Design assistance device, program, and design assistance method

Abstract

This design assistance device comprises: a data shaping unit that shapes data of a first dataset including manufacturing information and evaluation information pertaining to a first product, and a second dataset including manufacturing information and evaluation information pertaining to a second product, so that physical property information included in the evaluation information is used as an objective variable; a first training unit that preliminarily trains a machine learning model by using the first dataset subjected to the data shaping; a second training unit that performs transfer training of the preliminarily trained machine learning model by using the second dataset subjected to the data shaping; and an optimization unit that obtains manufacturing information pertaining to the second product satisfying desired physical property information by using the transfer-trained machine learning model.

Description

Design support device, program, and design support method

[0001] The present disclosure relates to a design support device, a program, and a design support method.

[0002] In recent years, there has been an increasing interest in materials informatics (MI) that analyzes a vast amount of material data by machine learning to efficiently design a material composition with desired performance.

[0003] For example, in Patent Document 1, explanatory variable data includes information such as raw materials of an adhesive composition and conditions for measuring physical property data of the adhesive composition, and a technique for performing machine learning including the influence of the measurement conditions on the physical property data is disclosed.

[0004] Japanese Patent Application Laid-Open No. 2024-59737

[0005] When implementing product development using a machine learning model, a machine learning model obtained by machine learning a dataset including product manufacturing information and evaluation information is used. When implementing new product development with a shape and evaluation information different from those of an existing product, it has been difficult to utilize a dataset of an existing product with different shape and evaluation information or a machine learning model obtained by machine learning a dataset of an existing product.

[0006] Therefore, when implementing new product development using a machine learning model, there has been a problem that the number of datasets used for machine learning is insufficient in the initial stage of development, and the effect of improving development efficiency becomes negligible. Note that Patent Document 1 does not solve the above problems.

[0007] An object of the present disclosure is to provide a design support device, a program, and a design support method for improving the development efficiency of a new product implemented using a machine learning model.

[0008] The present disclosure has the following configuration.

[0009] [1] A design support device comprising: a data formatting unit that formats a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information becomes the target variable; a first learning unit that pre-trains a machine learning model using the first dataset from which the data formatting has been performed; a second learning unit that transfer-learns the pre-trained machine learning model using the second dataset from which the data formatting has been performed; and an optimization unit that obtains the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer-learned model has been performed.

[0010] [2] The design support apparatus according to [1], wherein the data formatting unit decomposes the evaluation information of the first dataset into measurement condition information and physical property information, formats the decomposed measurement condition information and the manufacturing information of the first dataset as input data and the decomposed physical property information as output data, and decomposes the evaluation information of the second dataset into measurement condition information and physical property information, formats the decomposed measurement condition information and the manufacturing information of the second dataset as input data and the decomposed physical property information as output data.

[0011] [3] The design support apparatus according to [2], wherein the data formatting unit, when the evaluation information of the first dataset includes multiple types of measurement condition information, formats the data such that the measurement condition information and the manufacturing information of the first dataset are input data and the physical property information are output data for each of the multiple types of measurement condition information, and when the evaluation information of the second dataset includes multiple types of measurement condition information, formats the data such that the measurement condition information and the manufacturing information of the second dataset are input data and the physical property information is output data for each of the multiple types of measurement condition information.

[0012] [4] The design support apparatus according to [2] or [3], wherein the optimization unit obtains, by multi-objective optimization, the manufacturing information of the second product that satisfies the desired physical property information with respect to the multiple types of measurement condition information when the evaluation information of the second dataset includes multiple types of measurement condition information.

[0013] [5] A design support device according to any one of [1] to [4], further comprising: a prediction unit that predicts the physical property information of the second product from the manufacturing information of the second product using the machine learning model that has been transferred to learn the machine learning model.

[0014] [6] A design support device according to any one of [2] to [4], wherein the first product and the second product differ in shape or in the measurement condition information included in the evaluation information.

[0015] [7] A program that causes the information processing device to function as: a data formatting unit that formats a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information is used as the target variable; a first learning unit that pre-trains a machine learning model using the first dataset from which the data formatting has been performed; a second learning unit that transfer-learns the pre-trained machine learning model using the second dataset from which the data formatting has been performed; and an optimization unit that obtains the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer-learned model has been performed.

