Physical property prediction model generation device, physical property prediction model generation method, and physical property prediction model generation program

Abstract

This physical property prediction model generation device comprises: an actual measurement data acquisition unit that acquires actual measurement data including a manufacturing condition of a sample including a qualitative variable and an actual measurement value obtained by actually measuring the physical property of the sample manufactured on the basis of the manufacturing condition; a similarity acquisition unit that acquires similarity between a plurality of levels included in the qualitative variable; a digitization unit that digitizes, on the basis of the similarity acquired by the similarity acquisition unit, the qualitative variable in the actual measurement data acquired by the actual measurement data acquisition unit; and a model generation unit that generates a prediction model of the physical property on the basis of the actual measurement data after the qualitative variable is digitized by the digitization unit.

Description

Apparatus for generating material property prediction models, method for generating material property prediction models, and program for generating material property prediction models

[0001] This disclosure relates to a physical property prediction model generation apparatus, a physical property prediction model generation method, and a physical property prediction model generation program.

[0002] In recent years, a field of technology called materials informatics has been attracting attention, which utilizes information science techniques such as machine learning to improve the efficiency of materials development. In materials informatics, it is necessary to find the appropriate combination of numerous parameters related to the mixing ratio of raw materials and process conditions in order to manufacture materials with desired physical properties. When there are many parameters, the number of parameter combinations becomes enormous. When the number of parameter combinations becomes enormous, it becomes difficult to comprehensively prototype materials. When it is difficult to comprehensively prototype materials, a method is adopted in which the physical properties of a certain number of prototyped materials are measured, and a predictive model of the physical properties is constructed based on the measured data obtained from the measurements. By using a predictive model, it is possible to search for the optimal parameters even if the number of prototyped materials is limited.

[0003] Supervised machine learning methods are sometimes used to build predictive models. However, with supervised machine learning methods, it becomes difficult to build highly accurate predictive models when the amount of measured data is insufficient or when the relationship between parameters and physical properties is complex. To build highly accurate predictive models even when the amount of measured data is insufficient or the relationship between parameters and physical properties is complex, sequential experimental design methods such as Bayesian optimization are used. In Bayesian optimization, Bayesian estimation is used to build predictive models. In Bayesian estimation, in addition to the predicted values ​​of physical properties, the uncertainty of the predicted values ​​is calculated. By calculating the uncertainty, it is possible to calculate the probability that the physical properties will be at a specific value and the expected value of the physical properties. According to Bayesian estimation, the optimal parameters can be found based on limited measured data. In other words, according to Bayesian estimation, an experimental plan for the next prototype of a material can be made based on limited measured data. The predictive model is updated by making the next prototype of the material, measuring the physical properties of the prototyped material, and adding the measurement results to the measured data. In Bayesian optimization, this series of cycles consisting of prototyping, measuring physical properties, and updating the predictive model is repeated. By repeating this series of cycles, it is possible to minimize the number of prototypes required to obtain the desired physical properties.

[0004] However, the parameters that can be optimized by Bayesian optimization are limited to numerically quantifiable variables. Specifically, Bayesian optimization can find the optimal values ​​of quantitative variables that can be expressed numerically, such as the mixing ratio of raw materials. On the other hand, Bayesian optimization has difficulty finding the optimal values ​​of qualitative variables that cannot be expressed numerically, such as selecting the best raw material from a list of candidates. Optimization of qualitative variables has been limited to situations where the number of levels of the qualitative variable is small and the qualitative variable can be quantified using a simple method. Therefore, conventionally, it has been difficult to appropriately and efficiently develop materials with desired physical properties while considering qualitative variables.

[0005] WO2021 / 200281A1 discloses a sequential experimental design that applies machine learning to measured data. WO2020 / 145252A1 discloses a method for applying Gaussian process regression when material properties are obtained as interval values.

[0006] However, neither WO2021 / 200281A1 nor WO2020 / 145252A1 addresses cases where the parameters include qualitative variables that are difficult to quantify.

[0007] This disclosure has been made in consideration of the above points, and aims to provide a material property prediction model generation apparatus, a material property prediction model generation method, and a material property prediction model generation program that can appropriately and efficiently develop materials having desired physical properties.

[0008] Embodiments of the present disclosure relate to the following [1] to

[22] : [1] A physical property prediction model generation apparatus comprising: an actual measurement data acquisition unit that acquires actual measurement data including manufacturing conditions for a sample including qualitative variables and actual measurement values ​​of the physical properties of the sample manufactured based on the manufacturing conditions; a similarity acquisition unit that acquires similarity between a plurality of levels included in the qualitative variables; a digitization unit that digitizes the qualitative variables from the actual measurement data acquired by the actual measurement data acquisition unit based on the similarity acquired by the similarity acquisition unit; and a model generation unit that generates a prediction model of the physical properties based on the actual measurement data after the qualitative variables have been digitized by the digitization unit.

[0009] [2] The physical property prediction model generation apparatus further comprises a similarity setting unit for setting the similarity, and the similarity acquisition unit acquires the similarity set by the similarity setting unit, as described in [1].

[0010] [3] The physical property prediction model generation apparatus according to [2], wherein the similarity setting unit sets the similarity by at least one of inputting or modifying the similarity on an input screen, inputting a file in which the similarity is described, and reusing a similarity that has been input in the past, or by automatic setting based on a predetermined rule.

[0011] [4] The physical property prediction model generation apparatus further comprises a display unit that displays an input screen for receiving input operations for inputting or modifying the similarity, and an input interface for receiving the input operations for the input screen, wherein the similarity setting unit sets the similarity based on the input operations received via the input interface for the input screen displayed on the display unit, the physical property prediction model generation apparatus according to [2].

[0012] [5] A physical property prediction model generation apparatus according to any one of [1] to [4], further comprising a conversion table creation unit that creates a conversion table for quantifying the qualitative variables based on the similarity obtained by the similarity acquisition unit, wherein the quantification unit quantifies the qualitative variables using the conversion table created by the conversion table creation unit.

[0013] [6] The physical property prediction model generation apparatus according to [5], wherein the conversion table creation unit creates the conversion table based on multidimensional scaling.

[0014] [7] The physical property prediction model generation apparatus according to any one of [1] to [6], wherein the qualitative variable is expressed on a nominal scale and has three or more levels.

