Manufacturing condition selection method, material manufacturing method, manufacturing condition selection device, and program

Abstract

A manufacturing condition selection method according to the present invention comprises: an input step (S11) in which, with regard to a material to be manufactured, an initial data group including initial manufacturing condition data and characteristic value data of the material is input; a search range setting step (S12) in which a search range for a manufacturing condition of the material is set on the basis of the initial data group; a calculation step (S13) which uses a machine learning model to calculate a predicted characteristic value that is a predicted value of the characteristic value of the material in a search range on the basis of the initial data group, and a predicted error that is an estimated error of the predicted characteristic value; and a selection step (S14) in which a manufacturing condition of the material, the manufacturing condition being used as an experimental point, is selected on the basis of a calculated value of an acquisition function that uses a parameter based on the initial data group, the predicted characteristic value, and the predicted error.

Description

Manufacturing condition selection method, material manufacturing method, manufacturing condition selection device and program

[0001] The present disclosure relates to a manufacturing condition selection method, a material manufacturing method, a manufacturing condition selection device, and a program.

[0002] The properties of metal materials, for example, the mechanical properties of steel materials, are strongly dependent on the alloying elements and heat treatment conditions. For example, DP steel sheets (dual phase steel sheets) are made by adding alloying elements such as 0.1% by mass C, 1.0% by mass Si, and 1.0% by mass Mn to steel sheets, which are annealed at a temperature in the two-phase region of ferrite and austenite, and then tempered at a temperature of approximately 200 to 400 °C. This process achieves both high tensile strength and high ductility (high elongation). Furthermore, the higher the annealing temperature during DP steel sheet production, the higher the austenite volume fraction, resulting in a higher martensite volume fraction after quenching, which tends to result in a DP steel sheet with high tensile strength. Furthermore, the higher the tempering temperature and the longer the tempering time during DP steel sheet production, the more the martensite is tempered, and the strain in the martensite is relaxed, which tends to result in a DP steel sheet with low tensile strength and high elongation. Based on these general trends, steel materials are developed by adjusting manufacturing conditions such as annealing temperature, annealing time, tempering temperature, and tempering time to improve mechanical properties such as tensile strength and elongation. However, these manufacturing conditions do not have a clear linear relationship with the mechanical properties, making it difficult to predict the optimal manufacturing conditions. Furthermore, the optimal manufacturing conditions vary depending on the type and concentration of alloying elements. Therefore, in practice, a large number of experiments are conducted to comprehensively search for the optimal manufacturing conditions, which results in enormous costs for the development of steel materials. Therefore, technology for predicting material properties using machine learning has attracted attention.

[0003] For example, a method for developing materials using a technique called materials informatics is known. Materials informatics utilizes digital technology in materials development. Specifically, it predicts the manufacturing conditions for materials to obtain the desired properties from a vast amount of accumulated data on the manufacturing conditions and properties of materials.

[0004] Patent Document 1 also discloses a fabrication evaluation system that automatically performs sample fabrication, measurement, and optimization. The fabrication evaluation system in Patent Document 1 includes a fabrication device that fabricates a sample, a measurement device that measures material information representing the physical properties or structure of the sample fabricated by the fabrication device, and an estimation device connected to the fabrication device and the measurement device. The estimation device further includes an estimation unit that estimates optimal fabrication conditions based on a dataset containing the sample's fabrication conditions and material information, and a data addition unit that adds the fabrication conditions estimated by the estimation unit and material information measured on a sample fabricated according to the fabrication conditions to the dataset. Thus, the estimation unit can sequentially estimate fabrication conditions based on the added dataset. As a specific example, the results of an experiment to minimize the electrical resistance of a Nb-doped TiO thin film deposited on a glass substrate are disclosed. The film was deposited using a sputtering method, and two types of targets, Ti0.94Nb0.06O2 and Ti1.98Nb0.02O3, which have different Ti and O ratios, were used. By adjusting the mixture ratio of Ar gas to a mixture of Ar (99%) and O2 (1%), the oxygen content (oxygen partial pressure) in the thin film at which electrical resistance is minimized is sought. The lower confidence bound is used as the acquisition function (A), and the minimum value is sought as A = -E + 5σ (E: predicted value of the material information of the sample, σ: predicted error value). The upper and lower limits of the oxygen partial pressure, which correspond to the manufacturing conditions of the sample, are determined, and the sample is divided into a grid to search for the minimum value of the acquisition function. The minimum electrical resistance was found in 18 experiments.

[0005] Japanese Patent Application Laid-Open No. 2021-101465

[0006] Generally, materials development methods using materials informatics are based on the premise that a huge amount of data on the manufacturing conditions and properties of materials already exists, which results in a large number of experimental steps and low efficiency.

[0007] The method disclosed in Patent Document 1 uses an acquisition function (A) obtained by linearly adding a predicted value and a prediction error, and treats the probability (E) of achieving a predicted value as equivalent to the probability of achieving characteristics that are 5σ away from the predicted value. This results in a problem of increased experimental man-hours in searching for optimal sample manufacturing conditions, resulting in poor efficiency.

[0008] In view of the above circumstances, the purpose of the present disclosure is to provide a manufacturing condition selection method, a material manufacturing method, a manufacturing condition selection device, and a program that can efficiently select manufacturing conditions for improving the characteristic values ​​of a material.

[0009] [1] A manufacturing condition selection method according to one embodiment of the present disclosure includes: an input step for inputting an initial data group including initial manufacturing condition data and property value data of a material to be manufactured; a search range setting step for setting a search range for manufacturing conditions of the material based on the initial data group; a calculation step for calculating, using a machine learning model, a predicted property value that is a predicted value of a property value of the material in the search range based on the initial data group, and a prediction error that is an estimated error in the predicted property value; and a selection step for selecting manufacturing conditions for the material to be used as experimental points based on calculated values ​​of an acquisition function that uses parameters based on the initial data group, the predicted property value, and the prediction error.

[0010] [2] As one embodiment of the present disclosure, in [1], the method further includes: an acquisition step of acquiring characteristic values ​​of the material under the manufacturing conditions set as the experimental points; and a determination step of determining whether or not to select manufacturing conditions of the material as new experimental points by comparing calculated values ​​of the acquisition function before and after adding the manufacturing conditions set as the experimental points and the acquired characteristic values ​​of the material to the initial data group.

[0011] [3] As an embodiment of the present disclosure, in [1] or [2], the acquisition function is: A is the calculated value; R is a random number corresponding to the magnitude of the measurement error for correcting the measurement error when measuring the property value of the material; and y is the predicted property value. ａｖ , the error of the predicted characteristic value is y σ , the actual measured value of the characteristic value is Y, and the standard deviation of the actual measured value is Yσ , the maximum value of the actual measurement value is max(Y), and the parameter for correcting the prediction error is ε １ , ε ２ and ε ３ When this is the case, it is expressed by the following equation (1).