[0016] [8] A design support method executed by a design support device, comprising: a data shaping step of shaping a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information is used as the target variable; a first learning step of pre-training a machine learning model using the first dataset from which the data shaping has been performed; a second learning step of transfer learning the pre-trained machine learning model using the second dataset from which the data shaping has been performed; and an optimization step of obtaining the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer learning has been performed.

[0017] According to this disclosure, it is possible to improve the efficiency of developing new products using machine learning models.

[0018] This is a diagram illustrating an example of the design support system 1 according to this embodiment. This is a diagram illustrating an example of the computer according to this embodiment. This is a diagram illustrating an example of the design support system 1 according to this embodiment. This is a diagram illustrating an example of the dataset for product A. This is a diagram illustrating an example of the dataset for product B. This is a diagram illustrating an example of machine learning using the dataset for product A shown in Figure 4 and the dataset for product B shown in Figure 5. This is a flowchart illustrating an example of the processing of the design support device 10 according to this embodiment. This is an explanatory diagram illustrating an example of the processing of decomposing the evaluation information contained in the datasets for product A and product B into measurement condition information and physical property information. This is a diagram illustrating an example of data formatting for the dataset for product A. This is a diagram illustrating an example of data formatting for the dataset for product B. This is an explanatory diagram illustrating an example of the dataset for product A that has undergone the data formatting shown at the bottom of Figure 9. This is an explanatory diagram illustrating an example of the dataset for product B that has undergone the data formatting shown at the bottom of Figure 10. This is an example graph showing the relationship between the number of training data and prediction accuracy in a machine learning model that has not undergone transfer learning and a machine learning model that has undergone transfer learning. This is a diagram illustrating an example of the optimization process in step S16. This is a flowchart illustrating an example of the processing in step S10. This is a flowchart illustrating an example of the processing in step S10. This is a flowchart illustrating an example of the pre-training in step S12 and the transfer learning in step S14. This is a flowchart of an example of the optimization process in step S16.

[0019] Next, embodiments of the present invention will be described in detail. However, the present invention is not limited to the following embodiments.

[0020] [First Embodiment] <System Configuration> Figure 1 is a configuration diagram of an example of a design support system 1 according to this embodiment. The design support system 1 in Figure 1 has a design support device 10 and a worker terminal 12. The design support device 10 and the worker terminal 12 are connected via a communication network 18 such as a local area network (LAN) or the Internet, enabling data communication.

[0021] The worker terminal 12 is an information processing device such as a PC, tablet, or smartphone operated by the worker. The worker terminal 12 displays a screen and accepts operations from the worker. The worker terminal 12 transmits the content of the operations received from the worker to the design support device 10. The worker terminal 12 displays the results of the processing by the design support device 10 on its screen for the worker to confirm.

[0022] The design support device 10 is an information processing device such as a PC or workstation that performs processing to support product design based on the content of the operation received from the worker. The design support device 10 supports product design using a machine learning model that has been trained on a dataset including product manufacturing information and evaluation information. The design support device 10 may display the processing results on a display device such as the display of the worker terminal 12 for the worker to confirm, or it may display them on the display device of the design support device 10 for the worker to confirm.

[0023] In the design support system 1 according to this embodiment, the first dataset of the first product and the second dataset of the second product are reshaped so that the target variable is common. A machine learning model, which has been pre-trained using a sufficient amount of the reshaped first dataset, is then transferred and trained using the reshaped second dataset. In the design support system 1 according to this embodiment, the development efficiency of the second product is improved by obtaining manufacturing information for the second product that satisfies the desired physical property information, or by predicting physical property information from the manufacturing information of the second product, using the machine learning model that has been transferred and trained.

[0024] It should be noted that the design support system 1 in Figure 1 is merely an example, and there are various system configurations depending on the application and purpose. For example, the design support device 10 may be implemented using multiple computers, or it may be implemented as a cloud computing service. Furthermore, the design support system 1 in Figure 1 may be implemented using a standalone computer.

[0025] <Hardware Configuration> The design support device 10 and the worker terminal 12 in Figure 1 are implemented by a computer 500 with the hardware configuration shown in Figure 2, for example.

