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

Abstract

A physical property prediction device is provided with: an actual measurement data acquisition unit that acquires actual measurement data including a first manufacturing condition of a sample and an actual measurement value obtained by actual measurement of physical properties of the sample manufactured on the basis of the first manufacturing condition; a hypothetical data acquisition unit that acquires hypothetical data including a second manufacturing condition of the sample and a hypothetical value obtained by hypothesizing a physical property of the sample corresponding to the second manufacturing condition; and a model generation unit that generates a model on the basis of the measured data acquired by the measured data acquisition unit and the virtual data acquired by the virtual data acquisition unit. A prediction model of the physical properties is generated, the prediction model having a first uncertainty of the physical properties with respect to the measured data and a second uncertainty of the physical properties with respect to the virtual data, the second uncertainty being greater than the first uncertainty of the physical properties.

Description

Technical Field

[0001] This disclosure relates to a physical property prediction device, a physical property prediction method, and a physical property prediction procedure. Background Technology

[0002] In recent years, the field of materials informatics, which effectively utilizes information science technologies such as machine learning to improve the efficiency of materials development, has received increasing attention. In materials informatics, to manufacture materials with desired properties, it is necessary to find appropriate combinations of numerous parameters related to the proportions of raw materials and process conditions. When the number of parameters is large, the combinations of parameters become enormous. If the number of parameter combinations is enormous, it becomes difficult to comprehensively prototype materials. When comprehensively prototyping materials is difficult, the following method is adopted: measuring the properties of a certain quantity of the prototype material, and constructing a predictive model of the properties based on the measured data. By using the predictive model, even if the number of material prototypes is limited, the search for optimal parameters can be achieved.

[0003] To build predictive models, supervised machine learning methods are sometimes used. However, in supervised machine learning, constructing high-accuracy predictive models is difficult when the amount of measured data is insufficient or the relationship between parameters and property is complex. Even with insufficient measured data and complex relationships between parameters and property, successive experimental planning techniques, such as Bayesian optimization, are used to build high-accuracy predictive models. In Bayesian optimization, Bayesian estimation is used to construct the predictive model. Bayesian estimation, based on the idea of ​​Bayesian probability, infers the state to be estimated from observed events according to probabilistic meaning. Bayesian probability is interpreted as a reasonable expected value of a state of knowledge, or a quantified probability of an individual's belief. In Bayesian estimation, in addition to calculating the predicted value of a property, the uncertainty of the predicted value is also calculated. By calculating the uncertainty, the probability of a property becoming a specific value and the expected value of the property can be calculated. According to Bayesian estimation, optimal parameters can be obtained based on limited measured data. In other words, based on Bayesian estimation, an experimental plan for the next material trial can be formulated based on limited measured data. The predictive model is updated through the next material trial, the determination of the material's properties, and the addition of measured data to the results. In Bayesian optimization, a series of loops consisting of the aforementioned trial production, property determination, and predictive model updates are repeated. By repeating this series of loops, the number of trials required until the desired property values ​​are obtained can be minimized.

[0004] However, constructing an initial predictive model for Bayesian optimization is difficult when experimental data is limited or unavailable. This difficulty in constructing the initial predictive model makes developing a successive experimental plan challenging. Specifically, the number of experimental data points required to construct a predictive model is typically between 5 and 10. If the number of parameters is large, the required number of experimental data points may increase further. However, due to high monetary and time costs associated with material prototyping and property determination, and constraints on development budgets and timeframes, the number of prototyping operations that can be conducted is limited. This limited number of prototyping operations results in insufficient experimental data, making the construction of the initial predictive model and the application of Bayesian optimization difficult.

[0005] WO2021 / 200281A1 discloses a technique for applying machine learning to perform successive experimental planning on measured data. WO2020 / 188971A1 discloses a technique for pre-learning physical property values ​​by using simulations such as quantum chemical calculations to obtain the physical property values ​​corresponding to the parameters.

[0006] However, it is difficult to implement the techniques disclosed in WO2021 / 200281A1 without a sufficient amount of measured data in the initial stage. In particular, it is difficult to obtain the amount of measured data required to create the initial predictive model when the number of trial productions and property measurements is limited.

[0007] Furthermore, the technology disclosed in WO2020 / 188971A1 is limited to applications in fields where simulation methods are established as a part of the overall approach. In other words, the technology disclosed in WO2020 / 188971A1 is difficult to implement in many areas where simulation applications are challenging. Summary of the Invention

[0008] This disclosure was made with consideration of the above points in mind, and its purpose is to provide a property prediction device, property prediction method, and property prediction procedure that can appropriately and efficiently develop materials with desired physical properties.

[0009] The physical property prediction apparatus disclosed herein comprises: a measured data acquisition unit that acquires measured data including a first manufacturing condition of a sample and measured values ​​of physical properties of the sample manufactured based on the first manufacturing condition; a hypothetical data acquisition unit that acquires hypothetical data including a second manufacturing condition of the sample and hypothetical values ​​of physical properties of the sample corresponding to the second manufacturing condition; and a model generation unit that, based on the measured data acquired by the measured data acquisition unit and the hypothetical data acquired by the hypothetical data acquisition unit, generates a prediction model of the physical property such that it has a first uncertainty regarding the physical property with respect to the measured data and a second uncertainty regarding the physical property with respect to the hypothetical data that is greater than the first uncertainty regarding the physical property.

[0010] In the property prediction apparatus of this disclosure, the hypothetical data acquisition unit may acquire the property value of the sample predicted when the value of at least one of the multiple parameters included in the second manufacturing conditions is extremely small or large, and use it as the hypothetical value.

[0011] In the physical property prediction device disclosed herein, the hypothetical data acquisition unit may acquire the physical property values ​​of the raw materials or materials with processes similar to the sample as the hypothetical values.

[0012] In the property prediction device disclosed herein, the first uncertainty of the property may be 0.

[0013] In the physical property prediction apparatus disclosed herein, it may further include: an input screen generation unit that generates an input screen for inputting the measured data and the hypothetical data; a measured data acquisition unit that acquires the measured data input to the input screen generated by the input screen generation unit; and a hypothetical data acquisition unit that acquires the hypothetical data input to the input screen generated by the input screen generation unit.