[0015] [8] The physical property prediction model generation apparatus according to [7], wherein the qualitative variables include at least one of at least three types of materials usable in the production of the sample, at least three types of apparatus usable in the production of the sample, and at least three types of methods usable in the production of the sample.

[0016] [9] The physical property prediction model generation apparatus according to any one of [1] to [8], wherein the model generation unit generates the prediction model using Bayesian estimation.

[0017]

[10] The physical property prediction model generation apparatus according to any one of [1] to [9], wherein the model generation unit generates the prediction model according to the following formula.

[0018]

[11] The physical property prediction model generation apparatus according to any one of [1] to

[10] , further comprising a manufacturing condition calculation unit that calculates second manufacturing conditions for the sample based on the prediction model generated by the model generation unit.

[0019]

[12] The physical property prediction model generating apparatus according to

[11] , wherein the second manufacturing conditions include manufacturing conditions when the physical properties in the prediction model fall within a preset target range.

[0020]

[13] The physical property prediction model generation apparatus according to

[12] , wherein the target range is set by upper and lower limits entered by the user.

[0021]

[14] The physical property prediction model generation apparatus according to

[12] , wherein the target range is set based on the specification values ​​stored in the memory unit.

[0022]

[15] The physical property prediction model generation apparatus according to any one of

[12] to

[14] , wherein the measurement data acquisition unit further acquires second measurement data including the second manufacturing conditions calculated by the manufacturing condition calculation unit and second measurement values ​​obtained by actually measuring the physical properties of the sample manufactured based on the second manufacturing conditions, and the model generation unit updates the prediction model based on the second measurement data acquired by the measurement data acquisition unit.

[0023]

[16] The physical property prediction model generation apparatus according to

[15] , wherein the model generation unit determines whether the second measured value is a set optimal value by comparing the second measured value with the optimal value, and if the second measured value is not the optimal value, updates the prediction model based on the second measured data.

[0024]

[17] A method for generating a physical property prediction model performed by a computer, comprising: a step of acquiring measured data including manufacturing conditions for a sample that includes qualitative variables and measured values ​​of the physical properties of the sample manufactured based on the manufacturing conditions; a step of acquiring the similarity between a plurality of levels included in the qualitative variables; a step of quantifying the qualitative variables from the acquired measured data based on the acquired similarity; and a step of generating a physical property prediction model based on the measured data after the qualitative variables have been quantified.

[0025]

[18] The method for generating a physical property prediction model according to

[17] , further comprising the step of setting the similarity, wherein the step of obtaining the similarity is performed by obtaining the set similarity.

[0026]

[19] A method for generating a physical property prediction model according to

[17] or

[18] , further comprising the step of creating a conversion table for quantifying the qualitative variables based on the obtained similarity, wherein the quantification step is performed using the created conversion table.

[0027]

[20] A property prediction model generation program for causing a computer to perform the following steps: a procedure for acquiring measured data including manufacturing conditions for a sample containing qualitative variables and measured values ​​of the physical properties of the sample manufactured based on the manufacturing conditions; a procedure for acquiring the similarity between a plurality of levels included in the qualitative variables; a procedure for quantifying the qualitative variables from the acquired measured data based on the acquired similarity; and a procedure for generating a predictive model of the physical properties based on the measured data after the qualitative variables have been quantified.

[0028]

[21] The physical property prediction model generation program according to

[20] , wherein the computer further performs the procedure for setting the similarity, and the procedure for obtaining the similarity is performed by obtaining the set similarity.

[0029]

[22] The physical property prediction model generation program according to

[20] or

[21] , wherein the computer is further made to perform a procedure to create a conversion table for quantifying the qualitative variables based on the acquired similarity, and the procedure for quantification is performed using the created conversion table.

[0030] According to this disclosure, materials having desired physical properties can be developed appropriately and efficiently.

[0031] Figure 1 is a schematic diagram showing a physical property prediction model generation device according to the first embodiment. Figure 2 is a block diagram showing the physical property prediction model generation device according to the first embodiment. Figure 3 is a flowchart showing an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 4 is a diagram showing the display process of the input screen in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 5 is a diagram showing the input process of similarity data in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 6 is a diagram showing the creation process of the conversion table in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 7 is a diagram showing the input process of measured data in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 8 is a diagram showing the quantification process of categorical variables in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 9A is a diagram showing the prediction model in the case of raw material A in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 9B is a diagram showing the prediction model in the case of raw material B in an example of operation of the physical property prediction model generation device according to the first embodiment. Figure 9C shows the predicted model for raw material C in an example of operation of the physical property prediction model generation apparatus according to the first embodiment. Figure 9D shows the predicted model for raw material D in an example of operation of the physical property prediction model generation apparatus according to the first embodiment. Figure 10 is a block diagram of the physical property prediction model generation apparatus according to the second embodiment.

[0032] Embodiments of the present disclosure will now be described with reference to the drawings. In the drawings referenced in the embodiments, the same or similar reference numerals are used to denote identical parts or parts having similar functions, and repeated descriptions may be omitted.

[0033] [First Embodiment] First, a physical property prediction model generation apparatus according to the first embodiment will be described with reference to Figures 1 to 9D. Figure 1 is a conceptual diagram showing the physical property prediction model generation apparatus 1 according to the first embodiment.

[0034] As shown in Figure 1, the physical property prediction model generation apparatus 1 according to the first embodiment generates a prediction model of a sample based on measured data. The measured data includes the manufacturing conditions of the sample and the measured values. The sample is, for example, a material manufactured based on multiple raw materials. The manufacturing conditions of the sample include quantitative variables expressed numerically and qualitative variables that are not expressed numerically. In this specification, measured data includes (i) manufacturing conditions (including qualitative variables) and (ii) measured values, and is acquired by input.

[0035] Quantitative variables include, for example, the mixing ratio of raw materials and temperature. Qualitative variables can be expressed as either ordinal scales (which can be ordered) or nominal scales (which cannot be ordered). Examples of qualitative variables expressed as ordinal scales include "small," "medium," and "large," which can be ordered by size. Examples of qualitative variables expressed as nominal scales include raw materials, equipment, and methods used in the manufacture and synthesis of materials. In this specification, qualitative variables expressed as nominal scales with three or more levels are called categorical variables. A level is at least one subdivided type of variable belonging to a single type of variable. For example, a raw material, which is one of the qualitative variables, may have levels such as "raw material A," "raw material B," and "raw material C."