[0026] Figure 2 is a hardware configuration diagram of an example of a computer according to this embodiment. The computer 500 shown in Figure 2 includes an input device 501, a display device 502, an external interface 503, RAM 504, ROM 505, CPU 506, a communication interface 507, and an HDD 508, and each of these is interconnected via bus B. Note that the input device 501 and the display device 502 may be connected and used via the external interface 503.

[0027] The input device 501 includes a touch panel, operation keys and buttons, a keyboard and mouse, etc., used by the operator to input various signals. The display device 502 consists of a display such as a liquid crystal or organic EL that displays the screen, and a speaker that outputs sound data such as voice and sound.

[0028] The communication interface 507 is an interface for the computer 500 to perform data communication. The HDD 508 is an example of a non-volatile storage device that stores programs and data. The programs and data stored include the OS, which is the basic software that controls the entire computer 500, and applications that provide various functions on the OS. The computer 500 may also use a drive device that uses flash memory as a storage medium (for example, a solid-state drive: SSD, etc.) instead of the HDD 508.

[0029] The external I / F 503 is an interface to an external device. An external device may be a recording medium 503a, etc. This allows the computer 500 to read from and / or write to the recording medium 503a via the external I / F 503. The recording medium 503a may be a flexible disk, CD, DVD, SD memory card, or USB memory, etc.

[0030] ROM 505 is an example of a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. ROM 505 stores programs and data such as the BIOS, OS settings, and network settings that are executed when the computer 500 starts up. RAM 504 is an example of a volatile semiconductor memory (storage device) that temporarily holds programs and data.

[0031] The CPU 506 is an arithmetic unit that controls and implements the functions of the entire computer 500 by reading programs and data from storage devices such as ROM 505 and HDD 508 onto RAM 504 and executing processing. In this embodiment, the computer 500 can implement various functions of the design support device 10 and the worker terminal 12 described later by executing programs. The CPU 506 may also read and execute programs from the recording medium 503a where the programs are stored via an external I / F 503.

[0032] <Functional Configuration> The configuration of the design support system 1 according to this embodiment will now be described. Figure 3 is a functional configuration diagram of an example of the design support system 1 according to this embodiment. Note that parts of the configuration diagram in Figure 3 that are not necessary for the explanation of this embodiment have been appropriately omitted.

[0033] The design support device 10 of the design support system 1 shown in Figure 3 has a configuration comprising a request receiving unit 20, a response transmission unit 22, a data formatting unit 24, a first learning unit 26, a second learning unit 28, an optimization unit 30, a prediction unit 32, a dataset storage unit 40, a machine learning model storage unit 42, and a data formatting information storage unit 44. The worker terminal 12 has a configuration comprising an information display unit 50, an operation reception unit 52, a request transmission unit 54, and a response receiving unit 56.

[0034] The information display unit 50 of the worker terminal 12 displays a screen on the display device 502 that supports product development by the worker. The operation reception unit 52 receives various operations from the worker. The request transmission unit 54 transmits processing requests to the design support device 10 based on the various operations received from the worker. The response reception unit 56 receives responses from the design support device 10 to the processing requests transmitted by the request transmission unit 54.

[0035] The request receiving unit 20 of the design support device 10 receives processing requests from the worker terminal 12 based on the operator's operations. The response transmission unit 22 transmits a response to the processing request received by the request receiving unit 20 to the worker terminal 12.

[0036] The data formatting unit 24 formats the dataset containing product manufacturing information and evaluation information as described below. The data formatting unit 24 formats the first dataset for the first product and the second dataset for the second product. In this embodiment, an example is described in which a first dataset with a sufficient amount for machine learning and a second dataset with a smaller amount than the first dataset and insufficient to improve the accuracy of machine learning are used.

[0037] The first learning unit 26 pre-trains a machine learning model using the first dataset of the first product, which has undergone data formatting. The second learning unit 28 transfers the machine learning model, which has been pre-trained on the first dataset, using the second dataset of the second product, which has also undergone data formatting. Pre-training is a learning phase aimed at acquiring general knowledge of the first product. Transfer learning is a phase in which the knowledge of the first product, which has been pre-trained, is adapted to the new second product. In transfer learning, the machine learning model pre-trained using the first dataset of the first product is fine-tuned using the second dataset to obtain a machine learning model adapted to the second product. Transfer learning is, for example, fine-tuning.

