Training data generation device, training data generation method, and training data generation program

Abstract

The present invention reduces imbalance in simulation patterns when generating training data for constructing a surrogate model. This training data generation device generates training data for constructing a surrogate model, and has: a determination unit that randomly determines each of an initial value of input data and a final value of the input data; a path generation unit that generates a path by randomly allocating, to each step in a path from the determined initial value to the determined final value, the amount of change that is possible in said step; and a training data generation unit that generates training data including the determined initial value and final value, and the amount of change allocated to each step in the generated path.

Description

Learning data generation device, learning data generation method, and learning data generation program

[0001] The present disclosure relates to a training data generation device, a training data generation method, and a training data generation program.

[0002] In recent years, the use of predictive processing using surrogate models has been promoted as an alternative to physical simulation. A surrogate model is a model constructed by learning the simulation results when a physical simulation is performed using, for example, a simulation device that reproduces a manufacturing process. By performing predictive processing using the surrogate model, it is possible to significantly reduce calculation costs compared to performing physical simulation using a simulation device.

[0003] Here, when performing physical simulations and generating learning data for the purpose of constructing a surrogate model, it is important from the perspective of the predictive accuracy of the surrogate model to perform the physical simulations based on various simulation patterns generated based on, for example, random walk theory.

[0004] Yangzi He, Shabnam J. Semnani, "Machine learning based modeling of path-dependent materials for finite element analysis", Computers and Geotechnics 156 (2023) 105254, [Retrieved September 7, 2023], Internet<URL:https: / / doi.org / 10.1016 / j.compgeo.2023.105254>

[0005] However, when generating simulation patterns based on the random walk theory, as the number of simulation patterns increases, a problem arises in that newly generated simulation patterns tend to be biased toward the average value of the simulation patterns generated up to that point.

[0006] The present disclosure reduces bias in simulation patterns when generating training data for building a surrogate model.

[0007] A first aspect of the present disclosure is a training data generation device that generates training data for constructing a surrogate model, comprising: a determination unit that randomly determines an initial value of input data and a final value of the input data; a path generation unit that generates a path by randomly assigning to each step a change amount that can be taken by each step in a path from the determined initial value to the final value; and a training data generation unit that generates training data including the determined initial value and final value and the change amount assigned to each step of the generated path.

[0008] A second aspect of the present disclosure is a learning data generation device according to the first aspect, further comprising an extraction unit that extracts a path from the generated paths in which a total change amount obtained by adding up the change amounts assigned to each step is equal to the difference between the determined initial value and final value, and the learning data generation unit generates learning data including the determined initial value and final value and the change amount assigned to each step of the extracted path.

[0009] A third aspect of the present disclosure is a learning data generation device according to the first aspect, further comprising a first setting unit that accepts a range of initial values ​​of the input data and a range of final values ​​of the input data, and the determination unit randomly determines the initial values ​​of the input data and the final values ​​of the input data within the ranges accepted by the first setting unit.

[0010] A fourth aspect of the present disclosure is a learning data generation device according to the first aspect, further comprising a second setting unit that accepts a number of steps, and the path generation unit randomly assigns a possible amount of change for each step to each of the number of steps accepted by the second setting unit.

[0011] A fifth aspect of the present disclosure is a learning data generation device according to the first aspect, further comprising a third setting unit that accepts possible changes in each step, and the path generation unit randomly assigns the changes accepted by the third setting unit to each step.

[0012] A sixth aspect of the present disclosure is a learning data generation device according to the second aspect, further comprising an acquisition unit that acquires a simulation result when the determined initial value and final value and the change amount assigned to each step of the extracted path are input as input data into a simulation device, and the learning data generation unit generates learning data including the determined initial value and final value, the change amount assigned to each step of the extracted path, and the simulation result.

[0013] In a training data generation method according to a seventh aspect of the present disclosure, a computer of a training data generation device that generates training data for constructing a surrogate model performs the following steps: randomly determining an initial value of input data and a final value of the input data; generating a path from the determined initial value to the final value by randomly assigning to each step a possible change amount for each step; and generating training data including the determined initial value and final value and the change amount assigned to each step of the generated path.

[0014] A training data generation program according to an eighth aspect of the present disclosure causes a computer of a training data generation device that generates training data for constructing a surrogate model to execute the following steps: a step of randomly determining an initial value of input data and a final value of the input data; a step of generating a path from the determined initial value to the final value by randomly assigning to each step the amount of change that each step can take; and a step of generating training data that includes the determined initial value and final value and the amount of change assigned to each step of the generated path.

[0015] According to the present disclosure, bias in simulation patterns can be reduced when generating learning data for constructing a surrogate model.