[0014] The physical property prediction apparatus disclosed herein may further include: a communication unit that sends the input screen generated by the input screen generation unit to a user's terminal device, receives the measured data and the hypothetical data input to the input screen from the terminal device, the measured data acquisition unit acquires the measured data received by the communication unit, and the hypothetical data acquisition unit acquires the hypothetical data received by the communication unit.

[0015] In the property prediction device disclosed herein, it may further include: a hypothetical data supplementation unit that supplements the hypothetical data, and a hypothetical data acquisition unit that acquires the hypothetical data supplemented by the hypothetical data supplementation unit.

[0016] In the property prediction device disclosed herein, the hypothetical data supplementation unit may supplement the second manufacturing conditions of the hypothetical data by calculating the values ​​of parameters other than the portion of the multiple parameters when inputting a portion of the parameters with extremely small or large values ​​and the hypothetical values ​​predicted based on the portion of the parameters into the input screen.

[0017] In the property prediction apparatus disclosed herein, it may further include: an uncertainty setting unit that sets a first uncertainty and a second uncertainty of the property, and a model generation unit that further generates the prediction model based on the first uncertainty and the second uncertainty of the property set by the uncertainty setting unit.

[0018] In the property prediction apparatus disclosed herein, it may further include: a setting screen generation unit that generates a setting screen for setting a first uncertainty and a second uncertainty of the property, wherein the uncertainty setting unit sets the first uncertainty and the second uncertainty of the property based on an input operation to the setting screen generated by the setting screen generation unit.

[0019] In the property prediction apparatus disclosed herein, the model generation unit may use Bayesian estimation to generate the prediction model, wherein the first uncertainty is a first variance, and the second uncertainty is a second variance that is larger than the first variance.

[0020] In the physical property prediction apparatus disclosed herein, the model generation unit may generate the prediction model according to the following formula.

[0021] [Mathematical Expression 1]

[0022]

[0023] [Mathematical Expression 2]

[0024]

[0025] The physical property prediction apparatus disclosed herein may further include: a manufacturing condition calculation unit that calculates a third manufacturing condition for the sample based on the prediction model generated by the model generation unit.

[0026] In the property prediction apparatus of this disclosure, the third manufacturing condition may be at least one of the manufacturing conditions when the uncertainty of the property in the prediction model is maximized and the manufacturing conditions when the uncertainty of the property in the prediction model is minimized.

[0027] In the property prediction apparatus disclosed herein, the measured data acquisition unit may further acquire second measured data, which includes the third manufacturing conditions calculated by the manufacturing conditions calculation unit and the measured properties of the sample manufactured based on the third manufacturing conditions, and the model generation unit updates the prediction model based on the second measured data acquired by the measured data acquisition unit.

[0028] In the property prediction apparatus disclosed herein, the model generation unit does not update the prediction model if the second measured value is the set optimal value.

[0029] The property prediction method disclosed herein comprises the following steps: obtaining measured data including a first manufacturing condition of the sample and measured values ​​of the physical properties of the sample manufactured based on the first manufacturing condition; obtaining hypothetical data including a second manufacturing condition of the sample and hypothetical values ​​of the physical properties of the sample corresponding to the second manufacturing condition; and generating a prediction model of the physical property based on the obtained measured data and hypothetical data, wherein the measured data has a first uncertainty of the physical property and the hypothetical data has a second uncertainty of the physical property that is greater than the first uncertainty of the physical property.

[0030] The property prediction program disclosed herein enables a computer to perform the following steps: obtaining measured data comprising a first manufacturing condition of a sample and measured values ​​of the physical properties of the sample manufactured based on the first manufacturing condition; obtaining hypothetical data comprising a second manufacturing condition of the sample and hypothetical values ​​of the physical properties of the sample corresponding to the second manufacturing condition; and generating a prediction model of the physical property based on the obtained measured data and hypothetical data, such that the measured data has a first uncertainty regarding the physical property and the hypothetical data has a second uncertainty regarding the physical property that is greater than the first uncertainty regarding the physical property.

[0031] According to this disclosure, materials with desired physical properties can be developed appropriately and efficiently. Attached Figure Description

[0032] Figure 1 This is a schematic diagram illustrating the property prediction device of the first embodiment.

[0033] Figure 2 This is a block diagram showing the property prediction device of the first embodiment.

[0034] Figure 3 This is a flowchart illustrating an example of the operation of the property prediction device according to the first embodiment.

[0035] Figure 4This is a diagram showing an example of the input screen in an operational example of the property prediction device of the first embodiment.

[0036] Figure 5 This is a diagram illustrating an example of the prediction model in an operational example of the property prediction device of the first embodiment.

[0037] Figure 6 This is a diagram illustrating another example of the prediction model in the operation example of the property prediction device of the first embodiment.

[0038] Figure 7 This is a diagram illustrating an example of calculation based on the third manufacturing condition of the prediction model in the operation example of the property prediction device of the first embodiment.

[0039] Figure 8 This is a block diagram illustrating the property prediction device of the second embodiment.

[0040] Figure 9 This is a flowchart illustrating an example of the operation of the property prediction device according to the second embodiment.

[0041] Figure 10 This is a diagram showing an example of the input screen in an operational example of the property prediction device of the second embodiment.

[0042] Figure 11 This is a block diagram illustrating the property prediction device of the third embodiment.

[0043] Figure 12 This is a flowchart illustrating an example of the operation of the property prediction device according to the third embodiment.

[0044] Figure 13 This is a diagram showing an example of the input screen in an operational example of the property prediction device of the third embodiment.

[0045] Figure 14 This is a block diagram illustrating the property prediction device of the fourth embodiment. Detailed Implementation

[0046] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. In the drawings referred to in the embodiments, the same reference numerals or similar reference numerals are used for the same parts or parts having the same function, and sometimes repeated descriptions are omitted.

[0047] [First Embodiment]

[0048] First, refer to Figures 1 to 7 This describes the property prediction device of the first embodiment. Figure 1 This is a conceptual diagram showing the property prediction device 1 of the first embodiment.