[0036] Measured values ​​are the values ​​obtained by actually measuring the physical properties of a sample manufactured under manufacturing conditions that include qualitative variables. Examples of sample properties include conductivity and dielectric constant. A predictive model is information that shows the predicted physical properties of a sample based on its manufacturing conditions. In other words, a predictive model shows the correspondence between the manufacturing conditions and the physical properties of a sample.

[0037] Further, the physical property prediction model generation device 1 calculates new production conditions for a sample based on the generated prediction model, and creates a new experimental plan including the relevant production conditions as necessary. A new sample is prototyped in accordance with the created new experimental plan. An experiment for actually measuring the physical property values is performed on the prototyped new sample. Actual measurement data obtained by the experiment is added to the physical property prediction model generation device 1. The physical property prediction model generation device 1 updates the prediction model based on the actual measurement data updated by the addition.

[0038] The physical property prediction model generation device 1 optimizes the experimental plan by repeating the generation of the prediction model based on the actual measurement data, the creation of the experimental plan, and the addition of the actual measurement data.

[0039] FIG. 2 is a block diagram illustrating the physical property prediction model generation device 1 according to the first embodiment. More specifically, as shown in FIG. 2, the physical property prediction model generation device 1 includes a control unit 2, an input interface 3, a display unit 4, and a storage unit 5. The control unit 2, the input interface 3, the display unit 4, and the storage unit 5 are communicably connected via a bus 6.

[0040] The input interface 3 receives input operations for various instructions and information from a user. The input interface 3 converts the input operations received from the user into electrical signals. The input interface 3 outputs the input operations converted into electrical signals to the control unit 2. The input interface 3 may receive an input operation on an input screen for inputting or correcting the similarity between a plurality of levels included in a qualitative variable to be described later. The input interface 3 may be, for example, a mouse, a keyboard, a touchpad, a touch panel, a trackball, a switch button, a microphone, or the like.

[0041] The display unit 4 displays an image corresponding to the image data transmitted from the control unit 2. The display unit 4 may display an input screen that receives an input operation for inputting or correcting the similarity between a plurality of levels included in a qualitative variable to be described later. The display unit 4 may be, for example, a liquid crystal monitor, a CRT (Cathode Ray Tube) monitor, a touch panel, or the like. The display unit 4 may be provided with a speaker.

[0042] The storage unit 5 is a non-transitory storage device that stores various types of information. The storage unit 5 may be, for example, an HDD (Hard Disk Drive), an optical disk, a magnetic disk, a magneto-optical disk, a CD (Compact Disc), a DVD (Digital Versatile Disc), a semiconductor memory, or the like. The storage unit 5 stores, for example, a physical property prediction model generation program executed by the physical property prediction model generation apparatus 1, and various types of data used for executing the physical property prediction model generation program.

[0043] The control unit 2 controls the operation of the physical property prediction model generation apparatus 1 in response to a user's input operation using the input interface 3. The control unit 2 includes a measured data acquisition unit 21, a similarity acquisition unit 27, a quantification unit 22, a model generation unit 23, a similarity setting unit 24, a conversion table creation unit 25, and a manufacturing condition calculation unit 26.

[0044] The functions corresponding to each of the components 21 to 27 of the control unit 2 are recorded in the storage unit 5 in the form of a physical property prediction model generation program that can be executed by a computer. The control unit 2 is, for example, a processor. The processor constituting the control unit 2 reads and executes the physical property prediction model generation program from the storage unit 5, thereby executing the functions of the components 21 to 27 corresponding to the read physical property prediction model generation program. Alternatively, instead of storing the physical property prediction model generation program in the storage unit 5, the physical property prediction model generation program may be directly incorporated into the processor's circuitry. In this case, the processor executes the physical property prediction model generation program incorporated into the circuitry, thereby executing the functions corresponding to each of the components 21 to 27. The physical property prediction model generation program may be a program downloaded by the physical property prediction model generation device 1 from a server via a network. Alternatively, the physical property prediction model generation program may be a program imported into the physical property prediction model generation device 1 from a portable storage medium. The control unit 2 may consist of a single processor, or it may consist of a combination of multiple processors. The processor may be, for example, a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an Application Specific Integrated Circuit (ASIC), or a programmable logic device. The programmable logic device may be an SPLD (Simple Programmable Logic Device), a CPLD (Complex Programmable Logic Device), or an FPGA (Field Programmable Gate Array).

[0045] The measurement data acquisition unit 21 acquires measurement data that includes the manufacturing conditions of a sample, which include qualitative variables, and the measured values ​​of the physical properties of the sample manufactured based on those manufacturing conditions. The qualitative variables may include qualitative variables expressed on a nominal scale with three or more levels, i.e., categorical variables.

[0046] The similarity acquisition unit 27 acquires the similarity between multiple levels included in the qualitative variable.

[0047] The quantification unit 22 quantifies the qualitative variables among the measured data acquired by the measured data acquisition unit 21 based on the similarity obtained by the similarity acquisition unit 27.

[0048] The model generation unit 23 generates a predictive model of physical properties based on the measured data after the qualitative variables have been quantified by the quantification unit 22. The predictive model may be a probabilistic predictive model that shows the predicted physical properties in terms of probabilities. The model generation unit 23 may also generate a predictive model that has uncertainty about the physical properties with respect to the measured data.

[0049] In this specification, "similarity" refers to an index that quantitatively represents the closeness or relationship between multiple levels contained in a qualitative variable. Similarity is constructed as a value corresponding to each combination of levels and may be represented in matrix form. The matrix may include diagonal components representing the relationship between identical levels and off-diagonal components representing the relationship between different levels. Similarity may be constructed as a symmetric matrix on the premise that the relationships between two levels are equal, but is not limited to this. Furthermore, similarity may be set as a dissimilarity index, where a smaller value indicates closer levels, or as an index, where a larger value indicates closer levels. The range of values, scale, and method of representation of similarity are not particularly limited and may be set based on an ordinal scale, interval scale, or any arbitrary scale.

