprocessing shown in the flowchart of Figure 4. The control unit 11 may also use the characteristic measurement data preprocessing unit 111 to perform normalization processing on the characteristic measurement data as necessary. In such cases, the standardization processing in S470 may be omitted, and the subsequent processes may be executed.

[0044] Figure 5 is a flowchart detailing the regression model construction process.

[0045] In step S510, the control unit 11 selects the cross-validation conditions for evaluating the regression model to be constructed on the pre-processed data, using the regression model construction processing unit 112. In this embodiment, the prototype condition proposal system 1 performs evaluation for each regression model using K-fold cross-validation. The default value for this is set to K=10. In this case, the prototype condition proposal system 1 evaluates the regression model using 10-fold cross-validation. In this embodiment, the prototype condition proposal system 1 also allows the user to select the cross-validation conditions. That is, the regression model construction processing unit 112 can receive cross-validation conditions from the user via the input unit 131 or the communication unit 14. Once the processing in step S510 is complete, the control unit 11 proceeds to step S520.

[0046] In step S520, the control unit 11 uses the regression model construction processing unit 112 to select candidate regression models to be used as predictive models for exploring prototype conditions. At this time, the regression model construction processing unit 112 selects regression models as candidates that have been selected by the user via the input unit 131 or the communication unit 14. In the prototype condition proposal system 1 of this embodiment, the user can select multiple regression models as candidates from Gaussian process regression and various regression models such as linear regression, regression trees (including cases using ensemble methods), neural network regression, support vector regression, logistic regression, and LASSO regression. Once the processing in step S520 is complete, the control unit 11 proceeds to step S530.

[0047] In step S530, the control unit 11 uses the regression model construction processing unit 112 to calculate the weighting criteria, which are the basis for weighting the measured characteristic data. The weighting criteria are calculated based on the difference between the dependent variable included in the measured characteristic data and the target characteristic representing the target value of the material's properties, or on a statistic that represents the rarity of the explanatory variables included in the measured characteristic data (details will be described later). A specific example of a statistic that represents the rarity of an explanatory variable is the probability of an explanatory variable that satisfies predetermined conditions. Once the processing in step S530 is complete, the control unit 11 proceeds to step S540.

[0048] In step S540, the control unit 11 executes a process in which the regression model construction processing unit 112 weights the measured characteristic data based on the weight criteria calculated in step S530. In addition, in the prototype condition proposal system 1 of this embodiment, the regression model construction processing unit 112 can also directly weight the regression model based on the weight criteria calculated in step S530. Specifically, the process that the regression model construction processing unit 112 executes in step S540 based on the results of step S530 is one of the following: setting a loss function, which is a function that serves as an indicator for learning; oversampling, which amplifies rare or highly important measured characteristic data and adds it as training data; or undersampling, which removes redundant or less important measured characteristic data from the training data (details will be described later). By executing these processes, even if the number of training data used to construct the machine learning model is small or if there is a bias in the distribution of the training data, the weights of the training data are appropriately adjusted accordingly. As a result, estimation accuracy is improved, and prototype conditions for good materials can be proposed with accuracy. When the processing in step S540 is completed, the control unit 11 proceeds to step S550.

[0049] In step S550, the control unit 11 uses the regression model construction processing unit 112 to search for and set the optimal hyperparameters for each regression model of the method selected as a candidate in step S520. In the prototype condition proposal system 1 of this embodiment, for each regression model, the regression model construction processing unit 112 automatically searches for all parameters and automatically sets the ones that will best improve the generalization performance of the regression model when set as hyperparameters during the construction of the regression model. Once the processing in step S550 is complete, the control unit 11 proceeds to step S560.

[0050] In step S560, the control unit 11 performs the process of creating regression models for each method with optimal hyperparameters set by the regression model construction processing unit 112. After creating regression models for each method, the regression model construction processing unit 112 selects the regression model with the highest generalization performance from all the created regression models to determine the final regression model. Once the processing in step S560 is complete, the control unit 11 terminates the regression model construction process shown in the flowchart of Figure 5.