[0049] like Figure 1As shown, the property prediction device 1 of the first embodiment generates a prediction model for a sample based on measured data and prediction data. The measured data includes data on the first manufacturing conditions of the sample and measured values. The sample is, for example, a material manufactured based on multiple raw materials. The first manufacturing conditions include, for example, the proportions of the raw materials, temperature, and other process conditions. The measured values ​​are values ​​obtained by actually measuring the properties of the sample manufactured under the first manufacturing conditions. The properties of the sample include, for example, conductivity and dielectric constant. The prediction data includes data on the second manufacturing conditions of the sample and hypothetical values. The second manufacturing conditions are different from the first manufacturing conditions. The second manufacturing conditions are not the actual manufacturing conditions used to manufacture the sample, but rather hypothetical manufacturing conditions. The hypothetical values ​​are values ​​obtained by hypothesizing the properties of the sample corresponding to the second manufacturing conditions. The prediction model is information representing the predicted properties of the sample based on the sample's manufacturing conditions. In other words, the prediction model is information representing the correspondence between the sample's manufacturing conditions and its properties.

[0050] Furthermore, the property prediction device 1 generates a new experimental plan for the sample based on the generated prediction model. A new sample is then prepared according to the new experimental plan. Experiments are conducted on the prepared new sample to measure its physical properties. The measured data obtained from the experiments are added to the property prediction device 1. The property prediction device 1 updates the prediction model based on the updated measured data and hypothetical data obtained through this addition.

[0051] As mentioned above, the property prediction device 1 optimizes the experimental plan by generating a prediction model based on measured data and hypothetical data, creating an experimental plan, and repeating the experiment with additional measured data.

[0052] Figure 2 This is a block diagram illustrating the property prediction device 1 according to the first embodiment. More specifically, as shown... Figure 2 As shown, the property prediction device 1 includes a control unit 2, an input interface 3, a display unit 4, and a storage unit 5. The control unit 2, input interface 3, display unit 4, and storage unit 5 are connected in a communicative manner via a bus 6.

[0053] Input interface 3 receives various instructions and information input from the user. Input interface 3 converts the input received from the user into electrical signals. Input interface 3 outputs the converted electrical signals to control unit 2. Input interface 3 can be, for example, a mouse, keyboard, touchpad, touch panel, trackball, switch button, and microphone.

[0054] Display unit 4 displays an image corresponding to the image data sent from control unit 2. Display unit 4 may be, for example, an LCD monitor, a CRT (Cathode Ray Tube) monitor, or a touch panel. Display unit 4 may include a speaker.

[0055] Storage unit 5 is a non-temporary storage device that stores various types of information. Storage unit 5 may be, for example, an HDD (Hard Disk Drive), optical disc, magnetic disk, optical disc drive, CD (Compact Disc), DVD (Digital Versatile Disc), or semiconductor memory. Storage unit 5 may store, for example, the property prediction program executed by property prediction device 1 and various data used in the execution of the property prediction program.

[0056] The control unit 2 controls the operation of the physical property prediction device 1 according to the input operation of the user input interface 3. The control unit 2 includes a measured data acquisition unit 21, a hypothetical data acquisition unit 22, a model generation unit 23, an input screen generation unit 24, and a manufacturing condition calculation unit 25.

[0057] The functions corresponding to each structural unit 21-25 of the control unit 2 are recorded in the storage unit 5 as a property prediction program executable by a computer. The control unit 2 is, for example, a processor. The processor constituting the control unit 2 executes the functions of structural units 21-25 corresponding to the read property prediction program by reading from the storage unit 5 and executing the property prediction program. Alternatively, instead of storing the property prediction program in the storage unit 5, the property prediction program can be directly programmed into the processor's circuitry. In the case of directly programming the property prediction program, the processor executes the property prediction program programmed into the circuitry to perform the functions corresponding to each structural unit 21-25. The property prediction program can be a program downloaded from a server by the property prediction device 1 via a network. Alternatively, the property prediction program can be a program retrieved from a portable storage medium into the property prediction device 1. The control unit 2 can be composed of a single processor. The control unit 2 can also be composed of a combination of multiple processors. Processors can be, for example, CPUs (Central Processing Units), GPUs (Graphics Processing Units), Application Specific Integrated Circuits (ASICs), and programmable logic devices. Programmable logic devices can be SPLDs (Simple Programmable Logic Devices), CPLDs (Complex Programmable Logic Devices), or FPGAs (Field Programmable Gate Arrays).

[0058] The measured data acquisition unit 21 acquires measured data including the first manufacturing conditions of the sample and measured values ​​of the physical properties of the sample manufactured based on the first manufacturing conditions.

[0059] The hypothetical data acquisition unit 22 acquires hypothetical data including the second manufacturing conditions of the sample and hypothetical values ​​of the physical properties of the sample corresponding to the second manufacturing conditions.

[0060] The hypothetical data acquisition unit 22 can acquire the sample's physical property value as a hypothetical value when the value of one of the multiple parameters included in the second manufacturing condition is extremely small or large.

[0061] The case of an extremely small parameter value refers to a value that is the lower limit of the pre-determined search region or a value below the lower limit. The case of an extremely large parameter value refers to a value that is the upper limit of the pre-determined search region or a value exceeding the upper limit.

[0062] The hypothetical data acquisition unit 22 can also replace the predicted physical property values ​​when the values ​​of some parameters are extremely small or large, and obtain the physical property values ​​of the raw materials or processes similar to the sample as hypothetical values.

[0063] The model generation unit 23 generates a predictive model of the physical property based on the measured data obtained by the measured data acquisition unit 21 and the hypothetical data obtained by the hypothetical data acquisition unit 22. The predictive model can be a predictive model that expresses the probability of the physical property to be predicted as a probability. The model generation unit 23 generates a predictive model that has a first uncertainty regarding the physical property with respect to the measured data and a second uncertainty regarding the physical property with respect to the hypothetical data that is greater than the first uncertainty. The first uncertainty regarding the physical property is, for example, 0.