[0050] The similarity setting unit 24 sets the similarity between multiple levels included in a qualitative variable. The similarity setting unit 24 may set the similarity by presenting an input screen for inputting the similarity between multiple levels included in a qualitative variable. Specifically, the similarity setting unit 24 may generate an input screen that accepts input operations for entering similarity, and display the generated input screen on the display unit 4. When displaying an input screen, the similarity setting unit 24 sets the similarity entered by the user for the displayed input screen. That is, the similarity setting unit 24 sets the similarity based on the input operations received via the input interface 3 for the input screen displayed on the display unit 4. The user may input desired values ​​according to the user's judgment as the similarity between multiple levels for the input screen. Alternatively, the similarity setting unit 24 may obtain a similarity database showing the similarity between multiple levels from the storage unit 5 or an external network. The similarity setting unit 24 may display the obtained similarity database on the display unit 4. The user may refer to the displayed similarity database and select the similarity to be entered on the input screen.

[0051] As will be described later, the similarity score may be entered into the input screen as a value indicating the dissimilarity between two different levels.

[0052] Instead of inputting similarity into the input screen, or in addition to inputting similarity into the input screen, the similarity setting unit 24 may set the similarity by inputting a file containing similarity values ​​and reusing previously entered similarity values. The similarity setting unit 24 may input a file containing similarity values, for example, via communication over an external network or through a portable storage medium. The similarity setting unit 24 may, for example, read and use previously entered similarity values ​​from the storage unit 5.

[0053] Furthermore, the similarity setting unit 24 may automatically initialize the similarity between multiple levels included in the qualitative variable based on predetermined rules, either in lieu of or prior to user input. Examples of predetermined rules include calculating the distance or dissimilarity between levels based on the chemical structure, composition, physical properties, or feature quantities derived therefrom representing the material, and setting the similarity based on that distance or dissimilarity. The similarity setting unit 24 may also refer to an existing database of similar materials and estimate the similarity between levels based on the similarity relationship between the target level and previously registered materials. Additionally, the similarity setting unit 24 may calculate the similarity based on past measurement data, specifically the correlation of physical properties between levels or the closeness of their contributions in a prediction model. These automatically set similarities may be modified by the user, or they may be used as is by the digitization unit 22 without modification.

[0054] The similarity acquisition unit 27 acquires the similarity set by the similarity setting unit 24. In other words, the quantification unit 22 quantifies the qualitative variables based on the similarity set by the similarity setting unit 24. For example, the quantification unit 22 quantifies the qualitative variables based on the similarity entered using the input interface 3 to the input screen displayed on the display unit 4.

[0055] The conversion table creation unit 25 creates a conversion table that quantifies qualitative variables based on the similarity obtained by the similarity acquisition unit 27. In other words, the conversion table creation unit 25 creates a conversion table based on the similarity set by the similarity setting unit 24. For example, the conversion table creation unit 25 creates a conversion table based on the similarity entered on the input screen.

[0056] The quantification unit 22 quantifies the qualitative variables using the conversion table created by the conversion table creation unit 25.

[0057] The transformation table creation unit 25 may create a transformation table based on multidimensional scaling. Multidimensional scaling is a method that represents the similarity between multiple data as distance and depicts the relative positional relationships between the data based on distance in a map space.

[0058] The manufacturing condition calculation unit 26 calculates the second manufacturing conditions for the sample based on the predictive model generated by the model generation unit 23.

[0059] The second manufacturing conditions may include manufacturing conditions that bring the physical properties in the prediction model within the target range.

[0060] The target range may be set based on upper and lower limits entered by the user. Alternatively, the target range may be automatically set based on product specifications, design specifications, or quality standards pre-stored in the memory unit 5. Furthermore, the target range may be set as a range that satisfies predetermined probability conditions or evaluation indicators based on the distribution of physical properties, confidence intervals, or expected values ​​calculated by a predictive model.

[0061] The measurement data acquisition unit 21 further acquires second measurement data, which includes second manufacturing conditions calculated by the manufacturing condition calculation unit 26 and second measurement values ​​obtained by actually measuring the physical properties of a sample manufactured based on the second manufacturing conditions.

[0062] The model generation unit 23 updates the prediction model based on the second measured data acquired by the measured data acquisition unit 21. The model generation unit 23 determines whether the second measured value is the set optimal value by comparing the second measured value with the optimal value. If it is determined that the second measured value is not the optimal value, the model generation unit 23 updates the prediction model based on the second measured data.

[0063] The optimal value may be set as the value that maximizes or minimizes the expected value of the physical properties calculated by the prediction model within the target range. Alternatively, the optimal value may be set based on the condition that maximizes the probability of the physical properties falling within the target range, or on the condition that the variation in the physical properties is less than or equal to a predetermined value. Furthermore, the optimal value may be set based on an evaluation function or weighting condition specified by the user.

[0064] The model generation unit 23 may generate a prediction model, i.e., a probabilistic prediction model, using Bayesian estimation.

[0065] When generating a predictive model using Bayesian estimation, the model generation unit 23 may generate the predictive model according to the following equation.

[0066]

[0067]

[0068]

[0069]

[0070] Note that equation (1) is identical to the following equation, differing only in expression. That is, when the parameters, physical properties, basis functions, and error terms of the prediction model are expressed using vectors, equation (1) is identical to the following equation. In equation (3), θ is the column vector representation of the physical properties in the i-th measured data mentioned above. A is the row vector representation of the parameters of the prediction model mentioned above. ε is the column vector representation of the error (i.e., the error term) mentioned above.

[0071] Furthermore, when vectors are used, equation (2) is equivalent to the following equation.

[0072] [Operation Example] Next, an operation example of the physical property prediction model generation device 1 having the above configuration will be described. Figure 3 is a flowchart showing an operation example of the physical property prediction model generation device 1 according to the first embodiment. As a premise of Figure 3, the user creates measured data. Specifically, the user obtains measured values ​​by measuring the physical properties of a sample manufactured based on manufacturing conditions that include categorical variables. The user then creates measured data by associating the manufacturing conditions with the measured values. In creating measured data, the user may set one to five combinations of values ​​for multiple parameters that constitute the manufacturing conditions. The user then creates measured data by measuring the physical properties of a sample manufactured according to the set combination of parameter values. In this case, the user may randomly set the values ​​of multiple parameters of the manufacturing conditions within a predetermined range. Alternatively, the user may set the values ​​of multiple parameters of the manufacturing conditions using a method that uniformly selects the parameter values ​​within a predetermined range, such as Latin square sampling. Furthermore, in the example shown below, the similarity setting unit 24 sets the similarity entered on the input screen. As mentioned above, the similarity setting unit 24 may set the similarity using methods other than inputting data into the input screen.