[0038] The optimization unit 30 uses a transfer-learned machine learning model to obtain manufacturing information for a second product that satisfies the desired physical property information. The prediction unit 32 uses a transfer-learned machine learning model to predict the physical property information of the second product from the manufacturing information of the second product. The dataset storage unit 40 stores the first dataset and the second dataset. The machine learning model storage unit 42 stores the machine learning model. The data formatting information storage unit 44 stores the data formatting information necessary for data formatting performed by the data formatting unit 24.

[0039] <Dataset>The dataset according to this embodiment will be described. In the following, it will be described using Product A which is an example of the first product and Product B which is an example of the second product. FIG. 4 is a diagram for explaining an example of the dataset of Product A. FIG. 5 is a diagram for explaining an example of the dataset of Product B.

[0040] FIGS. 4 and 5 show examples of Product A and Product B as friction materials (brake pads). Product A and Product B are manufactured through blending a plurality of materials, molding, and heat treatment. The shapes shown in FIGS. 4 and 5 indicate the shapes of Product A and Product B respectively. As shown in FIGS. 4 and 5, the shapes of Product A and Product B are different.

[0041] Manufacturing information includes, as information, the blending of materials and the manufacturing conditions. Evaluation information is the evaluation index of Product A and Product B. When Product A and Product B are friction materials, the evaluation indexes of Product A and Product B are the coefficient of kinetic friction, wear, noise, etc. which indicate the evaluation performance of Product A and Product B. The evaluation information shown in FIGS. 4 and 5 indicates measurement condition information and physical property information.

[0042] For example, the labels (such as 50°C 1EFF) of the evaluation information in FIGS. 4 and 5 indicate measurement condition information. The label of the evaluation information can be converted into measurement condition information using the data shaping information stored in the data shaping information storage unit 44. The data shaping information is information indicating the correspondence between the label of the evaluation information and the measurement condition information.

[0043] For example, the label "50°C 1EFF" of the evaluation information indicates that it is the coefficient of kinetic friction μ at the first braking time and a temperature of 50°C. As shown in FIGS. 4 and 5, the measurement condition information of Product A and Product B is different.

[0044] Also, the values (such as 0.4) of the evaluation information in FIGS. 4 and 5 indicate physical property information. The value "0.4" corresponding to the label "50°C 1EFF" of the evaluation information indicates the coefficient of kinetic friction μ which is an example of physical property information.

[0045] FIG. 6 is a diagram for explaining an example of machine learning using the dataset of Product A shown in FIG. 4 and the dataset of Product B shown in FIG. 5.

[0046] As shown in FIG. 6, when performing machine learning with manufacturing information as an explanatory variable and evaluation information as an objective variable, since the objective variables of the data sets of Product A and Product B are different, it is difficult to use the machine learning model obtained by performing machine learning on the data set of Product A for optimizing the formulation of Product B or predicting physical property information. Further, since the shapes of Product A and Product B are different, it is difficult to use the machine learning model obtained by performing machine learning on the data set of Product A for optimizing the formulation of Product B or predicting physical property information.

[0047] Therefore, in the present embodiment, as described below, a machine learning model used for optimizing the formulation of Product B or predicting physical property information is machine-learned using the data sets of Product A and Product B that have been data-formatted.

[0048] <Processing> FIG. 7 is a flowchart of an example of the processing of the design support device 10 according to the present embodiment.

[0049] In step S10, the data formatting unit 24 of the design support device 10 formats the data sets of Product A and Product B so that the physical property information included in the evaluation information is used as the objective variable. Here, the processing of step S10 will be described using FIGS. 8 to 11A and FIG. 11B.

[0050] FIG. 8 is an explanatory diagram of an example of the process of decomposing the evaluation information included in the data sets of Product A and Product B into measurement condition information and physical property information.

[0051] As shown in FIG. 8, the names described in the labels of the evaluation information of the data sets of Product A and Product B can be converted into measurement condition information using the data formatting information stored in the data formatting information storage unit 44.

[0052] In FIG. 8, the name "50°C 1EFF" of the label of the evaluation information on the left side is converted into measurement condition information of the number of braking times "first time" and temperature "50°C", and as shown in the evaluation information on the right side, it can be decomposed into measurement condition information and physical property information. In FIG. 8, an example is shown in which the measurement condition information is the number of braking times and the temperature, and the physical property information is the coefficient of kinetic friction μ. The coefficient of kinetic friction μ is an example of general physical property information (physical property value).