[0064] The input screen generation unit 24 generates an input screen for inputting measured data and hypothetical data. The input screen generation unit 24 displays the generated input screen on the display unit 4. The measured data acquisition unit 21 acquires the measured data input to the input screen generated by the input screen generation unit 24. More specifically, the measured data acquisition unit 21 acquires the measured data input to the input screen displayed on the display unit 4 using the input interface 3. The hypothetical data acquisition unit 22 acquires the hypothetical data input to the input screen generated by the input screen generation unit 24. More specifically, the hypothetical data acquisition unit 22 acquires the hypothetical data input to the input screen displayed on the display unit 4 using the input interface 3.

[0065] The manufacturing condition calculation unit 25 calculates the third manufacturing condition of the sample based on the prediction model generated by the model generation unit 23.

[0066] The third manufacturing condition may also include at least one of the manufacturing conditions when the uncertainty of the physical properties in the prediction model is maximized and the manufacturing conditions when the uncertainty of the physical properties in the prediction model is minimized.

[0067] The measured data acquisition unit 21 further acquires second measured data, which includes the third manufacturing conditions calculated by the manufacturing conditions calculation unit 25 and the second measured value obtained by measuring the physical properties of the sample manufactured based on the third manufacturing conditions.

[0068] The model generation unit 23 updates the prediction model based on the second measured data obtained by the measured data acquisition unit 21. The model generation unit 23 does not update the prediction model if the second measured value is the set optimal value.

[0069] The model generation unit 23 can use Bayesian estimation to generate a predictive model, i.e., a probabilistic predictive model. The first uncertainty can be the first variance, and the second uncertainty can be a second variance that is larger than the first variance.

[0070] When using Bayesian estimation to generate a prediction model, the model generation unit 23 can generate the prediction model according to the following formula.

[0071] [Mathematical Expression 3]

[0072]

[0073] [Mathematical Expression 4]

[0074]

[0075]

[0076] In the generation of the prediction model according to formulas (1) and (2), the model generation unit 23 will generate the variance σ of the likelihood corresponding to the measured data. 2 Setting it small will reduce the variance σ of the likelihood corresponding to the hypothetical data. 2 The value is set large. The model generation unit 23 can generate the variance σ of the likelihood corresponding to the measured data. 2 Set to 0.

[0077]

[0078] [Action Example]

[0079] Next, an example of the operation of the property prediction device 1 with the above structure will be explained. Figure 3 This is a flowchart illustrating an example of the operation of the property prediction device 1 according to the first embodiment. First, as... Figure 3The premise is that the user creates hypothetical data. Users can create hypothetical data through experience-based predictions. For example, when multiple parameters included in the second manufacturing condition are each extremely small or large, the physical property values ​​of the sample are sometimes predicted based on experience. In the case of predicting the physical property values ​​of the sample based on experience, the user uses multiple parameters taking extremely small or large values ​​as the second manufacturing condition, and uses the physical property values ​​of the sample to be predicted as hypothetical values ​​to create hypothetical data. Alternatively, the user can also use known physical property values ​​of raw materials or substances with processes similar to the sample to create hypothetical data. When using known physical property values ​​of a substance, the user uses the process of that substance as the second manufacturing condition, and uses the known physical property values ​​of that substance as hypothetical values ​​to create hypothetical data.

[0080] Furthermore, the user generates measured data. Specifically, the user obtains measured values ​​by determining the physical properties of a sample manufactured under the first manufacturing conditions. The user generates measured data by establishing a correspondence between the first manufacturing conditions and the measured values. In generating measured data, the user can set 1 to 5 levels of combinations of values ​​for multiple parameters constituting the first manufacturing conditions. Furthermore, the user generates measured data by measuring the physical properties of a sample prepared according to the set combinations of parameter values. When setting combinations of multiple parameter values, the user can randomly set the values ​​of multiple parameters of the first manufacturing conditions within a given range. Alternatively, the user can set the values ​​of multiple parameters of the first manufacturing conditions by uniformly selecting parameter values ​​within a given range, such as through Latin square sampling.

[0081] After generating hypothetical and measured data, the input screen generation unit 24 generates an input screen based on the user's input operation using the input interface 3. The input screen generation unit 24 then displays the generated input screen on the display unit 4 (step S1).

[0082] Figure 4 This is a diagram showing an example of the input screen SC in an operational example of the property prediction device 1 of the first embodiment. Figure 4 In the example shown, the input screen SC has input fields F1 for inputting parameters 1 to 5, which are examples of the first manufacturing condition. Additionally, the input screen SC has input fields F2 for inputting the first and second physical properties, which are examples of measured and hypothetical values. Furthermore, the input screen SC has an input field F3 for inputting the type of data, i.e., whether it's hypothetical data or measured data. If the execute button B is pressed after inputting data into input fields F1 to F3, the data input into input fields F1 to F3 is confirmed.

[0083] After the input screen is displayed on display unit 4, such as Figure 3As shown, the user inputs hypothetical data onto the input screen displayed on the display unit 4 by using the input operation of the input interface 3 (step S2).

[0084] After inputting hypothetical data, the user inputs actual measured data onto the input screen displayed on the display unit 4 using the input operation of input interface 3 (step S3). Steps S2 and S3 can be interchanged.

[0085] After inputting hypothetical and measured data, the model generation unit 23 generates a prediction model based on the input hypothetical and measured data (step S4). Figure 3 In the example shown, the model generation unit 23 generates a prediction model using Bayesian estimation according to the above-described formulas (1) and (2). When generating the prediction model, the model generation unit 23 calculates the variance σ of the likelihood corresponding to the measured data. 2 The value is set to 0. Furthermore, the model generation unit 23 sets the variance σ of the likelihood corresponding to the hypothetical data to 0. 2 The value is set to be greater than 0. In addition, the model generation unit 23 sets the prior distribution p(A) to have a high probability density when the value of the physical property θ changes smoothly.

[0086] The model generation unit 23 can display the generated prediction model on the display unit 4.