[0051] As mentioned above, the weighting criteria are calculated based on the difference between the dependent variable included in the measured properties data and the target properties representing the target values ​​of the material properties, or on a statistical measure that represents the rarity of the explanatory variables included in the measured properties data.

[0052] Of these, the specific details of the process performed in step S530 when calculating the weighting criteria based on the difference between the objective variable included in the measured characteristic data and the target characteristic representing the target value of the material's properties are shown below as steps S532 to S534.

[0053] In step S532, the control unit 11 uses the regression model construction processing unit 112 to calculate the difference between the objective variable included in the measured characteristic data and the target characteristic representing the target value of the material's properties. Once the processing in step S532 is complete, the control unit 11 proceeds to step S534.

[0054] In step S534, the control unit 11 determines the weighting criteria using a function of the difference between the dependent variable and the target characteristic calculated in step S532, as determined by the regression model construction processing unit 112. Specific examples of this function include the absolute value or squared value of the difference between the dependent variable and the target characteristic, which represents the distance between the dependent variable and the target characteristic. The reason for determining the weighting criteria based on the distance between the dependent variable and the target characteristic is that, since regression problems are frequently solved in materials development, the frequency of occurrence of the dependent variable, which is commonly used when determining weighting criteria in classification problems, cannot be used. Therefore, in the prototype condition proposal system 1 of this embodiment, a weighting criterion based on the dependent variable, which is a continuous variable, is used. Once the processing in step S534 is complete, the control unit 11 terminates the process of calculating the weighting criteria shown in step S530 and proceeds to step S540.

[0055] On the other hand, when calculating weight criteria based on statistics representing the scarcity of explanatory variables included in the measured characteristic data, the specific details of the process performed in step S530 are shown below as steps S536 to S538.

[0056] In step S536, the control unit 11 uses the regression model construction processing unit 112 to calculate the number of occurrences of explanatory variables that take a specific value or fall within a specific range. Specifically, for example, if the explanatory variable relates to the raw material ratio of various raw materials, the control unit 11 calculates the number of occurrences of explanatory variables that take a value greater than 0, in other words, the number of occurrences of the explanatory variable that represents the raw material ratio actually used. Once the processing in step S536 is complete, the control unit 11 proceeds to step S538.

[0057] In step S538, the control unit 11 determines the weight criteria using a function of the frequency of occurrence of the explanatory variable calculated in step S536, with the regression model construction processing unit 112. A specific example of this function is the probability of occurrence of the explanatory variable in the entire training data. Alternatively, the regression model construction processing unit 112 may determine the frequency of occurrence of the explanatory variable as the weight criterion instead of the probability of occurrence. The reason for using a criterion based on explanatory variables is to incorporate the materials science knowledge of the user, who is a materials developer, especially regarding important parameters that do not appear in the objective variable. Once the processing in step S538 is complete, the control unit 11 terminates the process of calculating the weight criteria shown in step S530 and proceeds to step S540.

[0058] Furthermore, as described above, in step S540, the regression model construction processing unit 112 performs one of the following processes to weight the measured characteristic data or the regression model: setting a loss function, oversampling to amplify rare or important measured characteristic data and add it as training data, or undersampling to remove redundant or unimportant measured characteristic data from the training data. In other words, the regression model construction process further includes one of the following processes, which is performed based on the calculated weight criteria: setting a loss function, oversampling to amplify rare or important measured characteristic data and add it as training data, or undersampling to remove redundant or unimportant measured characteristic data from the training data.

[0059] Of these, the specific details of the process executed in step S530 when setting the loss function based on the weight criteria calculated in step S530 will be explained below as step S542.

[0060] In step S542, the control unit 11, using the regression model construction processing unit 112, determines machine learning based on the weight criteria calculated in step S530 to minimize the prediction error of rare or highly important measured characteristic data. This process is performed by weighting the measured characteristic data with the loss function used during training. Once the processing in step S538 is complete, the control unit 11 finishes the process of reflecting the weight criteria in the training data or regression model, as shown in step S540, and proceeds to step S550.