[0012] [4] A method for producing a material according to an embodiment of the present disclosure produces the material using production conditions selected by any one of the production condition selection methods [1] to [3].

[0013] [5] A manufacturing condition selection device according to an embodiment of the present disclosure includes: an input unit to which an initial data group including initial manufacturing condition data and property value data of a material to be manufactured is input; a search range setting unit that sets a search range for manufacturing conditions of the material based on the initial data group; a calculation unit that uses a machine learning model to calculate a predicted property value that is a predicted value of a property value of the material in the search range based on the initial data group and a prediction error that is an estimated error in the predicted property value; and a selection unit that selects manufacturing conditions for the material to be used as experimental points based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted property value, and the prediction error.

[0014] [6] A program according to an embodiment of the present disclosure causes a computer to function as: an input unit to which an initial data group including initial manufacturing condition data and property value data of a material to be manufactured is input; a search range setting unit that sets a search range for manufacturing conditions of the material based on the initial data group; a calculation unit that uses a machine learning model to calculate a predicted property value that is a predicted value of a property value of the material in the search range based on the initial data group and a prediction error that is an estimated error in the predicted property value; and a selection unit that selects manufacturing conditions for the material to be used as experimental points based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted property value, and the prediction error.

[0015] According to the present disclosure, it is possible to provide a manufacturing condition selection method, a material manufacturing method, a manufacturing condition selection device, and a program that can efficiently select manufacturing conditions for improving the characteristic values ​​of a material.

[0016] FIG. 1 is a schematic diagram showing an example configuration of a manufacturing condition selection system including a manufacturing condition selection device according to an embodiment of the present disclosure. FIG. 2 is a flowchart showing processing of a manufacturing condition selection method according to an embodiment of the present disclosure. FIG. 3 is a diagram showing distributions of characteristic values ​​based on measured values ​​obtained by an experiment in Example 1. FIG. 4 is a diagram showing the relationship between the number of steps and the maximum value of the characteristic value in Example 1. FIG. 5 is a diagram showing distributions of characteristic values ​​according to the number of steps in Example 1. FIG. 6 is a diagram showing distributions of characteristic values ​​according to the number of steps in Example 2. FIG. 7 is a diagram showing distributions of characteristic values ​​according to the number of steps in Comparative Example 1. FIG. 8A is a diagram showing distributions of multiple characteristic values ​​based on measured values ​​obtained by an experiment in Example 2. FIG. 8B is a diagram showing distributions of multiple characteristic values ​​based on measured values ​​obtained by an experiment in Example 2. FIG. 8C is a diagram showing distributions of multiple characteristic values ​​based on measured values ​​obtained by an experiment in Example 2. FIG. 9 is a diagram showing the relationship between the number of steps and the maximum value of the characteristic value in Example 2. FIG. 10 is a diagram showing distributions of characteristic values ​​according to the number of steps in Example 3. FIG. 11 is a diagram showing distributions of characteristic values ​​according to the number of steps in Example 4. 12 is a diagram showing the distribution of characteristic values ​​according to the number of steps in Comparative Example 2. FIG. 13 is a diagram showing the distribution of characteristic values ​​according to the number of steps in Comparative Example 3.

[0017] Hereinafter, a manufacturing condition selection method, a material manufacturing method, a manufacturing condition selection device, and a program according to embodiments of the present disclosure will be described with reference to the drawings. In each drawing, identical or corresponding parts are designated by the same reference numerals. In the following description of the embodiments, the description of identical or corresponding parts will be omitted or simplified as appropriate.

[0018] (Manufacturing Condition Selection System) FIG. 1 is a block diagram of a manufacturing condition selection system 1 including a manufacturing condition selection device 10 according to this embodiment. The manufacturing condition selection system 1 includes the manufacturing condition selection device 10. The manufacturing condition selection device 10 includes an input unit 11, an output unit 12, and a processing unit 13. The processing unit 13 includes a search range setting unit 14, a calculation unit 15, a selection unit 16, an acquisition unit 17, and a determination unit 18. The manufacturing condition selection system 1 may further include a process computer 30, as shown in the configuration example of FIG. 1 . The configuration of the manufacturing condition selection system 1 is not limited to that shown in FIG. 1 . For example, the manufacturing condition selection system 1 may further include a display (display device) that displays data from the output unit 12. The manufacturing condition selection system 1 may also include a server computer that aggregates data, together with or instead of the process computer 30. In this case, an operator may input data to the server computer via an input device (e.g., a mouse, a keyboard, etc.), and the input data may be transferred from the server computer to the input unit 11. Furthermore, the process computer 30 may be a computer positioned above a plurality of control computers, each of which controls a specific manufacturing device. In this case, data may be input to the input unit 11 directly from some or all of the control computers.

[0019] (Process Computer) The process computer 30 is a computer that controls manufacturing equipment on a line (production line) that manufactures a manufacturing object, or a computer that controls experimental conditions related to manufacturing in a lab (laboratory). The manufacturing object may be, for example, a metal material including steel, but is not limited to metal materials and may be a general material. The process computer 30 can also function as a data server that aggregates data related to manufacturing. Here, as described above, a server computer that aggregates data related to manufacturing may be used instead of the process computer 30. The data related to manufacturing includes, for example, manufacturing conditions and measurement values ​​obtained by measuring the manufactured material. The process computer 30 may be capable of communicating with the manufacturing condition selection device 10 via a network.

[0020] (Manufacturing Condition Selection Device) The manufacturing condition selection device 10 is a device that selects manufacturing conditions so as to improve the characteristic values ​​of the material to be manufactured. The characteristic values ​​are values ​​that represent the properties of the material, and may be, for example, strength, yield strength, elongation, hardness, or a value obtained by combining these in the case of a metal material.

[0021] (Input Unit) The input unit 11 is an input interface for the manufacturing condition selection device 10. An initial data group including initial manufacturing condition data and material characteristic value data for the material to be manufactured is input to the input unit 11. Here, "initial" indicates that the data has already been set, measured, or predicted. In other words, the initial data group includes initial manufacturing conditions, which are manufacturing conditions, and characteristic values ​​that have already been obtained by actual measurement or prediction (calculation) of the material. The input unit 11 executes an input process, which will be described later. The manufacturing conditions include not only conditions on the manufacturing line but also experimental conditions in the laboratory.

[0022] (Output Unit) The output unit 12 is an output interface of the manufacturing condition selection device 10. The output unit 12 may output the manufacturing conditions selected by the manufacturing condition selection device 10 to the process computer 30. The output unit 12 may also display the manufacturing conditions selected by the manufacturing condition selection device 10 on a display device such as various displays so that an operator of the manufacturing line can confirm the selected manufacturing conditions. The output unit 12 executes an output process, which will be described later.