[0087] Figure 5 This is a diagram illustrating an example of the prediction model in an operational example of the property prediction device 1 of the first embodiment. Figure 6 This diagram illustrates another example of the prediction model in the operation example of the property prediction device 1 of the first embodiment. The model generation unit 23, for example, generates... Figure 5 as well as Figure 6 The prediction model shown. Figure 5 as well as Figure 6 The horizontal axis is parameter x, which is an example of the first manufacturing condition. Figure 5 as well as Figure 6 The vertical axis represents the physical property value θ. Figure 5 as well as Figure 6 The solid line in the graph represents the average value of the physical property θ corresponding to the parameter x. Measured data and hypothetical data are located on the solid line graph representing the average value. Figure 5 as well as Figure 6 The two dashed lines in the graph represent the uncertainty of the physical property θ. Specifically, Figure 5 as well as Figure 6 The two dashed lines in the graph represent the upper and lower bounds of the 50% confidence interval for the normal distribution, which is the probability distribution of property θ. As an example, in... Figure 5The normal distribution G at x = 0.4 is represented by a graph with a solid line representing the mean and a graph with a dashed line representing uncertainty. The shaded area within the normal distribution G corresponds to the 50% confidence interval at x = 0.4.

[0088] from Figure 5 as well as Figure 6 It can be seen that the uncertainty becomes zero in the vicinity of the measured data. The uncertainty becomes zero because the variance σ of the likelihood corresponding to the measured data is generated by the model generation unit 23. 2 This is set to 0. On the other hand, the uncertainty increases near the hypothetical data. The uncertainty increases because the variance σ of the likelihood corresponding to the hypothetical data is generated by the model generation unit 23. 2 The case where it is set to be greater than 0.

[0089] After generating the prediction model, such as Figure 3 As shown, the experimental plan is created (step S5). In the creation of the experimental plan, firstly, the manufacturing condition calculation unit 25 calculates the third manufacturing condition of the sample based on the prediction model generated by the model generation unit 23. Figure 7 This diagram illustrates an example of calculation based on the third manufacturing condition from the prediction model, within an operational example of the property prediction device 1 according to the first embodiment. Figure 7 In the example shown, the manufacturing condition calculation unit 25 calculates the manufacturing conditions when the uncertainty of the physical property in the prediction model is maximized and the manufacturing conditions when the uncertainty of the physical property in the prediction model is minimized, and uses these as the third manufacturing conditions. More specifically, the manufacturing condition calculation unit 25 calculates the value of parameter x1 when the uncertainty of physical property θ is maximized and the value of parameter x2 when the uncertainty of physical property θ is minimized, and uses these as the third manufacturing conditions. Here, when the uncertainty of physical property is maximized, it is possible to obtain very good physical properties, although the probability is low. Furthermore, when the uncertainty of physical property is minimized, the probability of obtaining good physical properties is high. Therefore, by setting the value of parameter x1 when the uncertainty of physical property θ is maximized and the value of parameter x2 when it is minimized as the third manufacturing conditions, the number of sample trials required until the desired physical properties are obtained can be reduced. Next, in the creation of the experimental plan, the user uses the calculated third manufacturing conditions to create a specific plan for manufacturing samples.

[0090] After the experimental plan is formulated, such as Figure 3 As shown, the user conducts the experiment (step S6). In conducting the experiment, firstly, the user manufactures a sample according to the experimental plan, i.e., manufacturing conditions 3. The sample is manufactured using a manufacturing apparatus (not shown). After manufacturing the sample, the user measures the physical properties of the manufactured sample. The physical properties of the sample are measured using a measuring instrument (not shown) corresponding to the physical properties.

[0091] If the experiment yields the desired physical properties (step S7: Yes), the process ends. Conversely, if the experiment fails to yield the desired physical properties (step S7: No) but further experiments can be performed (step S8: Yes), the user inputs the experimentally obtained measured data into the input screen (step S3). With the input of the experimentally obtained measured data, the model generation unit 23 updates the prediction model (step S4).

[0092] As described above, in the first embodiment, the measured data acquisition unit 21 acquires measured data including the first manufacturing conditions of the sample and measured values ​​of the physical properties of the sample manufactured based on the first manufacturing conditions. Furthermore, the hypothetical data acquisition unit 22 acquires hypothetical data including the second manufacturing conditions of the sample and hypothetical values ​​of the physical properties of a sample hypothetically corresponding to the second manufacturing conditions. Furthermore, the model generation unit 23 generates a predictive model of the physical properties based on the measured data acquired by the measured data acquisition unit 21 and the hypothetical data acquired by the hypothetical data acquisition unit 22. More specifically, the model generation unit 23 generates a predictive model of the physical properties that has a first uncertainty regarding the physical properties with respect to the measured data and a second uncertainty regarding the physical properties with respect to the hypothetical data that is greater than the first uncertainty.

[0093] Therefore, even when measured data is limited or unavailable, hypothetical data reflecting tacit knowledge such as experience and intuition in materials development can be effectively utilized to appropriately formulate successive experimental plans. Furthermore, even when the number of experiments that can be performed is limited, the number of prototypes required to construct the initial predictive model is difficult, and measurement is challenging, materials with desired physical properties can still be developed. Moreover, when a sufficiently large number of experiments can be performed, the number of prototypes required until the desired physical properties are obtained can be reduced by effectively utilizing tacit knowledge not expressed in measured data. Therefore, according to the first embodiment, materials with desired physical properties can be developed appropriately and efficiently.

[0094] In addition, in the first embodiment, the hypothetical data acquisition unit 22 may acquire the predicted physical property values ​​of the sample when some of the parameters included in the second manufacturing conditions are extremely small or large, and use them as hypothetical values.

[0095] Therefore, it is possible to obtain hypothetical data that is useful for generating predictive models in a simple way.

[0096] In addition, in the first embodiment, the hypothetical data acquisition unit 22 may acquire the physical property values ​​of the raw materials or the sample, which are similar to the sample, and use them as hypothetical values.

[0097] Therefore, it is possible to obtain hypothetical data that is useful for generating predictive models in a simple way.

[0098] Furthermore, in the first embodiment, the first uncertainty of the physical properties is 0.

[0099] Therefore, it is possible to generate predictive models that can easily search for optimal manufacturing conditions.