[0053] Furthermore, in Figure 8, the label name "100℃1EFF" in the evaluation information on the left is converted into measurement condition information of the number of brakings "1st time" and temperature "100℃", and as shown in the evaluation information on the right, it can be decomposed into measurement condition information and physical property information.

[0054] Figure 9 shows an example of data formatting for the product A dataset. The product A dataset shown at the top of Figure 9 includes multiple types of measurement condition information in the evaluation information. The label name for the evaluation information is "5 th "Fade" can be converted to measurement condition information of initial velocity "30", number of braking strokes "5", and temperature "300". The label name "50℃1EFF" in the evaluation information can be converted to measurement condition information of initial velocity "10", number of braking strokes "1", and temperature "50". The label name "100℃1EFF" in the evaluation information can be converted to measurement condition information of initial velocity "10", number of braking strokes "1", and temperature "100".

[0055] The data formatting unit 24 converts the labels of the evaluation information in the product A dataset shown in the upper part of Figure 9 into measurement condition information, and decomposes the evaluation information into measurement condition information and physical property information. In the example in Figure 9, the decomposed measurement condition information is information on initial velocity, number of braking strokes, and temperature. The physical property information is the coefficient of dynamic friction.

[0056] The data formatting unit 24 formats the data set of product A shown in the upper part of Figure 9, using the manufacturing information and decomposed measurement condition information as input data (explanatory variables) and the decomposed physical property information as output data (dependent variable). The lower part of Figure 9 shows an example of the product A data set that has been formatted to use the manufacturing information and decomposed measurement condition information as input data and the decomposed physical property information as output data.

[0057] Figure 10 shows an example of data formatting for the product B dataset. The product B dataset shown at the top of Figure 10 includes multiple types of measurement condition information in the evaluation information. The label name for the evaluation information is "50 th"Burnish" can be converted into measurement condition information of initial velocity "20", number of braking strokes "50", and temperature "50". The label name of the evaluation information "50km / h initial μ" can be converted into measurement condition information of initial velocity "50", number of braking strokes "1", and temperature "30". The label name of the evaluation information "100km / h initial μ" can be converted into measurement condition information of initial velocity "100", number of braking strokes "1", and temperature "30".

[0058] The data formatting unit 24 converts the labels of the evaluation information in the product B dataset shown in the upper part of Figure 10 into measurement condition information, and decomposes the evaluation information into measurement condition information and physical property information. In the example of Figure 10, the decomposed measurement condition information is information on initial velocity, number of braking strokes, and temperature. The physical property information is the coefficient of dynamic friction.

[0059] The data formatting unit 24 formats the data so that the manufacturing information and decomposed measurement condition information of the product B dataset shown in the upper part of Figure 10 are input data (explanatory variables), and the decomposed physical property information is output data (dependent variable). The lower part of Figure 10 shows an example of the product B dataset after it has been formatted so that the manufacturing information and decomposed measurement condition information of the product B dataset shown in the upper part of Figure 10 are input data, and the decomposed physical property information is output data.

[0060] Figure 11A is an explanatory diagram illustrating an example of the dataset for product A after the data formatting shown below Figure 9. Figure 11B is an explanatory diagram illustrating an example of the dataset for product B after the data formatting shown below Figure 10.

[0061] As shown in Figures 11A and 11B, the datasets for product A and product B, after data formatting, share a common target variable. As shown in Figures 11A and 11B, in this embodiment, machine learning can be performed on product A and product B using manufacturing information and measurement condition information as explanatory variables and the common physical property information as the target variable.

[0062] Returning to step S12 in Figure 7, the first learning unit 26 pre-trains a machine learning model using the product A dataset shown in Figure 11A, for example, after data formatting. Pre-training is a machine learning process that aims to acquire general knowledge about product A using the product A dataset.

[0063] In step S14, the second learning unit 28 performs transfer learning on the machine learning model pre-trained in step S12 using a dataset of product B, for example, shown in Figure 11B, which has been processed. Transfer learning is a machine learning method that applies the knowledge of product A, pre-trained in step S12, to product B. Transfer learning can utilize, for example, a neural network. Transfer learning may also be achieved, for example, by fine-tuning all layers or partially fine-tuning of the neural network. According to the transfer learning in step S14, for example, knowledge of product A, which has a different shape but other common information, can be applied to the knowledge of product B.