[0073] After the measured data is created, as shown in Figure 3, the similarity setting unit 24 displays an input screen on the display unit 4 for inputting the similarity between multiple levels included in the categorical variable (step S1). Figure 4 is a diagram showing the display process of the input screen in an example of the operation of the physical property prediction model generation device 1 according to the first embodiment. In the example shown in Figure 4, the similarity setting unit 24 presents an input screen for inputting the similarity between multiple levels included in raw materials, which are one of the categorical variables. In the example shown in Figure 4, the multiple levels are raw material A, raw material B, raw material C, and raw material D.

[0074] In the example shown in Figure 4, the input screen allows input of k(k-1) / 2 positive values ​​for similarity, where k is the number of levels in the categorical variable. In the example shown in Figure 4, the number of levels k is 4, representing raw materials A through D. Therefore, the similarity can be entered as 6 positive values. In other words, in the example shown in Figure 4, the similarity can be entered in the 6 cells that are not shaded. Also, in the example shown in Figure 4, the higher the similarity between two different levels, the smaller the value that can be entered on the input screen. Conversely, the higher the dissimilarity between two different levels, the larger the value that can be entered on the input screen. Specifically, the input screen allows input of 5 levels of similarity. The highest similarity can be entered as the smallest value "1". The lowest similarity can be entered as the largest value "5". Categorical variables other than raw materials can also be entered on the input screen in the same manner as in Figure 4.

[0075] After the input screen is displayed, as shown in Figure 3, the user inputs similarity data indicating the similarity between multiple levels included in the categorical variable by performing input operations using the input interface 3 (step S2). Figure 5 is a diagram showing the input process of similarity data in an example of the operation of the physical property prediction model generation device 1 according to the first embodiment. In the example shown in Figure 5, five levels of similarity are entered on the input screen. In the example shown in Figure 5, the similarity between raw material A and raw material D is the highest, "1". Also, in the example shown in Figure 5, the similarity between raw material A and raw material B is the lowest, "5".

[0076] In response to the input of similarity data, the similarity acquisition unit 27 acquires the input similarity data. After the similarity data is acquired, as shown in Figure 3, the conversion table creation unit 25 creates a conversion table that quantifies the categorical variables based on the acquired similarity data (step S3). For example, the conversion table creation unit 25 creates the conversion table based on multidimensional scaling. Figure 6 is a diagram showing the conversion table creation process in an example of the operation of the physical property prediction model generation device 1 according to the first embodiment. In the example shown in Figure 6, for example, the similarity of raw material B to raw material A is quantified from the input value "5" to the first variable "0.308". Also, the similarity of raw material C to raw material A is quantified from the input value "2" to the second variable "1.556". Also, the similarity of raw material D to raw material A is quantified from the input value "1" to the third variable "0.036". In the example shown in Figure 6, 12 values ​​for the conversion table are calculated based on the 6 similarity values ​​entered in the input screen of Figure 5. If any one of the similarity values ​​changes, all 12 values ​​shown in the transformation table in Figure 6 will change. The transformation table is created, for example, based on classical multidimensional scaling.

[0077] In this embodiment, the "similarity" entered on the input screen is used as an index representing the closeness between two levels, and this index may be entered as a value representing the dissimilarity or distance between the two levels. For example, the input value may be set so that the higher the similarity between the levels, the smaller the value, in which case the input value is interpreted as a distance index where the higher the dissimilarity, the larger the value. In the following, the input value entered on the input screen, or the value after a monotonic transformation performed on the input value, is used as the dissimilarity d_ij between level i and level j, and the square dissimilarity matrix D = [d 2 It constitutes _ij].

[0078] Classical multidimensional scaling utilizes the fact that the coordinate matrix X, which is the transformation table, can be derived by eigenvalue decomposition. Specifically, first, the squared dissimilarity matrix D, which corresponds to the square of the dissimilarity, is derived: [d ２ ｉｊSet the squared dissimilarity D, then perform bicentering using the centering matrix C: B = -1 / 2・CDC. C is represented as I - 1 / n・Jn. I is the n×n identity matrix. Jn is an n×n matrix where all matrix elements are 1. After performing bicentering, determine the top m eigenvalues ​​λ1, λ2, ..., λm of the matrix B calculated by bicentering, and the eigenvectors e1, e2, ..., em corresponding to those eigenvalues. m is the number of dimensions required for the output. After determining the eigenvalues ​​and corresponding eigenvectors, calculate the coordinate matrix, i.e., the transformation table, using the formula: X = Em√Λm. Em is a matrix with e1, e2, ..., em as column vectors. √Λm is a diagonal matrix with √λ1, √λ2, ..., √λm as diagonal elements.

[0079] After the conversion table is created, as shown in Figure 3, the user inputs the measured data into the unshown input screen displayed on the display unit 4 by inputting using the input interface 3 (step S4). Figure 7 is a diagram showing the input process of measured data in an example of the operation of the physical property prediction model generation device 1 according to the first embodiment. In the example shown in Figure 7, the measured data includes categorical variables, raw materials A to C, and quantitative variables, blending amounts, as explanatory variables. Also, in the example shown in Figure 7, the measured data includes physical properties, i.e., measured values, associated with the raw materials and blending amounts as objective variables.