[0061] Furthermore, the specific details of the process called oversampling or upsampling, which is performed in step S530 when rare or important measured characteristic data is amplified based on the weight criteria calculated in step S530 and added as training data, will be explained below as step S544.

[0062] In step S544, the control unit 11, using the regression model construction processing unit 112, duplicates rare or highly important measured characteristic data, which are subject to oversampling, based on the weight criteria calculated in step S530, and adds them to the training data. In this case, the measured characteristic data is duplicated until its number exceeds a predetermined threshold. In addition to adding the duplicated measured characteristic data to the training data, oversampling processing may also be performed using various methods such as SMOTE, ADASYN, Borderline-SMOTE, and Safe-level SMOTE. This improves the balance of the training data because rare or highly important measured characteristic data is added to the training data. Figure 6 shows a comparison of the estimation accuracy before and after the weighting processing by oversampling performed in step S544. Figure 6 shows the measured characteristic data and its predicted value for bending strength, one of the characteristics of a ceramic composite material made from several raw materials, on a scatter plot. Figure 6 shows that, for a specific raw material whose estimation accuracy was insufficient due to the low number of times it was used in the overall measured data, the oversampling process performed in step S544 applied weighting to the measured characteristic data, resulting in a reduction of the error between the estimated and measured values ​​for this composite material. In other words, the estimation accuracy for this composite material was improved. Once the processing in step S544 is complete, the control unit 11 finishes the process of reflecting the weight criteria in the training data or regression model, as shown in step S540, and proceeds to step S550.

[0063] Furthermore, the specific details of the process called undersampling or downsampling, which is performed in step S540 when redundant or less important measured characteristic data is removed from the training data based on the weight criteria calculated in step S530, will be explained below as step S546.

[0064] In step S546, the control unit 11 uses the regression model construction processing unit 112 to remove redundant or low-importance measured characteristic data from the training data. This improves the balance of the training data by removing redundant or low-importance measured characteristic data. Once the processing in step S546 is complete, the control unit 11 finishes the process of reflecting the weight criteria in the training data or regression model, as shown in step S540, and proceeds to step S550.

[0065] Figure 7 is a flowchart detailing the process for proposing prototype conditions.

[0066] In step S710, the control unit 11 uses the prototype condition proposal processing unit 113 to perform a search for prototype conditions based on the regression model constructed in step S560 of Figure 5. The prototype condition proposal processing unit 113 performs this process using optimization. The prototype condition proposal system 1 of this embodiment is configured to use various optimization methods such as mathematical optimization (MO), Bayesian optimization (BO), genetic algorithm (GA), Newton's method (NM), and simplex method (SM). As a result, each explanatory variable that yields the best predicted value of the target variable (characteristic) is proposed to the user as a provisional prototype condition representing the result of the search process. At this time, the prototype condition proposal processing unit 113 also performs a sensitivity analysis on the provisional prototype condition to evaluate the importance of each explanatory variable constituting the provisional prototype condition and presents the evaluation results. In addition, in the prototype condition suggestion system 1 of this embodiment, if the regression model used is a Gaussian process regression, it is also possible to select the prototype conditions that maximize the acquisition function. When the processing in step S710 is completed, the control unit 11 proceeds to step S720.

[0067] In step S720, the control unit 11 receives a request from the user for modification of the provisional prototype conditions proposed to the user in step S710, via the prototype condition proposal processing unit 113. The prototype condition proposal processing unit 113 receives an input operation from the user via the input unit 131 or the communication unit 14 regarding the modification of the values ​​of each explanatory variable constituting the provisional prototype conditions, and modifies the provisional prototype conditions according to the modification. At this time, the control unit 11 uses the prototype condition proposal processing unit 113 to obtain predicted values ​​of the material properties when prototypes are made with the modified provisional prototype conditions using a regression model, and presents the calculation results to the user. Also at this time, similar to step S710, the prototype condition proposal processing unit 113 performs a sensitivity analysis on the modified provisional prototype conditions to evaluate the importance of the explanatory variables constituting the modified provisional prototype conditions, and presents the evaluation results as well. This evaluation result is updated each time the user modifies the provisional prototype conditions, so that the latest evaluation results are always presented to the user. When the processing in step S720 is completed, the control unit 11 proceeds to step S730.