[0023] (Processing Unit) The processing unit 13 performs calculations for the manufacturing condition selection device 10 to select manufacturing conditions. The processing unit 13 may also function as a control unit that controls the entire manufacturing condition selection device 10. The processing unit 13 may be one or more processors. The processor may be, for example, a general-purpose processor or a dedicated processor specialized for a specific process, but is not limited to these and may be any processor.

[0024] As described above, in this embodiment, the processing unit 13 includes the search range setting unit 14, the calculation unit 15, the selection unit 16, the acquisition unit 17, and the determination unit 18. The functions of the search range setting unit 14, the calculation unit 15, the selection unit 16, the acquisition unit 17, and the determination unit 18 may be realized by software. For example, one or more programs may be stored in a storage device accessible by the processing unit 13. When the program stored in the storage device is read by the processing unit 13, which is a processor, it may cause the processing unit 13 to function as the search range setting unit 14, the calculation unit 15, the selection unit 16, the acquisition unit 17, and the determination unit 18.

[0025] (Search Range Setting Unit) The search range setting unit 14 sets a search range for the manufacturing conditions of the material based on the initial data group. That is, the search range setting unit 14 sets a limit on the range of the manufacturing conditions to be selected. Details of the processing executed by the search range setting unit 14 will be described later. In addition, the search range setting unit 14 executes a search range setting step, which will be described later.

[0026] (Calculation unit) The calculation unit 15 uses a model to calculate a predicted property value, which is a predicted value of a property value of a material in a search range based on an initial data group, and a prediction error, which is an estimated error of the predicted property value. In this embodiment, the model is a machine learning model. Details of the process executed by the calculation unit 15 will be described later. The calculation unit 15 also executes a calculation step, which will be described later.

[0027] (Selection Unit) The selection unit 16 selects manufacturing conditions for a material to be used as an experimental point based on the calculated value of an acquisition function that uses parameters based on an initial data group, predicted characteristic values, and prediction errors. The experimental point is a specific manufacturing condition under which an experiment (manufacturing of a material in this embodiment) has been or will be performed. The experimental point selected by the selection unit 16 is preferably a manufacturing condition not included in the initial manufacturing condition data, i.e., a manufacturing condition that has not been used in an experiment so far. Details of the process performed by the selection unit 16 will be described later. The selection unit 16 also performs a selection process, which will be described later.

[0028] (Acquisition Unit) The acquisition unit 17 acquires the characteristic values ​​of the material under the manufacturing conditions set as the experimental points. Details of the process executed by the acquisition unit 17 will be described later. The acquisition unit 17 also executes an acquisition step, which will be described later.

[0029] (Determination Unit) The determination unit 18 determines whether or not to select further manufacturing conditions for the material to be used as new experimental points by comparing the calculated values ​​of the acquisition function before and after adding the manufacturing conditions set as experimental points and the acquired property values ​​of the material to the initial data group. Here, the fact that there is no need to select further manufacturing conditions for the material to be used as new experimental points is sometimes referred to as "pass." In other words, the determination unit 18 determines whether or not the result is "pass," meaning that no new experimental points need to be selected, or "fail," meaning that new experimental points need to be selected. Details of the processing performed by the determination unit 18 will be described later. The determination unit 18 also executes a determination process, which will be described later.

[0030] Here, the manufacturing condition selection device 10 is not limited to a specific device, but may be realized by a computer, for example. The computer may be, for example, a commercially available general-purpose computer. The computer includes, for example, a storage device such as a memory and a hard disk drive, a CPU, and an input / output device. The processing unit 13 may be realized by the CPU. The program read by the processing unit 13 may be stored in the storage device. Furthermore, the input unit 11 and the output unit 12 may be realized by the input / output device.

[0031] 2 is a flowchart showing the processing of the manufacturing condition selection method executed by the manufacturing condition selection device 10 according to this embodiment. The manufacturing condition selection method generally includes an input step, a search range setting step, a calculation step, and a selection step. The manufacturing condition selection method may further include an acquisition step and a determination step.

[0032] (Input Step) In the input step, an initial data group including initial manufacturing condition data and material characteristic value data is input (step S11).

[0033] As described above, the initial data group is a data group that includes existing initial manufacturing condition data and characteristic value data of materials obtained under those manufacturing conditions.

[0034] The manufacturing condition data are control variables used to manufacture the material to be manufactured. For example, if the material to be manufactured is a DP steel plate, the manufacturing condition data include the chemical composition, annealing temperature, annealing time, tempering temperature, etc.

[0035] The material characteristic value data is the characteristic value of the material to be improved. For example, if the material to be manufactured is a DP steel plate, the material characteristic value data includes the yield strength, tensile strength, elongation value, etc.

[0036] The number of initial data groups may be one or more sets. A set is a combination of manufacturing conditions and the property values ​​of the material manufactured under those manufacturing conditions. In order to efficiently predict manufacturing conditions that will result in better property values, the number of initial data groups is preferably two or more sets, and more preferably four or more sets. It is preferable that the number of sets of the initial data groups be as large as possible.

[0037] When the number of sets in the initial data group is small, in order to efficiently predict (i.e., with fewer experimental steps) manufacturing conditions that will result in better characteristic values, it is preferable that the initial data group include sets that correspond to the ends (near the boundaries) of the search range described below.

[0038] (Search Range Setting Step) In the search range setting step, a search range for the manufacturing conditions of the material is set based on the initial data group (step S12).

[0039] Specifically, the upper and lower limits of a search range for searching manufacturing conditions for a material to be used as an experimental point are set so that the characteristic values ​​of the material are improved over those of the initial data group. By setting the search range, it is possible to efficiently predict the manufacturing conditions (experimental points) for the material to achieve excellent characteristic values.

[0040] The calculated value of the acquisition function used in the selection process described below has the property that the further the data points deviate from the manufacturing conditions of the data points input as the initial data group, the larger the value. If a search range is not set, the property of this acquisition function may result in the selection of experimental points that are significantly different from the manufacturing conditions of the data points input as the initial data group, which may result in divergence of the calculation. Therefore, a search range is set. The upper and lower limits of the search range may be set based on the initial manufacturing condition data, and may also be set taking into account conventional knowledge (e.g., widely known theories or knowledge in materials manufacturing).