[0080] In response to the input of actual measurement data, the actual measurement data acquisition unit 21 acquires the input actual measurement data. After the actual measurement data is acquired, as shown in Figure 3, the digitization unit 22 uses the conversion table created by the conversion table creation unit 25 to digitize the categorical variables included in the input actual measurement data (step S5). In other words, the digitization unit 22 digitizes the categorical variables, which are qualitative variables, based on the similarity data acquired by the similarity acquisition unit 27. Figure 8 is a diagram showing the digitization process of categorical variables in an example of operation of the physical property prediction model generation device 1 according to the first embodiment. In the example shown in Figure 8, raw material A, which is one of the categorical variables, is converted to a set of numerical values ​​(0.308, 1.556, 0.036) shown as the first to third variables. Also in the example shown in Figure 8, raw material B, which is one of the categorical variables, is converted to a set of numerical values ​​(-1.932, 0.561, 0.067) shown as the first to third variables. Furthermore, in the example shown in Figure 8, one of the categorical variables, raw material C, is converted to a set of numerical values ​​(-0.680, -0.565, 0.326) shown in the first to third variables.

[0081] After the categorical variables are quantified, as shown in Figure 3, the model generation unit 23 generates a predictive model of the physical properties based on the measured data after the categorical variables have been quantified by the quantification unit 22 (step S6). In the example shown in Figure 3, the model generation unit 23 generates a predictive model using Bayesian estimation according to the above-described formulas (1) and (2). At this time, the model generation unit 23 calculates the variance σ of the likelihood corresponding to the measured data. ２ It is also possible to set this to 0. Furthermore, the model generation unit 23 may set the prior distribution p(A) to have a high probability density when the value of the physical property θ changes smoothly.

[0082] The model generation unit 23 may display the generated prediction model on the display unit 4.

[0083] Figure 9A shows the predicted model for raw material A in an example of operation of the physical property prediction model generation apparatus 1 according to the first embodiment. Figure 9B shows the predicted model for raw material B in an example of operation of the physical property prediction model generation apparatus 1 according to the first embodiment. Figure 9C shows the predicted model for raw material C in an example of operation of the physical property prediction model generation apparatus 1 according to the first embodiment. Figure 9D shows the predicted model for raw material D in an example of operation of the physical property prediction model generation apparatus 1 according to the first embodiment.

[0084] For example, the model generation unit 23 generates the prediction model shown in Figures 9A to 9D. The horizontal axis in Figures 9A to 9D represents the blending amount. The vertical axis in Figures 9A to 9D represents the predicted value of the physical property θ corresponding to the blending amount. The blending amount may be the blending ratio of the raw materials (e.g., mass%, parts by mass, molar ratio, etc.). The physical property θ may be a physical property value such as conductivity or dielectric constant, and may be a normalized (dimensionless) value if necessary. The solid line graph in Figures 9A to 9D shows the predicted average value of the physical property θ calculated by the prediction model. The measured data is plotted at the position of the blending amount corresponding to the predicted average value. The dashed and dotted line graphs in Figures 9A to 9D are confidence intervals showing the uncertainty of the predicted value of the physical property θ. Specifically, the dashed line graph in Figures 9A to 9D shows the upper and lower limits of the 50% confidence interval in the normal distribution, which is the probability distribution of the physical property, respectively. The dashed-dotted graphs in Figures 9A to 9D show the upper and lower limits of the 80% confidence interval for the normal distribution, which is the probability distribution of physical properties, respectively.

[0085] For example, the predictive model for raw material A shown in Figure 9A is generated by substituting the first to third variables (0.308, 1.556, 0.036), which are numerical representations of raw material A, as categorical variables from the measured data, into formula (1). According to the predictive model for raw material A, predicted values ​​of the physical properties for any given amount of raw material A, along with confidence intervals, can be obtained. Predictive models for raw materials other than raw material A can also be generated by substituting the first to third variables corresponding to each raw material into formula (1).

[0086] As shown in Figure 9A, raw material A has a relatively large number of measured data points, four in total. Due to the large number of measured data points, the predictive model for raw material A has a narrow confidence interval and low uncertainty.

[0087] On the other hand, as shown in Figures 9B and 9C, raw materials C and D have a limited amount of measured data. However, according to the first embodiment, in the case of raw material C, even though there is only one measured data point, a predictive model can be constructed using the measured data of raw material A, which has a relatively high similarity to raw material C. Similarly, in the case of raw material D, even though there is no measured data, a predictive model can be obtained that estimates the general trend, albeit with a wide confidence interval, by using the measured data of raw material A, which has a high similarity to raw material D.

[0088] After the predictive model is generated, an experimental plan is created as shown in Figure 3 (step S7). In creating the experimental plan, first, the manufacturing condition calculation unit 26 calculates second manufacturing conditions for the sample based on the predictive model generated by the model generation unit 23. For example, when determining raw materials and their proportions such that the physical property value is 90 or higher, in the examples shown in Figures 9A to 9D, using raw material C in proportion to 100 is calculated as the second manufacturing condition. Next, in creating the experimental plan, the user creates a specific plan for manufacturing the sample using the calculated second manufacturing conditions.

[0089] After the experimental plan is created, the user conducts the experiment (step S8), as shown in Figure 3. In conducting the experiment, the user first manufactures a sample according to the experimental plan, i.e., the second manufacturing conditions. The sample is manufactured using a manufacturing apparatus (not shown). After the sample is manufactured, the user measures the physical properties of the manufactured sample. The physical properties of the sample are measured using measuring instruments (not shown) appropriate to the physical properties.

[0090] If the desired physical properties are obtained as a result of the experiment (Step S9: Yes), the process is terminated. On the other hand, if the desired physical properties are not obtained as a result of the experiment (Step S9: No), and additional experiments are possible (Step S10: Yes), the user inputs the measured data obtained from the experiment into the input screen (Step S4). Upon input of the measured data obtained from the experiment, the digitization unit 22 digitizes the categorical variables included in the measured data (Step S5).

[0091] As described above, in the first embodiment, the measurement data acquisition unit 21 acquires measurement data including the manufacturing conditions of a sample containing qualitative variables, and the measured values ​​of the physical properties of the sample manufactured based on the manufacturing conditions. The similarity acquisition unit 27 acquires the similarity between multiple levels included in the qualitative variables. The digitization unit 22 digitizes the qualitative variables from the measurement data acquired by the measurement data acquisition unit 21 based on the similarity acquired by the similarity acquisition unit 27. The model generation unit 23 generates a predictive model of the physical properties based on the measurement data after the qualitative variables have been digitized by the digitization unit 22.

[0092] Therefore, even when measured data includes qualitative variables that cannot be expressed numerically, materials with desired physical properties can be developed appropriately and efficiently by quantifying these qualitative variables.