[0064] Figure 12 is an example graph showing the relationship between the number of training data points and prediction accuracy for machine learning models that have not undergone transfer learning and for machine learning models that have undergone transfer learning.

[0065] Method 1 shows the relationship between the number of training data points and prediction accuracy in a machine learning model that has not undergone transfer learning. Method 2 shows the relationship between the number of training data points and prediction accuracy in a machine learning model that has undergone transfer learning. As shown in Figure 12, Method 2, which uses a machine learning model that has undergone transfer learning, has better initial accuracy than Method 1, which uses a machine learning model that has not undergone transfer learning.

[0066] In step S16, the optimization unit 30 obtains manufacturing information for product B that satisfies the desired physical property information using the machine learning model that underwent transfer learning in step S14. The optimization process in step S16 is performed, for example, as shown in Figure 13. Figure 13 is a diagram illustrating an example of the optimization process in step S16.

[0067] As shown in the upper part of Figure 13, in this embodiment, even if the evaluation information includes multiple types of measurement condition information, the data is formatted into a machine learning model with the coefficient of dynamic friction μ, which is an example of physical property information, as the target variable, and manufacturing information and measurement condition information (initial velocity, number of braking cycles, and temperature, etc.) as explanatory variables.

[0068] The optimization process in step S16 involves substituting the values ​​of the measurement condition information in the dataset as fixed values, as shown in the lower part of Figure 13, and then performing multi-objective optimization or single-objective optimization to determine the material formulation x1 to x that satisfies the target dynamic friction coefficient μ. ｎ To obtain.

[0069] The optimization process in step S16 can utilize a variety of optimization methods, including simulated annealing (SA), genetic algorithms (GA), and random search. The appropriate method should be selected depending on the type of machine learning model used for transfer learning and the optimization problem setting. For example, if there are many target variables (e.g., more than 4), an example of GA such as NSGA-III may be used.

[0070] Figure 14 is a flowchart of an example of the process in step S10. The input and output dataset shown as an example in Figure 14 uses "Variable 1" to "Variable 3" as input data and "Physical property X under condition A" as output data. The "Physical property X under condition A" in the output data represents the physical property "X" at a temperature of "100 degrees" and a processing time of "10 minutes". Figure 14 shows the case where the target variable is "1".

[0071] In step S20, the data formatting unit 24 decomposes the output data label (output label) into measurement condition information and physical property information. For example, in Figure 14, the output label "Physical property X under condition A" is decomposed into measurement condition information "Temperature 100 degrees, processing time 10 minutes" and physical property information "X".

[0072] In step S22, the data formatting unit 24 adds the measurement condition information decomposed in step S20 to the input data.

[0073] In step S24, the data formatting unit 24 can format the dataset shown in the upper part of Figure 14 into the dataset shown in the lower part of Figure 14 by removing measurement condition information from the output data.

[0074] The process in step S10 may be carried out as shown in Figure 15. Figure 15 is a flowchart of an example of the process in step S10. In the input and output dataset shown as an example in Figure 15, "Variable 1" to "Variable 3" are input data, and "Physical property X under condition A" and "Physical property X under condition B" are output data.

[0075] The "Physical Property X under Condition A" in the output data represents the physical property "X" at a temperature of "100 degrees" and a processing time of "10 minutes". The "Physical Property X under Condition B" in the output data represents the physical property "X" at a temperature of "150 degrees" and a processing time of "15 minutes". Figure 15 shows the case where the number of dependent variables is "2", but the same applies when the number of dependent variables is "3" or more.

[0076] In the flowchart of Figure 15, the output label "Physical property X under condition A" is referred to as output label (1). Also, in the flowchart of Figure 15, the output label "Physical property X under condition B" is referred to as output label (2). Furthermore, in the flowchart of Figure 15, the output data is divided into a dataset with output label (1) and a dataset with output label (2), and processing is performed. Steps S30 to S34 are the processing for the dataset with output label (1). Steps S36 to S40 are the processing for the dataset with output label (2).

[0077] In step S30, the data formatting unit 24 decomposes the output label (1) into measurement condition information and physical property information. For example, in Figure 15, the output label (1) is decomposed into measurement condition information "temperature 100 degrees, processing time 10 minutes" and physical property information "X".