[0041] For example, in the development of steel materials, the search range may be set so that an annealing temperature that is so low that carbon diffusion cannot occur and as a result the material structure does not change at all is not selected. Also, for example, in the development of a DP steel plate with an excellent balance between strength and ductility, upper and lower limit values ​​of the search range for the manufacturing conditions such as the composition and annealing temperature are set. For example, based on conventional knowledge such as literature on DP steel plates, A is set as the lower limit of the annealing temperature. １ A point is selected as the upper limit of the annealing temperature. ４ A point may be selected. １ It is known that if the annealing temperature is below the A point, the phase transformation to austenite does not occur, so the structure does not change at all, and an improvement in the balance between strength and ductility due to annealing cannot be expected. ４ If the annealing temperature exceeds the A point, δ-ferrite will precipitate. ４ It is known that, even if the structure is converted to full austenite by subsequent cooling, the structure becomes full martensite by quenching, and an improvement in the balance between strength and ductility cannot be expected. Thus, by eliminating conditions that do not result in improved properties based on conventional knowledge of materials, it is possible to predict the manufacturing conditions for materials that efficiently achieve excellent property values. Therefore, it is preferable to select experimental conditions within a reasonable range based on material knowledge. However, when conducting research in areas that deviate from conventional knowledge, i.e., common knowledge in materials systems, it is acceptable to set a search range without being bound by conventional knowledge.

[0042] Here, the search range may be a continuous space, but it is preferably a discrete space in order to achieve an improvement in the characteristic values ​​of the material with a small number of experimental steps. When the search range is a discrete space, the number of manufacturing conditions (search points) to be searched for can be reduced compared to when the search range is a continuous space.

[0043] (Calculation Step) In the calculation step, the initial data group is used to calculate the predicted property values ​​and prediction errors of the material in the search range using a model (step S13).

[0044] Specifically, using the initial data set input in the input step, the manufacturing conditions of the material, which are control variables, are changed within the search range set in the search range setting step to calculate predicted property values ​​and prediction errors for the material. For example, in the case of a DP steel plate, predicted property values ​​of the material, such as yield strength and tensile strength, are calculated for control variables such as chemical composition, annealing temperature, and annealing time.

[0045] A model such as a machine learning model may be used to calculate the predicted property values ​​of the material. The machine learning model may be generated using a neural network model, a Gaussian process model, a random forest, or the like. For example, multiple sets of manufacturing conditions and property values ​​of materials manufactured under those manufacturing conditions may be extracted from actual manufacturing data collected in the process computer 30 and used as learning data. The model used to calculate the predicted property values ​​of the material may be an interpolation function generated based on a group of initial data, such as a known linear approximation or spline interpolation. When multiple models are used in the calculation process, the average value of the predicted property values ​​calculated using the multiple models may be used as the final predicted property value.

[0046] Furthermore, an error (prediction error) in the calculated predicted property value of the material is calculated. For example, the prediction error may be the standard deviation of predicted property values ​​calculated using multiple models. Alternatively, for example, the prediction error may be the standard deviation of predicted property values ​​calculated using one model with different combinations of input initial data groups. Furthermore, when a prediction error is obtained from one model, such as in calculations using a Gaussian process model, the obtained prediction error may be used as is.

[0047] (Selection Step) In the selection step, manufacturing conditions for a material to be used as an experiment point are selected based on the calculated value of the acquisition function obtained using the initial data group, the predicted property values ​​of the material, and the prediction error (step S14).

[0048] The selected experimental point is, for example, a manufacturing condition that is an optimal solution where the acquisition function exhibits a maximum or minimum value (maximum value in this embodiment). However, the selected experimental point may be a manufacturing condition where the acquisition function exhibits a suboptimal solution (e.g., the top 10% or more of the acquisition function). For example, for a problem with a small measurement error, selecting an experimental point where an experiment has already been conducted is generally meaningless and inefficient. Therefore, if the optimal solution where the acquisition function exhibits a maximum or minimum value matches an experimental point where an experiment has already been conducted, the suboptimal solution may be selected as the experimental point. However, for a problem with a large measurement error, if the optimal solution matches an experimental point where an experiment has already been conducted, the optimal solution may be selected as the experimental point to confirm reproducibility.

[0049] The acquisition function generally uses UCB (Upper Confidence Bound) or PI (Probability of Improvement). In the manufacturing condition selection method according to this embodiment, the acquisition function shown in Equation (1) is used to improve the efficiency of selection. A is the calculated value of the acquisition function.

[0050]

[0051] Here, R is a random number. More specifically, R is a parameter for correcting measurement errors when measuring the characteristic values ​​of a material, and is a random number corresponding to the magnitude of the measurement error. ａｖ is the predicted value (predicted characteristic value). σ is the error of the predicted value (predicted characteristic value). Y is the actual measured value of the characteristic value. Y σ is the standard deviation of the measured values. max(Y) is the maximum value of the measured values. ε １ , ε ２ and ε ３ is a parameter. In detail, ε １ , ε ２ and ε ３ is a parameter that corrects the prediction error.

[0052] For example, in steel materials, there can generally be a measurement error of about 1% in total elongation. Therefore, when total elongation is used as a characteristic value, the measurement error can be corrected by generating a random number R in the range of 0.99 to 1.01. If the measurement error is small, R may be set to 1.0.

[0053] Also, ε １ , ε ２ and ε ３ is preferably 0.01 or more, more preferably 0.1 or more, and even more preferably 0.5 or more. １ , ε ２ and ε ３ If ε is set to less than 0.1, the influence of the prediction error becomes excessively small, and the number of experimental steps required to make a pass / fail judgment, which will be described later, increases. １ , ε ２ and ε ３ is preferably 10.0 or less, more preferably 5.0 or less. １ , ε ２ and ε ３ If it is set to be greater than 10.0, the influence of the prediction error becomes excessively large, and the number of experimental steps required to make a pass / fail judgment, which will be described later, increases.

[0054] When two or more characteristic values ​​need to be improved, the manufacturing condition selection method according to this embodiment preferably uses the acquisition functions shown in equations (3) and (4).

[0055]

[0056] Here, i is each parameter corresponding to the characteristic value i. ｉ The magnitude of the absolute value of the obtained acquisition function also changes depending on the magnitude of the characteristic value. Therefore, if an acquisition function is calculated by multiplying acquisition functions corresponding to each characteristic value, the influence of characteristic values ​​with large absolute values ​​will be large, and characteristic values ​​with small absolute values ​​may not be taken into account. Therefore, when the acquisition function A corresponding to each characteristic value is calculated, ｉ The maximum value max(A ｉ ) and then calculate the acquisition function A.