[0068] In step S730, the control unit 11, using the prototype condition proposal processing unit 113, determines whether the predicted values ​​of the characteristics obtained in step S720 are insufficient as characteristic values ​​for the material to be prototyped for the modified provisional prototype conditions. This determination is made, for example, if the regression model used is a Gaussian process regression, by obtaining an acquisition function for each prototype condition that represents the expected improvement in the material's characteristics when prototyped under that condition, and determining whether the difference between the value of the acquisition function and the maximum value of the acquisition function is within a predetermined range. The acquisition function is calculated based on the predicted value μ of the material's characteristics when prototyped under any given prototype condition and the standard deviation σ that represents the variability of the predicted value. If it is determined that the predicted characteristic values ​​are insufficient, the process returns to step S720 to receive instructions from the user to modify the prototype conditions again. If it is determined that the predicted characteristic values ​​are not insufficient, the predicted values ​​of the material properties when prototyping with the modified provisional prototype conditions are sufficient, so the provisional prototype conditions are finalized and proposed to the user as finalized prototype conditions. In other words, this determination process is repeated until it is determined that the predicted values ​​of the material properties related to the provisional prototype conditions are not insufficient. Once the processing in step S730 is completed, the control unit 11 terminates the prototype condition proposal process shown in the flowchart of Figure 7.

[0069] As mentioned above, in this embodiment, when a user prototypes a material based on the prototype conditions proposed in step S730 of Figure 7, and data representing the measured properties of the resulting material is input as new measured property data, the prototype condition proposal system 1 of this embodiment proposes more optimized prototype conditions to the user based on the newly input measured property data. In this case, the control unit 11 of the prototype condition proposal system 1 determines whether or not the newly input measured property data contains missing values. If it is determined that there are missing values, the system performs preprocessing of the newly input measured property data to fill in these missing values, starting from step S440 of Figure 4. On the other hand, if it is determined that there are no missing values, preprocessing of the newly input measured property data is not necessary. In such cases, the control unit 11 determines whether or not it is necessary to update the regression model. If it is determined that the regression model needs to be updated, the control unit 11 starts processing from step S510 in Figure 5 to construct a new regression model and executes the regression model construction process for the newly input characteristic measurement data. On the other hand, if it is determined that the regression model does not need to be updated, the control unit 11 omits the regression model construction process and the regression model evaluation process, and uses the previously constructed regression model to perform the prototype condition suggestion process for the newly input characteristic measurement data, starting processing from step S710 in Figure 7. Even if the newly input characteristic measurement data contains missing values, if the regression model does not need to be updated, the control unit 11 similarly omits the regression model construction process and the regression model evaluation process.

[0070] According to the embodiments of the present invention described above, the following effects and advantages are achieved.

[0071] (1) The prototype condition proposal system 1 is a system that proposes prototype conditions for materials to material developers, and comprises a regression model construction processing unit 112 and a prototype condition proposal processing unit 113. The regression model construction processing unit 112 performs regression model construction processing on the measured characteristic data representing the measured results of the material's properties (step S320). The prototype condition proposal processing unit 113 searches for the optimal prototype conditions for the material using the constructed regression model and performs prototype condition proposal processing based on the search results (step S340). The regression model construction processing (Figure 5) includes a process to calculate a weighting standard, which is the basis for weighting the measured characteristic data (step S530), and a process to weight the measured characteristic data based on the calculated weighting standard (step S540). In this way, even if the number of training data used to construct the machine learning model is small or if there is a bias in the distribution of the training data, the weights of the training data are appropriately adjusted accordingly. As a result, estimation accuracy is improved, and good prototype conditions for materials can be proposed with accuracy.

[0072] (2) The weighting criteria are calculated based on the difference between the objective variable included in the measured characteristic data and the target characteristic which represents the target value of the material's properties (steps S532-S534). In this way, even in material development where regression problems are often solved and the weighting criteria cannot be determined based on the frequency of occurrence of the objective variable, the weighting criteria can be determined based on the objective variable.