[0100] Furthermore, in the first embodiment, the property prediction device 1 also includes an input screen generation unit 24 that generates an input screen for inputting measured data and hypothetical data. Additionally, the measured data acquisition unit 21 acquires the measured data input to the input screen generated by the input screen generation unit 24. Furthermore, the hypothetical data acquisition unit 22 acquires the hypothetical data input to the input screen generated by the input screen generation unit 24.

[0101] Therefore, a predictive model that reflects the user's intentions can be calculated.

[0102] Furthermore, in the first embodiment, the model generation unit 23 generates a prediction model using Bayesian estimation based on formulas (1) and (2). That is, the model generation unit 23 generates a prediction model that has a first variance for the measured data and a second variance for the hypothetical data that is larger than the first variance.

[0103] Therefore, it is possible to generate predictive models that can search for optimal manufacturing conditions with high accuracy.

[0104] Furthermore, in the first embodiment, the manufacturing condition calculation unit 25 calculates the third manufacturing condition of the sample based on the prediction model generated by the model generation unit 23.

[0105] Therefore, it is possible to search for optimal manufacturing conditions based on a predictive model.

[0106] Furthermore, in the first embodiment, the third manufacturing condition includes at least one of the manufacturing conditions when the uncertainty of the physical property in the prediction model is maximized and the manufacturing conditions when the uncertainty of the physical property in the prediction model is minimized.

[0107] Therefore, it can reduce the number of sample preparations required to obtain the desired physical properties, and can efficiently develop materials with the desired physical properties.

[0108] Furthermore, in the first embodiment, the measured data acquisition unit 21 further acquires second measured data, which includes the third manufacturing conditions calculated by the manufacturing conditions calculation unit 25 and the second measured values ​​obtained by measuring the physical properties of the sample manufactured based on the third manufacturing conditions. Additionally, the model generation unit 23 updates the prediction model based on the second measured data acquired by the measured data acquisition unit 21.

[0109] Therefore, by optimizing the experimental plan, it is possible to develop materials with the desired physical properties.

[0110] Furthermore, in the first embodiment, the model generation unit 23 does not update the prediction model if the second measured value is the set optimal value.

[0111] Therefore, it can reduce the number of times samples need to be prepared.

[0112] [Second Embodiment]

[0113] Next, the second implementation method, which generates a prediction model based on the set uncertainty, will be described, focusing on the differences from the first implementation method. Figure 8 This is a block diagram showing the property prediction device 1 of the second embodiment.

[0114] like Figure 8 As shown, the control unit 2 of the property prediction device 1 in the second embodiment, in addition to the structure of the first embodiment, also includes an uncertainty setting unit 26 and a setting screen generation unit 27.

[0115] The uncertainty setting unit 26 sets the first uncertainty and the second uncertainty of the property. The model generation unit 23 also generates a prediction model based on the first uncertainty and the second uncertainty of the property set by the uncertainty setting unit 26.

[0116] The setting screen generation unit 27 generates a setting screen for setting a first uncertainty and a second uncertainty of a property. The setting screen generation unit 27 displays the generated setting screen on the display unit 4. The uncertainty setting unit 26 sets the first uncertainty and the second uncertainty of the property according to the input operation performed on the setting screen generated by the setting screen generation unit 27. More specifically, the uncertainty setting unit 26 sets the first uncertainty and the second uncertainty of the property according to the input operation performed on the setting screen displayed on the display unit 4 using the input interface 3. The setting screen generation unit 27 can display the setting screen simultaneously with the input screen displayed by the input screen generation unit 24. Alternatively, the setting screen generation unit 27 can display the setting screen at a different time than the input screen displayed by the input screen generation unit 24.

[0117] Figure 9 This is a flowchart illustrating an example of the operation of the property prediction device 1 according to the second embodiment. For example... Figure 9 As shown, in the second embodiment, the setting screen generation unit 27 displays the setting screen simultaneously with the input screen displayed by the input screen generation unit 24 (step S11).

[0118] Figure 10 This is a diagram showing an example of the input screen SC in an operational example of the property prediction device of the second embodiment. Figure 10In the example shown, the setting screen generation unit 27 displays a setting screen SC2 on the input screen SC. The setting screen SC2 has a slider S for inputting the ratio of the weights of the measured data to the weights of the hypothetical data. By sliding the slider S, the ratio of the weights of the measured data to the weights of the hypothetical data can be input, thereby setting a first uncertainty and a second uncertainty corresponding to the input ratio.

[0119] like Figure 9 As shown, in the second embodiment, the uncertainty setting unit 26 sets the first uncertainty and the second uncertainty according to the input operation for the setting screen (step S9).

[0120] The model generation unit 23 generates a prediction model based on the input hypothetical data, measured data, and the set first and second uncertainties (step S4).

[0121] As described above, in the second embodiment, the uncertainty setting unit 26 sets a first uncertainty and a second uncertainty of the property. Furthermore, the model generation unit 23 generates a prediction model based on the first uncertainty and the second uncertainty of the property set by the uncertainty setting unit 26.

[0122] Therefore, it is possible to calculate a predictive model that reflects the user's intentions, thereby increasing the degree of freedom in searching for manufacturing conditions.

[0123] Furthermore, in the second embodiment, the setting screen generation unit 27 generates a setting screen for setting a first uncertainty and a second uncertainty of the property. Additionally, the uncertainty setting unit 26 sets the first uncertainty and the second uncertainty of the property according to an input operation on the setting screen generated by the setting screen generation unit 27.

[0124] Therefore, the first uncertainty and the second uncertainty can be easily set using the settings screen.

[0125] [Third Embodiment]

[0126] Next, the third implementation method for supplementing hypothetical data will be described, focusing on the differences from the first implementation method. Figure 11 This is a block diagram showing the property prediction device 1 of the third embodiment.

[0127] like Figure 11 As shown, the control unit 2 of the property prediction device 1 in the third embodiment, in addition to the structure of the first embodiment, also includes a hypothetical data supplementation unit 28.