[0057] Here, R ｉis a random number corresponding to the characteristic value i. ｉ is a parameter for correcting the measurement error when measuring the characteristic value i of the material, and is a random number according to the magnitude of the measurement error. ａｖ，ｉ is the predicted value (predicted characteristic value) of characteristic value i. σ，ｉ is the error of the predicted value (predicted characteristic value) of characteristic value i. ｉ is the measured value of characteristic value i. σ，ｉ is the standard deviation of the measured values ​​of characteristic value i. ｉ ) is the maximum measured value of characteristic value i. １，ｉ , ε ２，ｉ and ε ３，ｉ are parameters corresponding to the characteristic value i. １，ｉ , ε ２，ｉ and ε ３，ｉ is a parameter that corrects the prediction error.

[0058] Here, for example, in steel materials, there may generally be a measurement error of about 1% in the total elongation. Therefore, when the total elongation is used as a characteristic value, R ｉ The measurement error can be corrected by generating a random number in the range of 0.99 to 1.01 as R. ｉ may be set to 1.0.

[0059] Also, ε １，ｉ , ε ２，ｉ and ε ３，ｉ is preferably 0.01 or more, more preferably 0.1 or more, and even more preferably 0.5 or more. １，ｉ , ε ２，ｉ and ε ３，ｉ If ε is set to less than 0.1, the influence of the prediction error becomes excessively small, and the number of experimental steps required to make a pass / fail judgment, which will be described later, increases. １，ｉ , ε ２，ｉ and ε ３，ｉ is preferably 10.0 or less, more preferably 5.0 or less. １，ｉ , ε ２，ｉ and ε ３，ｉ If it is set to be greater than 10.0, the influence of the prediction error becomes excessively large, and the number of experimental steps required to make a pass / fail judgment, which will be described later, increases.

[0060] (Acquisition Step) In the acquisition step, the characteristic values ​​of the material under the manufacturing conditions of the material to be used as the experiment point selected in the selection step are acquired (step S15).

[0061] In the acquisition step, the data points obtained in the selection step, i.e., the selected experimental points and the property values ​​(data sets) of the material corresponding to these experimental points, are added to the initial data group. In order to improve the prediction accuracy of the property values, a machine learning model may be generated (updated) using the initial data group to which the data sets have been added.

[0062] In the acquisition step, it is preferable that a material is actually manufactured at the selected experimental points (manufacturing conditions) and the property values ​​of the manufactured material are actually measured. In other words, it is preferable that the property values ​​added to the initial data group are actually measured values. However, the property values ​​added to the initial data group may be predicted property values ​​calculated using a machine learning model or the like.

[0063] (Determination process) In the determination process, it is determined whether or not to select manufacturing conditions for a material that will become a new experimental point based on the calculated values ​​of the acquisition function obtained before and after adding the selected experimental point (manufacturing conditions) and the characteristic values ​​acquired in the acquisition process to the initial data group (step S16).

[0064] The calculated value of the acquisition function obtained from the initial data group input in the input step using equation (1) is (A １ ) is calculated. The calculation value of the acquisition function obtained by using the formula (1) from the initial data group after adding the data points representing the characteristic values ​​and manufacturing conditions obtained in the acquisition process is (A ２ In the determination step, (A １ ) and (A ２ ) to determine whether or not to search for manufacturing conditions for the next experimental point.

[0065] For example, for the manufacturing conditions in the search range, (A １ ) and (A ２) is calculated, and the difference is divided by the number of data points (the number of search points) to obtain a normalized difference. If the normalized difference is less than the threshold, a sufficient improvement in the characteristic value has been obtained, and the test is determined to be pass (Yes in step S17), and a search for a new experimental point is not performed. The threshold is 0.001, for example, but is not limited to a specific value and may be set depending on, for example, the type of material being tested. If the normalized difference is equal to or greater than the threshold, the test is determined to be fail (No in step S17), and a search for a new experimental point is performed. In other words, if the test is failing, the process returns to the input process (step S11), and the series of processes is executed again. Here, if the test is failing, the initial data group input in the input process will be the initial data group after the data set has been added in the acquisition process.

[0066] (Output Step) If the determination step determines that the product is pass (Yes in step S17), the output step is executed (step S18). The output step may output the pass / fail result of the determination step, as well as the characteristic values ​​obtained in the acquisition step and the manufacturing conditions thereof.

[0067] The manufacturing condition selection method according to the present embodiment may be realized by installing software for executing the processing of the manufacturing condition selection method on a computer, or by installing the software on a cloud computer on a network.

[0068] In this way, the manufacturing condition selection method executed by the manufacturing condition selection device 10 according to this embodiment can efficiently select manufacturing conditions for improving the characteristic values ​​of a material. The manufacturing condition selection method may be executed as part of a manufacturing method for a material to be manufactured. For example, the process computer 30 may acquire the manufacturing conditions selected by the manufacturing condition selection method executed by the manufacturing condition selection device 10, and control the manufacturing of the material using the acquired manufacturing conditions.

[0069] EXAMPLES The effects of the present disclosure will be specifically described below based on examples, but the present disclosure is not limited to these examples. In the examples, the manufacturing condition selection method using the manufacturing condition selection device 10 described above was performed.

[0070] (Example 1) In Example 1, the material to be manufactured is mild steel, which is required to be soft and elongate. In Example 1, manufacturing conditions were selected to achieve one characteristic of mild steel that is more soft and elongate. The soft and elongate characteristic is determined to be better when the ratio (El / YS), which is the ratio of the total elongation (El) to the yield strength (YS), which are characteristic values ​​of the material, is greater. The manufacturing conditions for mild steel were changed in tempering temperature and tempering time. There were four combinations of tempering temperature and tempering time (conditions 1 to 4) as follows, and test materials 1 to 4 were manufactured under each of the conditions.

[0071] In the first condition, (normalized tempering temperature, normalized tempering time) was set to (-0.3, -0.3). In the second condition, (normalized tempering temperature, normalized tempering time) was set to (-0.3, -0.1). In the third condition, (normalized tempering temperature, normalized tempering time) was set to (0.3, 0.1). In the fourth condition, (normalized tempering temperature, normalized tempering time) was set to (0.3, 0.3). In order to efficiently find the maximum value of the characteristics, these initial points were set to include the edges of the search space, which will be described later.

[0072] Here, the normalized tempering temperature is the normalized tempering temperature. Normalization was performed by determining the average value and standard deviation of the tempering temperatures at 30 experimental points in the mild steel database obtained so far, subtracting the determined average value from the tempering temperature, and dividing by the standard deviation. Furthermore, the normalized tempering time is the normalized tempering time. Normalization was performed by determining the average value and standard deviation of the tempering times at 30 experimental points in the mild steel database obtained so far, subtracting the determined average value from the tempering time, and dividing by the standard deviation.