[0016] FIG. 1 is a diagram illustrating an example of the system configuration of a learning system. FIG. 2 is a diagram illustrating an example of the hardware configuration of a training data generation device. FIG. 3 is a diagram illustrating specific examples of time-series data of various simulation patterns. FIG. 4 is a diagram illustrating an example of the detailed functional configuration of a generation unit. FIG. 5 is an example of a flowchart illustrating the flow of training data generation processing. FIG. 6 is a first diagram illustrating a specific example of training data. FIG. 7 is a diagram illustrating an example of the distribution of input data of training data generated by the training data generation processing. FIG. 8 is a diagram illustrating an example of a path generated by the training data generation processing. FIG. 9 is a first diagram illustrating an example of the prediction accuracy of a trained model trained using training data generated by the training data generation processing. FIG. 10 is a second diagram illustrating a specific example of training data. FIG. 11 is a second diagram illustrating an example of the prediction accuracy of a trained model trained using training data generated by the training data generation processing.

[0017] Hereinafter, each embodiment will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant description will be omitted.

[0018] [First embodiment] <System configuration of a learning system> First, a description will be given of the system configuration of a learning system including a learning data generation device according to the first embodiment. Fig. 1 is a diagram showing an example of the system configuration of a learning system.

[0019] As shown in FIG. 1, the learning system 100 includes a simulation device 110, a learning data generation device 120, and a learning device .

[0020] The simulation device 110 is a device that reproduces the manufacturing device 10. Prerequisites (e.g., the structure, size, material, etc. of the device) for reproducing the manufacturing device 10 are set for the simulation device 110. Furthermore, by setting the prerequisites, various manufacturing conditions (e.g., temperature, pressure, flow rate, and other physical quantities controlled by the manufacturing device 10 during manufacturing) are input into the simulation device 110 that reproduces the manufacturing device 10, and a physical simulation is performed.

[0021] In the first embodiment, the manufacturing conditions input to the simulation device 110 are assumed to be time-series data, and include time-series data of various simulation patterns. Specifically, the time-series data of various simulation patterns includes: time-series data of various simulation patterns in which the combinations of the initial values ​​and final values ​​of the manufacturing conditions are different from each other; and time-series data of various simulation patterns in which the combinations of the initial values ​​and final values ​​of the manufacturing conditions are the same but the intermediate paths are different from each other.

[0022] A learning data generation program is installed in the learning data generation device 120, and the learning data generation device 120 functions as a generation unit 121 by executing the program.

[0023] The generator 121 generates learning data for each precondition set for the simulation device 110 .

[0024] Specifically, the generation unit 121 generates time-series data of various simulation patterns as manufacturing conditions under each precondition. The generation unit 121 then inputs the generated time-series data of the various simulation patterns into the simulation device 110 for which the corresponding preconditions are set, thereby executing a physical simulation and acquiring simulation results. The generation unit 121 then associates the time-series data of the various simulation patterns with the respective simulation results for each precondition, thereby generating learning data for each precondition. The generation unit 121 then stores the generated learning data in the learning data storage unit 122.

[0025] A learning program is installed in the learning device 130, and the learning device 130 functions as a learning unit 131 by executing the program.

[0026] The learning unit 131 reads out the learning data from the learning data storage unit 122 and uses the read out learning data to learn a model, thereby constructing a surrogate model, which is a trained model.

[0027] <Hardware Configuration of Training Data Generation Device> Next, the hardware configuration of the training data generation device 120 will be described. Fig. 2 is a diagram showing an example of the hardware configuration of a training data generation device. As shown in Fig. 2, each training data generation device 120 includes a processor 201, a memory 202, an auxiliary storage device 203, an I / F (Interface) device 204, a communication device 205, and a drive device 206. Note that the respective hardware components of the training data generation device 120 are connected to each other via a bus 207.

[0028] The processor 201 has various arithmetic devices such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc. The processor 201 executes various programs (for example, a learning data generation program) by reading them into the memory 202.

[0029] The memory 202 has a main storage device such as a read-only memory (ROM) or a random access memory (RAM). The processor 201 and the memory 202 form a so-called computer, and the processor 201 executes various programs read onto the memory 202, thereby enabling the computer to realize various functions.

[0030] The auxiliary storage device 203 stores various programs and various data used when the various programs are executed by the processor 201. For example, the learning data storage unit 122 is realized in the auxiliary storage device 203.

[0031] The I / F device 204 is a connection device for connecting an operation device 211, which is an example of a user interface device, and a display device 212. The communication device 205 is a communication device for communicating with external devices via a network (not shown).

[0032] The drive device 206 is a device for loading a recording medium 213. The recording medium 213 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, a magneto-optical disk, etc. The recording medium 213 may also include semiconductor memory that records information electrically, such as a ROM, a flash memory, etc.