[0128] The hypothetical data supplementation unit 28 supplements hypothetical data. When the hypothetical data input to the input screen is not complete and some data is missing, the hypothetical data supplementation unit 28 supplements the hypothetical data by compensating for the missing data.

[0129] The hypothetical data acquisition unit 22 acquires the hypothetical data supplemented by the hypothetical data supplementation unit 28.

[0130] The hypothetical data supplementation unit 28 can supplement hypothetical data when a portion of the multiple parameters included in the second manufacturing condition, which have extremely small or large values, and hypothetical values ​​predicted based on a portion of the parameters are input to the input screen. Specifically, the hypothetical data supplementation unit 28 can supplement the hypothetical data of the second manufacturing condition by calculating the values ​​of parameters other than a portion of the multiple parameters.

[0131] Figure 12 This is a flowchart illustrating an example of the operation of the property prediction device 1 according to the third embodiment. Figure 13 This is a diagram showing an example of the input screen in an operational example of the property prediction device 1 of the third embodiment.

[0132] like Figure 12 As shown, in the third embodiment, the hypothetical data supplementation unit 28 supplements hypothetical data when the hypothetical data input to the input screen is not complete (step S10).

[0133] For example, the second manufacturing condition includes m parameters x1, x2, ..., x... m Among them, the value of one parameter x1 is either extremely small or extremely large. ＊ Sometimes, the physical property value of a sample can be predicted based on the user's tacit knowledge. ＊ When the predictable physical property value is θ ＊ In this case, the user inputs x1 on the input screen. ＊ Use θ as the value of parameter x1. ＊ Use this as the property value, without inputting values ​​for parameters other than x1. Figure 13 In the example shown, an extremely small 0 is input to the input screen SC as the value of the first parameter. Furthermore, 0.0 is input to the input screen SC as the first and second properties, respectively. Additionally, in the input screen SC, the values ​​of parameters other than the first parameter are set to undetermined, i.e., ***.

[0134] Without inputting values ​​for parameters other than x1, the hypothetical data is incomplete. In the case of incomplete hypothetical data, the hypothetical data completion unit 28 calculates parameters x2, ..., x... other than x1. m The values ​​are added to complete the hypothetical data. For example, the hypothetical data completion unit 28 adds the predetermined parameters x2, ..., x...m Each search region is divided into k levels. Furthermore, the hypothetical data completion unit 28 calculates k by combining the values ​​selected from each of the k-level segmented regions. m-1 Types of parameters x2, ..., x m The value of can also be used instead of the k-level calculation method. Hypothetical data completion part 28 calculates the parameters x2, ..., x by generating m-1 dimensional uniform random numbers. m The value of .

[0135] Furthermore, the second manufacturing condition includes m parameters x1, x2, ..., x... m When two or more but fewer than m-1 parameters have extremely small or large values, it is sometimes possible to predict the physical properties of a sample based on the user's tacit knowledge. In cases where the physical property can be predicted, the user inputs the values ​​of two or more but fewer than m-1 parameters into the input screen, along with the physical property to be predicted, excluding any parameters other than the two or more but fewer than m-1 parameters. Figure 13 In the example shown, an extremely large value of 100 is input to the input screen SC as the value of the third parameter. Furthermore, an extremely large value of 100 is input to the input screen SC as the value of the fourth parameter. Additionally, 100.0 is input to the input screen SC as the first property. Furthermore, 133.3 is input to the input screen SC as the second property. In addition, in the input screen SC, the values ​​of parameters other than the third and fourth parameters are set to undetermined.

[0136] If no values ​​for any parameters other than two to m-1 parameters are input, the hypothetical data is incomplete. In the case of incomplete hypothetical data, the hypothetical data completion unit 28, for example, divides the search region for each parameter other than two to m-1 parameters into k levels. Furthermore, the hypothetical data completion unit 28 calculates the parameter values ​​obtained by combining the values ​​selected from each of the k-level segmented regions.

[0137] When measured data of a substance similar to the sample is used as a reference, the user can input the measured data of that substance into the input screen as hypothetical data.

[0138] If complete hypothetical data is input to the input screen, the hypothetical data supplementation unit 28 does not supplement the hypothetical data.

[0139] As described above, in the third embodiment, the hypothetical data supplementation unit 28 supplements the hypothetical data. Furthermore, the hypothetical data acquisition unit 22 acquires the hypothetical data supplemented by the hypothetical data supplementation unit 28.

[0140] Therefore, it can efficiently generate hypothetical data.

[0141] Furthermore, in the third embodiment, when the hypothetical data supplementation unit 28 inputs a portion of the parameters with extremely small or large values ​​from among the multiple parameters included in the second manufacturing conditions and a hypothetical value predicted based on a portion of the parameters into the input screen, it supplements the second manufacturing conditions of the hypothetical data by calculating the values ​​of the parameters other than a portion of the multiple parameters.

[0142] Therefore, it can efficiently supplement hypothetical data.

[0143] [Fourth Implementation]

[0144] Next, the fourth embodiment, which provides an input screen to the user's terminal device, will be described, focusing on the differences from the first embodiment. Figure 14 This is a block diagram showing the property prediction device 1 of the fourth embodiment.

[0145] exist Figure 1 In this example, an example will be described where the input interface 3 and the display unit 4 are installed in the property prediction device 1. Compared to... Figure 1 For example, in Figure 14 In the example shown, the input interface 3 and the display unit 4 are installed in the user's terminal device 9. The property prediction device 1 and the terminal device 9 are connected in a communicable manner via the network 8.

[0146] The property prediction 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 can be a network card or a network adapter. The communication unit 7 sends an input screen generated by the input screen generation unit 24 to the terminal device 9. The communication unit 7 receives measured data and hypothetical data input to the input screen from the terminal device 9.

[0147] The measured data acquisition unit 21 acquires the measured data received by the communication unit 7. The hypothetical data acquisition unit 22 acquires the hypothetical data received by the communication unit 7.

[0148] The model generation unit 23 can send the prediction model generated by the model generation unit 23 to the terminal device 9 via the communication unit 7. The prediction model sent to the terminal device 9 can be displayed on the display unit 4 of the terminal device 9.