[0073] JIS No. 5 test pieces 1 to 4 were manufactured for test materials 1 to 4, and tensile tests were performed to measure the yield strength (YS) and total elongation (El) of test materials 1 to 4. The manufacturing conditions of the obtained materials and the characteristic value (El / YS) of the material manufactured under those manufacturing conditions were used as the initial data group (corresponding to step S11). The characteristic value (El / YS) was also normalized. In normalizing (El / YS), El was first calculated by finding the average value of 30 El experimental points in the mild steel database and dividing by that average value to calculate the normalized El (norm.El). Similarly, YS was calculated by dividing the average value of 30 YS experimental points to calculate the normalized YS (norm.YS). Subsequently, norm.El / norm.YS was calculated to normalize (El / YS). Hereinafter, all characteristic values ​​(El / YS) will be referred to as this normalized norm. El / norm. YS.

[0074] Next, a search range was set (corresponding to step S12). In this embodiment, the ends of the normalized tempering temperature and normalized tempering time in the initial data group were set as the upper and lower limits of the search range for the normalized tempering temperature and normalized tempering time. That is, the upper limit of the search range for the normalized tempering temperature was set to 0.3, and the lower limit was set to -0.3. Also, the upper limit of the search range for the normalized tempering time was set to 0.3, and the lower limit was set to -0.3.

[0075] In addition, in order to prevent an increase in the number of experimental steps, the search range was set to be a discrete space divided into increments of 0.01 for the normalized tempering temperature and the normalized tempering time.

[0076] Next, the predicted characteristic values ​​and prediction errors were calculated (corresponding to step S13). The predicted characteristic values ​​and prediction errors were calculated using the initial data set and a Gaussian process model.

[0077] Next, the initial data set and the obtained predicted characteristic values ​​and prediction errors are used to calculate the acquisition function (A １ The calculated value of the acquisition function (A １ The manufacturing conditions under which the value of σ (σ) becomes the maximum value were selected as the experimental points (corresponding to step S14). Hereinafter, the selected manufacturing conditions may be referred to as "manufacturing conditions (1)."

[0078] Next, the characteristic value (El / YS) under the manufacturing conditions of the selected experimental point was obtained (corresponding to step S15). Here, the obtained characteristic value (El / YS) may be referred to as "characteristic value (1)" hereinafter. The characteristic value (1) here was obtained from FIG. 3 shown below.

[0079] FIG. 3 is a diagram showing the distribution of characteristic values ​​based on measured values. The distribution in FIG. 3 was calculated by creating a spline interpolation function based on 30 characteristic values ​​(El / YS) measured in an experiment. In FIG. 3, the characteristic maximum value, which is the maximum value of the characteristic value (El / YS), was calculated to be "1.16." Furthermore, in FIG. 3 and in FIGS. 5 to 7 described below, the distribution of characteristic values ​​is shown as a graph with the horizontal axis representing the normalized tempering temperature and the vertical axis representing the normalized tempering time. In FIG. 3, the coordinates (normalized tempering temperature, normalized tempering time) corresponding to the characteristic maximum value are (-0.11, 0.14). Here, since there is no measurement error, the R of the acquisition function defined by Equation (1) is set to 1.0.

[0080] Next, the manufacturing conditions (1) and characteristic values ​​(1) of the obtained material are newly added to the initial data group, and the calculated value (A) of the acquisition function is obtained. ２ ) and the calculated value of the above acquisition function (A １ ), it was determined whether or not to search for the manufacturing conditions to be used as the next experimental point (corresponding to step S16). Specifically, for the manufacturing conditions in the search range, (A １ ) and (A ２ ) was calculated, and the difference was divided by the data points to calculate a normalized difference. Thereafter, if the normalized difference was 0.001 or more (above the threshold), the sample was determined to be unacceptable (corresponding to No in step S17), and the characteristic value (1) and the manufacturing condition (1) were added to the initial data group, and the processes corresponding to steps S11 to S16 were executed. Thereafter, the processes corresponding to steps S11 to S16 were repeatedly executed until the normalized difference became less than 0.001 (below the threshold).

[0081] In the invention example 1, the acquisition function of the formula (1) used in the selection process is ε １ 1.0, ε ２ and ε ３In Example 2, the acquisition function of the formula (1) used in the selection process was set to 2.0. １ 1.0, ε ２ and ε ３ was set to 0.1.

[0082] In Comparative Example 1, UCB (Equation (2)) is used as the acquisition function, and the coefficient term of the prediction error value (ε Ｕ ) was set to 6.0. ａｖ is the predicted value (predicted characteristic value), and y σ is the error of the predicted value (predicted characteristic value).

[0083]

[0084] Figure 4 shows the relationship between the number of steps and the maximum characteristic value when manufactured under the manufacturing conditions proposed by this model. The characteristic values ​​here were obtained using a function interpolated with the spline function shown in Figure 3. The number of steps refers to the number of times the process from the input step to the judgment step was executed, and can also be referred to as the number of trials. In Figure 4, for example, if the number of steps is 6, the maximum characteristic value obtained in steps 1 to 6 is shown. While it took 12 steps (12 steps) to reach the maximum characteristic value in Comparative Example 1, Inventive Example 1 approached the maximum characteristic value in 4 steps (4 steps) and reached the maximum characteristic value in 8 steps (8 steps). Inventive Example 2, the maximum characteristic value was reached in 9 steps (9 steps), indicating that the inventive examples have fewer steps than the comparative examples.

[0085] The distribution of predicted characteristic values ​​according to the number of steps is shown for Example 1 in Fig. 5, Example 2 in Fig. 6, and Comparative Example 1 in Fig. 7. Each figure shows the cases where the number of steps is 4 (4 steps) and 9 (9 steps). In each figure, experimental points are indicated by dots.

[0086] In Example 1, the maximum characteristic value was found quickly and there was little bias in the selection of experimental points. In addition, Example 1 was able to largely reproduce the distribution in Figure 3, which was given as the correct answer.

[0087] In Example 2, although slower than Example 1, the maximum characteristic value was found after 9 iterations (9 steps).

[0088] On the other hand, in Comparative Example 1, it was confirmed that the experimental points were biased and an efficient search was not possible.

[0089] As described above, in this example, it was possible to appropriately select manufacturing conditions for a material having better characteristic values ​​than those of the initial material. Furthermore, it was possible to select manufacturing conditions for a material that efficiently maximizes characteristic values ​​with a small number of experimental steps. Because manufacturing conditions for a material having better characteristic values ​​can be efficiently selected, it is expected that efficiency will be improved in, for example, material development. In this example, in the acquisition step (corresponding to step S15), characteristic values ​​were acquired from a spline interpolation function created based on 30 characteristic values ​​(El / YS). Here, the properties may be maximized using a similar procedure by actually creating samples at the selected experimental points and conducting tensile tests on the samples to obtain the yield strength (YS) and total elongation (El).