[0033] The various programs to be installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 213 in the drive device 206 and reading the various programs recorded on the recording medium 213 by the drive device 206. Alternatively, the various programs to be installed in the auxiliary storage device 203 may be installed when downloaded from a network via the communication device 205.

[0034] 2 shows the hardware configuration of the training data generation device 120, one of the devices constituting the training system 100. However, the other devices constituting the training system 100, the simulation device 110 and the training device 130, also have the same hardware configuration as the training data generation device 120. Therefore, a description of the hardware configurations of the simulation device 110 and the training device 130 will be omitted.

[0035] <Specific Examples of Time Series Data> Next, a description will be given of specific examples of time series data of various simulation patterns generated as manufacturing conditions by the generation unit 121 of the learning data generation device 120. Fig. 3 is a diagram showing specific examples of time series data of various simulation patterns.

[0036] 3, the horizontal axis represents steps, and the vertical axis represents manufacturing conditions. A step is a time width obtained by dividing the time width from the start to the end of a simulation by the simulation device 110 by a predetermined number (referred to as the number of steps), and refers to the minimum time width when changing manufacturing conditions over time.

[0037] 3 shows how time-series data for three simulation patterns when the number of steps is 23 is generated as manufacturing conditions. Of these, the time-series data indicated by reference numerals 301 and 302 have the same combination of initial and final values ​​of the manufacturing conditions (initial value 1, final value 1), but are time-series data with different intermediate paths. Furthermore, the time-series data indicated by reference numeral 303 has a different combination of initial and final values ​​of the manufacturing conditions (initial value 2, final value 2) from the time-series data indicated by reference numeral 301, but are time-series data with the same intermediate path.

[0038] If the length of the double-headed arrow shown in FIG. 3 is taken to mean the amount of change in the manufacturing conditions per step = +1, the example in FIG. 3 indicates that the amount of change in the manufacturing conditions per step includes three types of change: ±0, +1, and +2. Also, all of the examples in FIG. 3 indicate that the total amount of change in the time-series data is 11. However, the amount of change in the manufacturing conditions per step is not limited to three types. Furthermore, the amount of change in the manufacturing conditions per step may include not only positive values ​​but also negative values. Furthermore, the total amount of change in the time-series data is not limited to 11. Furthermore, the number of steps is not limited to 23.

[0039] If the number of initial values ​​of the manufacturing conditions is M, the number of steps is S, and the number of types of change in the manufacturing conditions per step is n, then the number of patterns of time-series data that can be generated as manufacturing conditions = M x n Ｓ The learning data generation device 120 generates learning data by randomly generating time-series data and extracting time-series data that reach a predetermined final value for each initial value (details will be described later).

[0040] <Functional Configuration of Generator> Next, a detailed description will be given of the functional configuration of the generator 121. Fig. 4 is a diagram showing an example of the detailed functional configuration of the generator.

[0041] 4 , the generation unit 121 includes a step number setting unit 411, a change amount setting unit 412, an initial value range setting unit 413, a final value range setting unit 414, a determination unit 420, a path generation unit 430, and a difference calculation unit 440. The generation unit 121 also includes an extraction unit 450, a simulation result acquisition unit 460, and a learning data generation unit 470.

[0042] The step number setting unit 411 is an example of a second setting unit, and receives the number of steps input by the user and notifies the path generation unit 430 of the number.

[0043] The variation setting unit 412 is an example of a third setting unit, and receives the variation amount that can be taken by the manufacturing conditions per step input by the user, and notifies the path generating unit 430 of the variation amount.

[0044] The initial value range setting unit 413 receives the maximum and minimum possible values ​​for the initial value input by the user and notifies the determination unit 420. The final value range setting unit 414 receives the maximum and minimum possible values ​​for the final value input by the user and notifies the determination unit 420. The initial value range setting unit 413 and the final value range setting unit 414 are examples of a first setting unit.

[0045] The determination unit 420 determines a random value as the initial value within the range between the maximum and minimum initial values ​​notified by the initial value range setting unit 413 and notifies the difference calculation unit 440 of the value.

[0046] Furthermore, the determination section 420 determines a random value as the final value within the range between the maximum and minimum final values ​​notified by the final value range setting section 414 , and notifies the difference calculation section 440 of the final value.

[0047] The path generation unit 430 generates various paths by randomly assigning the amount of change notified by the amount of change setting unit 412 to each step of the number of steps notified by the step number setting unit 411, and notifies the extraction unit 450. Furthermore, the path generation unit 430 calculates the total amount of change for each of the various paths by adding up the amount of change assigned to each step for each step, and notifies the extraction unit 450.