[0073] (3) The weighting criteria are calculated based on statistics that represent the rarity of the explanatory variables included in the measured characteristic data (steps S536-S538). In this way, the weighting criteria can be calculated based on the explanatory variables. As a result, the materials science knowledge of the user, who is a materials developer, can be incorporated, especially for important parameters that do not appear in the objective variable.

[0074] (4) The regression model construction process (Figure 5) further includes a process (step S542) to set a loss function based on the calculated weight criteria. In this way, the prediction error can be reduced for rare or highly important measured characteristic data.

[0075] (5) The regression model construction process (Figure 5) further includes an oversampling process (step S544) in which rare or important observed characteristic data are amplified based on the calculated weight criteria and added as training data. In this way, rare or important observed characteristic data are added to the training data. As a result, the sensitivity of rare or important observed characteristic data can be improved, thus improving the balance of the training data.

[0076] (6) The regression model building process (Figure 5) further includes an undersampling process (step S546) which removes redundant or less important observed characteristic data from the training data based on the calculated weight criteria. In this way, redundant or less important observed characteristic data is removed from the training data. As a result, the sensitivity of redundant or less important observed characteristic data can be reduced, thereby improving the balance of the training data.

[0077] It should be noted that the present invention is not limited to the embodiments described above, and can be implemented using any components without departing from the spirit of the invention.

[0078] The embodiments and modifications described above are merely examples, and the present invention is not limited to these, as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these. Other embodiments conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention. [Explanation of symbols]

[0079] 1: Prototype Condition Proposal System 11: Control Unit 12: Storage section 13: User Interface Section 14: Communications Department 111: Preprocessing unit for measured characteristic data 112: Feature Selection Processing Unit 113: Regression Model Construction Processing Unit 114: Prototype Condition Proposal Processing Unit 131: Input section 132: Output section 400: Internet

Claims

1. A prototype condition proposal system that proposes prototype conditions for materials to material developers, A regression model construction processing unit that performs a regression model construction process on the measured characteristic data representing the measured characteristics of the aforementioned material, A prototype condition proposal processing unit searches for the optimal prototype conditions for the material using the constructed regression model and executes a prototype condition proposal process based on the search results. Equipped with, The regression model construction process described above is: A process for calculating a weighting criterion for the measured characteristics data, which corresponds to the distance between the objective variable and the target characteristic, based on the difference between the objective variable included in the measured characteristics data and the target characteristic representing the target value of the material's properties. A process to weight the measured characteristic data based on the weighting criteria calculated above. A prototype condition suggestion system, including the above.

2. In the prototype condition proposal system described in claim 1, The regression model construction process further includes a process for setting a loss function based on the calculated weight criteria, comprising a prototype condition proposal system.

3. In the prototype condition proposal system described in claim 1, The regression model construction process further includes an oversampling process that amplifies rare or highly important measured characteristic data based on the calculated weight criteria and adds it as training data, thereby providing a prototype condition proposal system.

4. In the prototype condition proposal system described in claim 1, The regression model construction process further includes an undersampling process that removes redundant or less important measured characteristic data from the training data based on the calculated weight criteria, in a prototype condition proposal system.

5. A method for proposing prototype conditions, which uses a computer to propose prototype conditions for materials to material developers, A regression model construction process is performed to construct a regression model for the measured property data representing the measured properties of the aforementioned material, Using the constructed regression model, the optimal prototype conditions for the material are searched, and a prototype condition proposal process is performed to propose prototype conditions for the material based on the search results. Have the computer run it, The regression model construction process described above is: A process for calculating a weighting criterion for the measured characteristics data, which corresponds to the distance between the objective variable and the target characteristic, based on the difference between the objective variable included in the measured characteristics data and the target characteristic representing the target value of the material's properties. A process to weight the measured characteristic data based on the weighting criteria calculated above. A method for proposing prototype conditions, including the above.

Images