[0078] In step S32, the data formatting unit 24 adds the measurement condition information decomposed in step S30 to the input data.

[0079] In step S34, the data formatting unit 24 can format the dataset shown in the upper part of Figure 15 into the dataset shown in the middle part of Figure 15 by removing measurement condition information from the output data.

[0080] In step S36, the data formatting unit 24 decomposes the output label (2) into measurement condition information and physical property information. For example, in Figure 15, the output label (2) is decomposed into measurement condition information "temperature 150 degrees, processing time 15 minutes" and physical property information "X".

[0081] In step S38, the data formatting unit 24 adds the measurement condition information decomposed in step S36 to the input data.

[0082] In step S40, the data formatting unit 24 can format the dataset shown in the upper part of Figure 15 into the dataset shown in the middle part of Figure 15 by removing measurement condition information from the output data.

[0083] In step S42, the data formatting unit 24 determines whether the output label (1) and output label (2) after data formatting are the same. If the output label (1) and output label (2) after data formatting are the same, the data formatting unit 24 proceeds to step S44 and combines the dataset of output label (1) and the dataset of output label (2) after data formatting to obtain the dataset shown in the lower part of Figure 15.

[0084] If the output labels (1) and (2) after data formatting are not the same, the data formatting unit 24 proceeds to step S46 and does not combine the dataset of output label (1) and the dataset of output label (2) after data formatting.

[0085] Figure 16 is a flowchart of an example of the pre-training process in step S12 and the transfer learning process in step S14. In the flowchart of Figure 16, the pre-formatted dataset used for pre-training is referred to as dataset (1). Also, in the flowchart of Figure 16, the pre-formatted dataset used for transfer learning is referred to as dataset (2). Dataset (1) and dataset (2) are assumed to have the same labels for their input and output data.

[0086] In step S50, the first learning unit 26 pre-trains a machine learning model using the dataset (1). In step S52, the first learning unit 26 evaluates the pre-trained machine learning model (pre-trained model) and obtains a pre-trained model with an acceptable prediction accuracy.

[0087] In step S54, the second learning unit 28 transfers the pre-trained model using the dataset (2). In step S56, the second learning unit 28 evaluates the transferred machine learning model (transfer-trained model) and obtains a transfer-trained model with acceptable prediction accuracy.

[0088] According to the flowchart in Figure 16, a pre-trained model, which has been pre-trained using dataset (1), can be fine-tuned using dataset (2) to obtain a transfer-trained model adapted to the products corresponding to dataset (2).

[0089] Figure 17 is a flowchart of an example of the optimization process in step S16. The target dataset shown as an example in Figure 17 represents the case where the number of target variables is "1".

[0090] The trained model for the target dataset is the transfer-trained model explained using Figure 16. In step S60, the optimization unit 30 substitutes the values ​​of the measurement conditions corresponding to the output data (physical properties X) into the measurement conditions (temperature and processing time) of the trained model for the target dataset.

[0091] In step S62, the optimization unit 30 uses the trained model into which the measurement condition values ​​were substituted in step S60 to optimize "variable 1" to "variable 3" so that the physical property X under condition A approaches the target value. The processing in step S62 can use simulated annealing, random search, or a genetic algorithm, etc.

[0092] Figure 18 is a flowchart of an example of the optimization process in step S16. The target dataset shown as an example in Figure 18 represents the case where the number of target variables is "2" or more.

[0093] The trained model for the target dataset is the transfer-trained model explained using Figure 16. In step S70, the optimization unit 30 obtains a trained model for condition A by substituting the values ​​of the measurement conditions (condition A) corresponding to the output data (physical property X) into the measurement conditions (temperature and processing time) of the trained model for the target dataset. The optimization unit 30 also obtains a trained model for condition B by substituting the values ​​of the measurement conditions (condition B) corresponding to the output data (physical property X) into the measurement conditions of the trained model for the target dataset.

[0094] In step S72, the optimization unit 30 performs multi-objective optimization on "variable 1" to "variable 3" so that the physical property X in the trained model for condition A and the trained model for condition B obtained in step S70 approaches the target value.

[0095] According to this embodiment, the utilization of datasets and machine learning models across products with different shapes and evaluation metrics is promoted, and the efficiency of developing new products using machine learning models can be improved. For example, in the early stages of product B's development, the number of datasets for product B may be small, but since the dataset for product A, which has a sufficient number of datasets, can be utilized, the development efficiency of product B is improved.