[0090] (Example 2) In Example 2, the material to be manufactured is a high-tensile steel that requires high tensile strength and high ductility (high elongation value). In Example 2, manufacturing conditions for a high-tensile steel that can achieve both of these two properties, high tensile strength and high elongation value, were selected. The manufacturing conditions for the high-tensile steel were varied by changing the annealing temperature and tempering temperature. The following two combinations of annealing temperature and tempering temperature (conditions 5 and 6) were used, and test materials 5 and 6 were manufactured under each of these conditions.

[0091] In the fifth condition, (normalized annealing temperature, normalized tempering temperature) was set to (-2.36, -1.41). In the sixth condition, (normalized annealing temperature, normalized tempering temperature) was set to (1.41, 1.41). These initial points were set to include the edges of the search space, which will be described later, in order to efficiently find the maximum value of the characteristics.

[0092] Here, the normalized annealing temperature is the normalized annealing temperature. Normalization was performed by finding the average value of the annealing temperatures at 25 experimental points in the database of high-tensile steels obtained up to that point, subtracting the obtained average value from the annealing temperature, and dividing by the standard deviation. Furthermore, the normalized tempering temperature is the normalized tempering temperature. Normalization was performed by finding the average value of the tempering temperatures at 25 experimental points in the database of high-tensile steels obtained up to that point, subtracting the obtained average value from the tempering temperature, and dividing by the standard deviation.

[0093] For test materials 5 and 6, JIS No. 5 test pieces 5 and 6 were manufactured and subjected to tensile tests to measure the tensile strength (TS) and total elongation (El) of test materials 5 and 6. The manufacturing conditions of the obtained materials and the characteristic values ​​TS and El of the materials manufactured under those manufacturing conditions were used as an initial data group (corresponding to step S11). The characteristic values ​​TS and El were also normalized in the same way as the annealing temperature and tempering temperature. Below, the characteristic values ​​TS and El are each shown as normalized numerical values.

[0094] Next, a search range was set (corresponding to step S12). In Example 2, the ends of the normalized annealing temperature and normalized tempering temperature in the initial data group were set as the upper and lower limits of the search range for the normalized annealing temperature and normalized tempering temperature. That is, the upper limit of the search range for the normalized annealing temperature was set to 1.41, and the lower limit was set to -2.36. Also, the upper limit of the search range for the normalized tempering temperature was set to 1.41, and the lower limit was set to -1.41.

[0095] In addition, in order to suppress an increase in the number of experimental steps, the search range was set to be a discrete space in which the normalized annealing temperature and the normalized tempering temperature were divided into increments of 0.236 and 0.141, respectively.

[0096] Next, predicted characteristic values ​​and prediction errors were calculated for TS and El (corresponding to step S13). The predicted characteristic values ​​and prediction errors were calculated using the initial data set by a Gaussian process model.

[0097] Next, the initial data set and the predicted characteristic values ​​and prediction errors of TS and El obtained are used to calculate the acquisition function (A １ The calculated value of the acquisition function (A１ The manufacturing conditions under which the value of σ (σ) becomes the maximum value were selected as the experimental points (corresponding to step S14). Hereinafter, the selected manufacturing conditions may be referred to as "manufacturing conditions (1)."

[0098] Next, the characteristic values ​​of tensile strength (TS) and total elongation (El) under the manufacturing conditions of the selected experimental point were obtained (corresponding to step S15). Hereinafter, the obtained characteristic values ​​may be referred to as "characteristic value (1)." Characteristic value (1) was obtained from FIGS. 8A and 8B shown below.

[0099] 8A to 8C are diagrams showing the distribution of characteristic values ​​based on the measured values ​​in Example 2, where Fig. 8A shows normalized TS, Fig. 8B shows normalized El, and Fig. 8C shows normalized TS×El obtained by multiplying TS and El. Normalized TS×El is normalized by subtracting the average value of TS×El obtained after multiplying TS and El and dividing by its standard deviation.

[0100] The distributions in Figures 8A to 8C were calculated by creating a spline interpolation function based on the 25 characteristic values ​​(TS and El) measured in the experiment. In Figures 8A to 8C and Figures 10 to 13 described below, the distribution of characteristic values ​​is shown as a graph with the normalized annealing temperature on the horizontal axis and the normalized tempering temperature on the vertical axis. In Figure 8A, the maximum value of TS was calculated to be "1.83," and the coordinates (normalized annealing temperature, normalized tempering temperature) corresponding to the characteristic maximum value at this time were (-0.94, 0.14). In Figure 8B, the maximum value of El was calculated to be "2.30," and the coordinates (normalized annealing temperature, normalized tempering temperature) corresponding to the characteristic maximum value at this time were (0.71, 0.57). Furthermore, in Figure 8C, the maximum value of TS x El was calculated to be "2.7," and the coordinates (normalized annealing temperature, normalized tempering temperature) corresponding to the characteristic maximum value at this time were (0.71, 0.57). Here, since there is no measurement error, the R of the acquisition function defined by Equation (4) ｉ was set to 1.0.

[0101] Next, the manufacturing conditions (1) and characteristic values ​​(1) of the obtained material are newly added to the initial data group, and the calculated value (A) of the acquisition function is obtained. ２ ) and the calculated value of the above acquisition function (A １), it was determined whether or not to search for the manufacturing conditions to be used as the next experimental point (corresponding to step S16). Specifically, for the manufacturing conditions in the search range, (A １ ) and (A ２ ) and divide the difference by the data points to calculate the normalized difference. －６ If the difference was greater than or equal to the threshold, the product was judged to be unacceptable (corresponding to No in step S17), and the characteristic value (1) and the manufacturing condition (1) were added to the initial data group, and the processes corresponding to steps S11 to S16 were executed. After that, if the normalized difference was 1.0×10 －６ The processes corresponding to steps S11 to S16 were repeatedly executed until the difference became less than the threshold value.

[0102] In Example 3, the acquisition functions of Equation (3) and Equation (4) used in the selection process are １，ｉ 1.0, ε ２，ｉ and ε ３，ｉ In Example 4, the acquisition functions of Equation (3) and Equation (4) used in the selection process were set to 2.0. １，ｉ 0.5, ε ２，ｉ and ε ３，ｉ was set to 2.0.

[0103] In Comparative Example 2, UCB (Equation (5) and Equation (6)) is used as the acquisition function, and the coefficient term of the prediction error value (ε Ｕ，ｉ ) was set to 0.0. In Comparative Example 3, UCB (Equation (5)) was used as the acquisition function, and the coefficient term (ε Ｕ，ｉ ) was set to 6.0. ａｖ，ｉ is the predicted value (predicted characteristic value) of characteristic value i, and y σ，ｉ is the error in the predicted value (predicted characteristic value) of characteristic value i. Also, max(UCBi) is the maximum value of the acquisition function UCBi corresponding to characteristic value i.