[0093] In the first embodiment, the similarity setting unit 24 sets the similarity between multiple levels included in the qualitative variable. The similarity acquisition unit 27 acquires the similarity set by the similarity setting unit 24.

[0094] Therefore, similarity scores used to quantify qualitative variables can be obtained simply and appropriately.

[0095] In the first embodiment, the similarity setting unit 24 sets the similarity by at least one of the following: inputting or modifying the similarity on the input screen, inputting a file containing the similarity, and reusing previously entered similarity values, or by automatic setting based on predetermined rules.

[0096] Therefore, it is possible to calculate a predictive model that reflects the user's intentions.

[0097] In the first embodiment, the conversion table creation unit 25 creates a conversion table that quantifies qualitative variables based on the similarity obtained by the similarity acquisition unit 27. The quantification unit 22 then quantifies the qualitative variables using the conversion table created by the conversion table creation unit 25.

[0098] Therefore, qualitative variables can be appropriately and efficiently quantified based on the conversion table.

[0099] In the first embodiment, the conversion table creation unit 25 creates a conversion table based on multidimensional scaling.

[0100] Therefore, a conversion table can be created simply and appropriately.

[0101] Furthermore, in the first embodiment, the qualitative variable is a qualitative variable with three or more levels expressed on a nominal scale. That is, in the first embodiment, the qualitative variable is a categorical variable. The categorical variable may include at least one of the materials, apparatus, and methods used in the preparation of the sample.

[0102] Conventionally, Bayesian optimization could not be applied to categorical variables, which are qualitative variables represented by a nominal scale and have three or more levels. For qualitative variables represented by an ordinal scale, Bayesian optimization can be applied by performing simple quantification using rank values. For example, the ordinal scales "small", "medium", and "large" can be simply quantified as "1", "2", and "3". In addition, for a qualitative variable represented by a nominal scale with 2 levels, the two levels can be simply replaced with different numerical values, such as quantifying one level value as "0" and the other as "1". That is, even when a parameter is a qualitative variable, optimization using conventional simple quantification is possible for ordinal scales and for cases where the number of levels is 2. On the other hand, when a parameter includes a categorical variable with 3 or more levels, conventionally it is necessary to optimize the parameter by performing case division for each level. In optimization with case division, it was necessary to perform optimization for each level while also taking into account the parameters of other levels. For example, when there is one type of categorical variable and the number of levels is m (≧3), in order to optimize the parameters for each of the m levels, it was necessary to also consider the parameters of the remaining levels. Especially when the number of categorical variables is large, the number of case divisions increases sharply, making it realistically difficult to examine all cases. Note that there are k types of categorical variables, and the number of levels is m １ (≧3), m ２ (≧3), ..., m ｋ (≧3), m １ ×m ２ ×...×m ｋ case divisions are required. In contrast, according to the first embodiment, even when measured data includes categorical variables with a large number of levels, a prediction model that takes data of all levels into account can be constructed by quantification based on similarity. Accordingly, it is possible to obtain optimal parameter values for achieving desired physical properties, and to improve search efficiency.

[0103] In the first embodiment, the model generation unit 23 generates a predictive model using Bayesian estimation according to equations (1) and (2). That is, the model generation unit 23 generates a predictive model using Bayesian estimation according to equations (3) and (4).

[0104] Therefore, it is possible to generate a predictive model that can accurately search for optimal manufacturing conditions.

[0105] In the first embodiment, the manufacturing condition calculation unit 26 calculates second manufacturing conditions for the sample based on the predictive model generated by the model generation unit 23.

[0106] Therefore, optimal manufacturing conditions can be explored based on the predictive model.

[0107] Furthermore, in the first embodiment, the second manufacturing conditions include manufacturing conditions in which the physical properties in the prediction model fall within a predetermined target range.

[0108] Therefore, the number of prototypes required to obtain the desired physical properties can be reduced, and materials with the desired properties can be developed efficiently.

[0109] In the first embodiment, the target range may be set by upper and lower limits entered by the user. Alternatively, the target range may be automatically set based on specification values ​​stored in the storage unit 5.

[0110] Therefore, the target range can be set simply and appropriately.

[0111] In the first embodiment, the measurement data acquisition unit 21 further acquires second measurement data, which includes second manufacturing conditions calculated by the manufacturing condition calculation unit 26 and second measurement values ​​obtained by measuring the physical properties of a sample manufactured based on the second manufacturing conditions. The model generation unit 23 updates the prediction model based on the second measurement data acquired by the measurement data acquisition unit 21.

[0112] Therefore, by optimizing the experimental design, it becomes possible to appropriately develop materials with the desired physical properties.

[0113] In the first embodiment, the model generation unit 23 determines whether the second measured value is the set optimal value by comparing the second measured value with the optimal value. If it is determined that the second measured value is not the optimal value, the model generation unit 23 updates the prediction model based on the second measured data. On the other hand, if the second measured value is the optimal value, the model generation unit 23 does not update the prediction model.

[0114] Therefore, the number of sample prototypes can be reduced.

[0115] (Second Embodiment) Next, a second embodiment, which provides an input screen to the user's terminal device, will be described, focusing on the differences from the first embodiment. Figure 10 is a block diagram showing the physical property prediction model generation device 1 according to the second embodiment.

[0116] Figure 1 illustrates an example in which the input interface 3 and display unit 4 are provided on the physical property prediction model generation device 1. In contrast, in the example shown in Figure 10, the input interface 3 and display unit 4 are provided on the user's terminal device 9. The physical property prediction model generation device 1 and the terminal device 9 are connected via a network 8 for communication.

[0117] The physical property prediction model generation device 1 includes a communication unit 7. The communication unit 7 is an interface for communicating with a terminal device 9 via a network 8. For example, the communication unit 7 may be a network card and a network adapter. The communication unit 7 transmits the input screen generated by the similarity setting unit 24 to the terminal device 9. The communication unit 7 receives the similarity data entered on the input screen from the terminal device 9.

[0118] The conversion table creation unit 25 acquires similarity data received by the communication unit 7. The conversion table creation unit 25 creates a conversion table based on the similarity data acquired from the communication unit 7.