[0096] [Other Embodiments] In this embodiment, the prediction unit 32 can predict the physical properties of product B from the manufacturing information of product B using a machine learning model that has undergone transfer learning. For this reason, for example, the optimization unit 30 may generate comprehensive search points for material formulations randomly or with a predetermined step size in the inverse problem analysis process, and obtain a material formulation that satisfies the target physical properties from the physical properties of the product corresponding to each comprehensive search point.

[0097] The material composition information and manufacturing conditions of a new product developed by the design support system 1 according to this embodiment may be used, for example, as input information to a manufacturing apparatus that manufactures the product according to the material composition information and manufacturing conditions. The manufacturing apparatus can manufacture the product by blending multiple materials according to the material composition information and manufacturing conditions.

[0098] Although this embodiment has been described above, it will be understood that various modifications to the form and details are possible without departing from the spirit and scope of the claims. Although the present invention has been described above based on examples, the present invention is not limited to the above examples, and various modifications are possible within the scope described in the claims. This application claims priority to Basic Application No. 2024-195408 filed with the Japan Patent Office on November 7, 2024, the entire contents of which are incorporated herein by reference.

[0099] 1 Design support system 10 Design support device 12 Operator terminal 18 Communication network 24 Data formatting unit 26 First learning unit 28 Second learning unit 30 Optimization unit 32 Prediction unit

Claims

1. A design support device comprising: a data formatting unit that formats a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information becomes the target variable; a first learning unit that pre-trains a machine learning model using the first dataset from which the data formatting has been performed; a second learning unit that transfer-learns the pre-trained machine learning model using the second dataset from which the data formatting has been performed; and an optimization unit that obtains the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer-learning has been performed.

2. The design support apparatus according to claim 1, wherein the data formatting unit decomposes the evaluation information of the first dataset into measurement condition information and physical property information, formats the decomposed measurement condition information and the manufacturing information of the first dataset as input data and the decomposed physical property information as output data, and decomposes the evaluation information of the second dataset into measurement condition information and physical property information, formats the decomposed measurement condition information and the manufacturing information of the second dataset as input data and the decomposed physical property information as output data.

3. The design support apparatus according to claim 2, wherein the data formatting unit, when the evaluation information of the first dataset includes multiple types of measurement condition information, formats the data such that the measurement condition information and the manufacturing information of the first dataset are input data and the physical property information are output data for each of the multiple types of measurement condition information, and when the evaluation information of the second dataset includes multiple types of measurement condition information, formats the data such that the measurement condition information and the manufacturing information of the second dataset are input data and the physical property information is output data for each of the multiple types of measurement condition information.

4. The design support apparatus according to claim 2 or 3, wherein the optimization unit obtains, by multi-objective optimization, the manufacturing information of the second product that satisfies the desired physical property information with respect to the multiple types of measurement condition information when the evaluation information of the second dataset includes multiple types of measurement condition information.

5. The design support device according to any one of claims 1 to 4, further comprising: a prediction unit that predicts the physical property information of the second product from the manufacturing information of the second product using the machine learning model that has undergone transfer learning.

6. The design support device according to any one of claims 2 to 4, wherein the first product and the second product differ in shape or in the measurement condition information included in the evaluation information.

7. A program that causes an information processing device to function as: a data formatting unit that formats a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information becomes the target variable; a first learning unit that pre-trains a machine learning model using the first dataset from which the data formatting has been performed; a second learning unit that transfer-learns the pre-trained machine learning model using the second dataset from which the data formatting has been performed; and an optimization unit that obtains the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer-learned model has been performed.

8. A design support method executed by a design support device, comprising: a data shaping step of shaping a first dataset containing manufacturing information and evaluation information of a first product and a second dataset containing manufacturing information and evaluation information of a second product so that the physical property information included in the evaluation information is used as the target variable; a first learning step of pre-training a machine learning model using the first dataset from which the data shaping has been performed; a second learning step of transfer learning the pre-trained machine learning model using the second dataset from which the data shaping has been performed; and an optimization step of obtaining the manufacturing information of the second product that satisfies the desired physical property information using the machine learning model from which the transfer learning has been performed.

Images