[0104]

[0105] FIG. 9 shows the relationship between the number of steps and the maximum value of the characteristic value TS×El when manufactured under the manufacturing conditions proposed by this model. The characteristic value TS×El is calculated by multiplying TS and El calculated using the spline interpolation functions shown in FIGS. 8A and 8B, respectively. Performance can be evaluated based on whether the characteristic value TS×El reaches its maximum value. The number of steps is the number of times the process from the input step to the judgment step is executed, and can also be referred to as the number of trials. In FIG. 9, for example, if the number of steps is 6, the maximum value of the characteristic value obtained in steps 1 to 6 is shown. Comparative Example 2 took 40 steps (40 steps) to reach the maximum characteristic value, and Comparative Example 3 did not find the maximum characteristic value. In contrast, Example 3 approached the maximum characteristic value in 13 steps (13 steps) and reached the maximum characteristic value in 17 steps (17 steps). In Example 4, the maximum characteristic value was reached in 21 steps (21 steps), indicating that the number of steps in the inventive examples is fewer than that in the comparative examples.

[0106] FIG. 10 shows the distribution of predicted characteristic values ​​according to the number of steps for Example 3. FIG. 11 shows the distribution of predicted characteristic values ​​according to the number of steps for Example 4. FIG. 12 shows the distribution of predicted characteristic values ​​according to the number of steps for Comparative Example 2. FIG. 13 shows the distribution of predicted characteristic values ​​according to the number of steps for Comparative Example 3. Each figure shows cases where the number of steps is 17 (17 steps), 21 (21 steps), and 41 (41 steps). In Example 3 of FIG. 10, processing is completed before the 21st step, so figures for the 21st step and the 41st step are not shown. In Example 4 of FIG. 11, processing is completed before the 41st step, so figures for the 41st step are not shown. 13, no improvement in the characteristic values ​​was observed even after 41 repetitions (41 steps), so the graph for 41 repetitions (41 steps) is not shown. Also, the experimental points are indicated by dots in each graph.

[0107] In Example 3, the maximum characteristic value was found quickly, and the distribution of FIG. 8C given as the correct answer was largely reproduced.

[0108] In Example 4, although slower than Example 3, the maximum characteristic value was found after 21 iterations (21 steps).

[0109] On the other hand, in Comparative Examples 2 and 3, the experimental points were biased, and it was confirmed that efficient search was not possible.

[0110] As described above, in Example 2, it was possible to appropriately select manufacturing conditions for a material having two characteristic values ​​superior to those of the initial material. Furthermore, it was possible to select manufacturing conditions for a material that efficiently maximizes the characteristic values ​​with a small number of experimental steps. Because manufacturing conditions for a material having superior characteristic values ​​can be efficiently selected, it is expected that efficiency will be improved in, for example, material development. In this example, in the acquisition step (corresponding to step S15), the characteristic values ​​were acquired from a spline interpolation function created based on 25 characteristic values. Here, the characteristics may be maximized using a similar procedure by actually creating samples at the selected experimental points and conducting tensile tests on the samples to obtain the tensile strength (TS) and total elongation (El).

[0111] Although the embodiments of the present disclosure have been described based on the drawings and examples, it should be noted that those skilled in the art would easily be able to make various modifications or alterations based on the present disclosure. Therefore, it should be noted that these modifications and alterations are included within the scope of the present disclosure. For example, the functions included in each component or step can be rearranged so as not to cause logical inconsistencies, and multiple components or steps can be combined or divided into one. The embodiments of the present disclosure can also be realized as a storage medium on which a program executed by a processor included in an apparatus is recorded. It should be understood that these are also included within the scope of the present disclosure.

[0112] REFERENCE SIGNS LIST 1 Manufacturing condition selection system 10 Manufacturing condition selection device 11 Input unit 12 Output unit 13 Processing unit 14 Search range setting unit 15 Calculation unit 16 Selection unit 17 Acquisition unit 18 Determination unit 30 Process computer

Claims

1. an input step of inputting an initial data group including initial manufacturing condition data and characteristic value data of the material as a manufacturing object; a search range setting step of setting a search range for manufacturing conditions of the material based on the initial data group; a calculation step of calculating a predicted characteristic value, which is a predicted value of a characteristic value of the material in the search range based on the initial data group, and a prediction error, which is an estimated error of the predicted characteristic value, using a machine learning model; a selection step of selecting manufacturing conditions for the material to be used as experimental points based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted characteristic value, and the prediction error.

2. an acquisition step of acquiring characteristic values ​​of the material under the manufacturing conditions set as the experimental points; 2. The manufacturing condition selection method according to claim 1, further comprising a determination step of determining whether or not to select manufacturing conditions for the material that will become new experimental points by comparing calculated values ​​of the acquisition function before and after the manufacturing conditions set as the experimental points and the acquired characteristic values ​​of the material are added to the initial data group.

3. The acquisition function is a function of A, a random number R according to the magnitude of the measurement error for correcting the measurement error when measuring the property value of the material, and y ａｖ , the error of the predicted characteristic value is y σ , the actual measured value of the characteristic value is Y, and the standard deviation of the actual measured value is Y σ , the maximum value of the actual measurement value is max(Y), and the parameter for correcting the prediction error is ε １ , ε ２ and ε ３ The manufacturing condition selection method according to claim 1 or 2, which is represented by the following formula (1) when [0010]

4. A method for producing a material, comprising the steps of: producing the material using production conditions selected by the method for selecting production conditions according to claim 1 or 2.

5. an input unit for inputting an initial data group including initial manufacturing condition data and characteristic value data of the material as a manufacturing object; a search range setting unit that sets a search range for manufacturing conditions of the material based on the initial data group; A calculation unit that calculates a predicted characteristic value that is a predicted value of a characteristic value of the material in the search range based on the initial data group using a machine learning model, and a prediction error that is an estimated error of the predicted characteristic value; a selection unit that selects manufacturing conditions for the material to be used as experimental points based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted characteristic value, and the prediction error.

6. Computer, an input unit for inputting an initial data group including initial manufacturing condition data and characteristic value data of the material as a manufacturing object; a search range setting unit that sets a search range for manufacturing conditions of the material based on the initial data group; A calculation unit that calculates a predicted characteristic value that is a predicted value of a characteristic value of the material in the search range based on the initial data group using a machine learning model, and a prediction error that is an estimated error of the predicted characteristic value; A program for functioning as a selection unit that selects manufacturing conditions for the material to be used as an experimental point based on a calculated value of an acquisition function that uses parameters based on the initial data group, the predicted characteristic value, and the prediction error.

Images