[0149] The manufacturing conditions calculation unit 25 can transmit the manufacturing conditions calculated by the manufacturing conditions calculation unit 25 to the terminal device 9 via the communication unit 7. The manufacturing conditions transmitted to the terminal device 9 can be displayed on the display unit 4 of the terminal device 9.

[0150] According to the fourth embodiment, a user using the terminal device 9 can search for optimal manufacturing conditions by means of an input screen provided by the property prediction device 1, i.e., the server.

[0151] Several embodiments of this disclosure have been described, but these embodiments are given by way of example and are not intended to limit the scope of the disclosure. These embodiments can be implemented in various other ways, and various omissions, substitutions, and modifications can be made without departing from the spirit of this disclosure. The embodiments and their variations are included within the scope and spirit of this disclosure, as are the claims and their equivalents.

Claims

1. A physical property prediction device, characterized by comprising: Possessing: a measured data acquisition section that acquires measured data including a first manufacturing condition of a sample and a measured value obtained by measuring a property of the sample manufactured based on the first manufacturing condition; a hypothetical data acquisition section that acquires hypothetical data including a second manufacturing condition of the sample and a hypothetical value obtained by assuming a property of the sample corresponding to the second manufacturing condition; and a model generation section that generates a prediction model of the property having a first uncertainty of the property for the measured data and a second uncertainty of the property greater than the first uncertainty of the property for the hypothetical data based on the measured data acquired by the measured data acquisition section and the hypothetical data acquired by the hypothetical data acquisition section.

2. The property prediction device according to claim 1, wherein the hypothetical data acquisition section acquires a property value of the sample predicted in a case where a value of at least one parameter among a plurality of parameters included in the second manufacturing condition is extremely small or large, as the hypothetical value.

3. The property prediction device according to claim 1, wherein the hypothetical data acquisition section acquires a property value of a material or a substance similar to the sample in a process, as the hypothetical value.

4. The property prediction device according to claim 1, wherein the first uncertainty of the property is 0.

5. The property prediction device according to claim 1, wherein the property prediction device further possesses an input screen generation section that generates an input screen for inputting the measured data and the hypothetical data, the measured data acquisition section acquires the measured data input to the input screen generated by the input screen generation section, the hypothetical data acquisition section acquires the hypothetical data input to the input screen generated by the input screen generation section.

6. The property prediction device according to claim 5, wherein the property prediction device further possesses a communication section that transmits the input screen generated by the input screen generation section to a terminal device of a user and receives the measured data and the hypothetical data input to the input screen from the terminal device, the measured data acquisition section acquires the measured data received by the communication section, the hypothetical data acquisition section acquires the hypothetical data received by the communication section.

7. The property prediction device according to claim 5, wherein the property prediction device further possesses a hypothetical data complement section that complements the hypothetical data, the hypothetical data acquisition section acquires the hypothetical data complemented by the hypothetical data complement section.

8. The property prediction device according to claim 7, wherein the hypothetical data complement section, when a part of parameters having a value extremely small or large among a plurality of parameters included in the second manufacturing condition and the hypothetical value predicted based on the part of parameters are input to the input screen, complements the second manufacturing condition of the hypothetical data by calculating values of parameters other than the part of parameters among the plurality of parameters.

9. The property prediction device according to claim 1, wherein ​ The property prediction device further includes an uncertainty setting section that sets a first uncertainty of the property and a second uncertainty of the property, The model generation section generates the prediction model further based on the first uncertainty of the property and the second uncertainty of the property set by the uncertainty setting section.

10. The property prediction device according to claim 9, wherein The property prediction device further includes: a setting screen generation section that generates a setting screen for setting the first uncertainty of the property and the second uncertainty of the property, The uncertainty setting section sets the first uncertainty of the property and the second uncertainty of the property in accordance with an input operation to the setting screen generated by the setting screen generation section.

11. The property prediction device according to claim 1, wherein The model generation section generates the prediction model using Bayesian estimation, the first uncertainty is a first variance, and the second uncertainty is a second variance that is larger than the first variance.

12. The property prediction device according to claim 11, wherein The model generation section generates the prediction model in accordance with the following equation, [Math. 1] [Math. 2] 13. The property prediction device according to any one of claims 1 to 12, wherein The property prediction device further includes a manufacturing condition calculation section that calculates a third manufacturing condition of the sample based on the prediction model generated by the model generation section.

14. The property prediction device according to claim 13, wherein The third manufacturing condition includes at least one of a manufacturing condition at which the uncertainty of the property in the prediction model is largest and a manufacturing condition at which the uncertainty of the property in the prediction model is smallest.

15. The property prediction device according to claim 13, wherein The measured data acquisition section further acquires second measured data including the third manufacturing condition calculated by the manufacturing condition calculation section and a second measured value measured based on a property of the sample manufactured based on the third manufacturing condition, The model generation section updates the prediction model based on the second measured data acquired by the measured data acquisition section.

16. The property prediction device according to claim 15, wherein The model generation section does not update the prediction model in a case where the second measured value is an optimal value set.

17. A physical property prediction method characterized by, acquiring measured data including a first manufacturing condition of a sample and a measured value measured based on a property of the sample manufactured based on the first manufacturing condition; acquiring hypothetical data including a second manufacturing condition of the sample and a hypothetical value hypothetically obtained based on a property of the sample corresponding to the second manufacturing condition; and generating a prediction model of the property having a first uncertainty of the property with respect to the measured data and a second uncertainty of the property that is larger than the first uncertainty of the property with respect to the hypothetical data, based on the measured data and the hypothetical data acquired. for causing a computer to execute: ​ 18. A physical property prediction program, characterized by, ​ acquiring measured data including a first manufacturing condition of a sample and a measured value measured based on a property of the sample manufactured based on the first manufacturing condition; acquiring hypothetical data including a second manufacturing condition of the sample and a hypothetical value hypothetically obtained based on a property of the sample corresponding to the second manufacturing condition; and generating a prediction model of the property based on the acquired measured data and hypothetical data, the prediction model having a first uncertainty of the property for the measured data and a second uncertainty of the property greater than the first uncertainty of the property for the hypothetical data.

Images