[0119] The model generation unit 23 may transmit the prediction model it has generated to the terminal device 9 via the communication unit 7. The prediction model transmitted to the terminal device 9 may be displayed on the display unit 4 of the terminal device 9.

[0120] The manufacturing condition calculation unit 26 may transmit the second manufacturing conditions calculated by the manufacturing condition calculation unit 26 to the terminal device 9 via the communication unit 7. The second manufacturing conditions transmitted to the terminal device 9 may be displayed on the display unit 4 of the terminal device 9.

[0121] According to the second embodiment, a user using the terminal device 9 can search for optimal manufacturing conditions through an input screen provided by the physical property prediction model generation device 1, i.e., the server.

[0122] While several embodiments of this disclosure have been described, these embodiments are presented as examples and are not intended to limit the scope of the disclosure. These embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of this disclosure, as well as in the claims of this disclosure and its equivalents.

Claims

1. A physical property prediction model generation apparatus comprising: a measurement data acquisition unit that acquires measurement data including manufacturing conditions for a sample containing qualitative variables and measured values ​​of the physical properties of the sample manufactured based on the manufacturing conditions; a similarity acquisition unit that acquires the similarity between a plurality of levels included in the qualitative variables; a digitization unit that digitizes the qualitative variables from the measurement data acquired by the measurement data acquisition unit based on the similarity acquired by the similarity acquisition unit; and a model generation unit that generates a prediction model of the physical properties based on the measurement data after the qualitative variables have been digitized by the digitization unit.

2. The physical property prediction model generation apparatus according to claim 1, further comprising a similarity setting unit for setting the similarity, and the similarity acquisition unit for acquiring the similarity set by the similarity setting unit.

3. The physical property prediction model generation apparatus according to claim 2, wherein the similarity setting unit sets the similarity by at least one of inputting or modifying the similarity on an input screen, inputting a file in which the similarity is described, and reusing a similarity that has been input in the past, or by automatic setting based on a predetermined rule.

4. The physical property prediction model generation apparatus further comprises: a display unit that displays an input screen for receiving input operations for inputting or modifying the similarity; and an input interface for receiving the input operations for the input screen, wherein the similarity setting unit sets the similarity based on the input operations received via the input interface for the input screen displayed on the display unit, the physical property prediction model generation apparatus according to claim 2.

5. The physical property prediction model generation apparatus according to claim 1, further comprising a conversion table creation unit that creates a conversion table for quantifying the qualitative variables based on the similarity obtained by the similarity acquisition unit, wherein the quantification unit quantifies the qualitative variables using the conversion table created by the conversion table creation unit.

6. The physical property prediction model generation apparatus according to claim 5, wherein the conversion table creation unit creates the conversion table based on multidimensional scaling.

7. The physical property prediction model generation apparatus according to any one of claims 1 to 6, wherein the qualitative variable is expressed on a nominal scale and has three or more levels.

8. The physical property prediction model generation apparatus according to claim 7, wherein the qualitative variables include at least one of at least three types of materials usable in the production of the sample, at least three types of apparatus usable in the production of the sample, and at least three types of methods usable in the production of the sample.

9. The physical property prediction model generation apparatus according to claim 1, wherein the model generation unit generates the prediction model using Bayesian estimation.

10. The physical property prediction model generation apparatus according to claim 1, wherein the model generation unit generates the prediction model according to the following formula.

11. The physical property prediction model generation apparatus according to claim 1, further comprising a manufacturing condition calculation unit that calculates second manufacturing conditions for the sample based on the prediction model generated by the model generation unit.

12. The physical property prediction model generating apparatus according to claim 11, wherein the second manufacturing conditions include manufacturing conditions when the physical properties in the prediction model fall within a preset target range.

13. The physical property prediction model generation apparatus according to claim 12, wherein the target range is set by upper and lower limits entered by the user.

14. The physical property prediction model generation apparatus according to claim 12, wherein the target range is set based on specification values ​​stored in the memory unit.

15. The physical property prediction model generation apparatus according to claim 12, wherein the measurement data acquisition unit further acquires second measurement data including the second manufacturing conditions calculated by the manufacturing condition calculation unit and second measurement values ​​obtained by measuring the physical properties of the sample manufactured based on the second manufacturing conditions, and the model generation unit updates the prediction model based on the second measurement data acquired by the measurement data acquisition unit.

16. The physical property prediction model generation apparatus according to claim 15, wherein the model generation unit determines whether the second measured value is a set optimal value by comparing the second measured value with the optimal value, and if the second measured value is not the optimal value, updates the prediction model based on the second measured data.

17. A method for generating a physical property prediction model performed by a computer, comprising: a step of acquiring measured data including manufacturing conditions for a sample that includes qualitative variables, and measured values ​​obtained by actually measuring the physical properties of the sample manufactured based on the manufacturing conditions; a step of acquiring the similarity between a plurality of levels included in the qualitative variables; a step of quantifying the qualitative variables from the acquired measured data based on the acquired similarity; and a step of generating a prediction model of the physical properties based on the measured data after the qualitative variables have been quantified.

18. The method for generating a physical property prediction model according to claim 17, further comprising the step of setting the similarity, wherein the step of obtaining the similarity is performed by obtaining the set similarity.

19. A method for generating a physical property prediction model according to claim 18, further comprising the step of creating a conversion table for quantifying the qualitative variables based on the obtained similarity, wherein the quantification step is performed using the created conversion table.

20. A property prediction model generation program for causing a computer to execute the following steps: a procedure for acquiring measured data including manufacturing conditions for a sample containing qualitative variables and measured values ​​of the physical properties of the sample manufactured based on the manufacturing conditions; a procedure for acquiring the similarity between multiple levels included in the qualitative variables; a procedure for quantifying the qualitative variables from the acquired measured data based on the acquired similarity; and a procedure for generating a predictive model of the physical properties based on the measured data after the qualitative variables have been quantified.

21. The physical property prediction model generation program according to claim 20, wherein the computer is further made to perform the procedure for setting the similarity, and the procedure for obtaining the similarity is performed by obtaining the set similarity.

22. The physical property prediction model generation program according to claim 21, wherein the computer is further made to perform a procedure to create a conversion table for quantifying the qualitative variables based on the acquired similarity, and the quantification procedure is performed using the created conversion table.

Images