[0048] The difference calculation unit 440 calculates the difference between the initial value and the final value notified by the determination unit 420 and notifies the extraction unit 450 of the difference.

[0049] The extraction unit 450 extracts, from among the various paths notified by the path generation unit 430, a path whose total amount of change is equal to the difference value notified by the difference calculation unit 440. The extraction unit 450 also notifies the simulation device 110 and the learning data generation unit 470 of time-series data including the extracted path and the corresponding initial value and final value.

[0050] As a result, in the simulation device 110, a physical simulation is executed for the time-series data of various simulation patterns generated by the generation unit 121.

[0051] The simulation result acquisition unit 460 acquires the simulation results notified from the simulation device 110 in response to the extraction unit 450 notifying the simulation device 110 of time-series data of various simulation patterns. The simulation result acquisition unit 460 also notifies the learning data generation unit 470 of the acquired simulation results.

[0052] The learning data generation unit 470 generates learning data by associating the simulation results notified by the simulation result acquisition unit 460 with the time-series data of various simulation patterns notified by the extraction unit 450. The learning data generation unit 470 stores the generated learning data in the learning data storage unit 122.

[0053] <Learning Data Generation Process Flow> Next, a description will be given of the flow of learning data generation process by the generation unit 121 of the learning data generation device 120. Fig. 5 is an example of a flowchart showing the flow of the learning data generation process.

[0054] In step S501, the learning data generation device 120 receives the maximum and minimum possible values ​​of the initial value and the maximum and minimum possible values ​​of the final value input by the user.

[0055] In step S502, the learning data generating device 120 receives the number of steps and the amount of change that the manufacturing conditions can have per step, input by the user.

[0056] In step S503, the training data generation device 120 randomly determines an initial value within a range between a maximum and minimum initial value. The training data generation device 120 also randomly determines a final value within a range between a maximum and minimum final value. Furthermore, the training data generation device 120 calculates the difference between the randomly determined initial value and the final value.

[0057] In step S504, the learning data generation device 120 generates various paths by randomly allocating the received amount of change to each step of the received number of steps.

[0058] In step S505, the learning data generation device 120 extracts time-series data of a route in which the total amount of change is equal to the difference value from among the various routes that have been generated.

[0059] In step S506, the learning data generation device 120 determines whether a predetermined number of time-series data for various simulation patterns have been generated for each precondition. If it is determined in step S506 that the predetermined number of time-series data for various simulation patterns have not been generated (NO in step S506), the process returns to step S503. In this case, steps S503 to S505 are executed for other randomly determined initial values ​​and final values ​​to generate new time-series data for various simulation patterns.

[0060] On the other hand, if it is determined in step S506 that a predetermined number of time-series data of various simulation patterns have been generated (YES in step S506), the process proceeds to step S507.

[0061] In step S507, the training data generation device 120 notifies the simulation device 110 of time-series data of a predetermined number of various simulation patterns. In response to the notification of the time-series data of the predetermined number of various simulation patterns, the training data generation device 120 acquires simulation results from the simulation device 110. The training data generation device 120 generates training data by associating the acquired simulation results with the time-series data of the various simulation patterns, and stores the training data in the training data storage unit 122.

[0062] In this way, by randomly determining the initial value and the final value and then randomly generating a path, the learning data generation device 120 can generate learning data that includes time series data of a simulation pattern with little bias.

[0063] <Specific Example of Training Data> Next, a description will be given of a specific example of training data generated by the training data generation device 120 by executing the training data generation process shown in Fig. 5. Fig. 6 is a first diagram showing a specific example of training data.

[0064] For the sake of simplicity, the following conditions were used to generate the learning data 600: number of steps = 5; possible changes in manufacturing conditions per step: -1, -0.5, ±0, +0.5, +1; maximum and minimum initial values: ±0; maximum and minimum final values: 5, -5.

[0065] Also, for the sake of simplicity, the final value is used as the correct answer data of the learning data 600 instead of obtaining the simulation result.

[0066] 6, the training data 600 includes information items such as "input data" and "correct answer data." The "input data" further includes information items such as "manufacturing condition ID," "initial value," and "step 1" to "step 5."

[0067] The "manufacturing condition ID" stores an identifier for identifying each of the time-series data of the various generated simulation patterns.

[0068] The "initial value" stores an initial value that is randomly determined within the range of the maximum and minimum possible values ​​for the initial value. In this case, ±0 is set as the maximum and minimum initial values, so "0" is stored in the "initial value."

[0069] "Step 1" to "Step 5" store randomly assigned amounts of change from among the amounts of change that the manufacturing conditions can take per step.

[0070] On the other hand, the "correct answer data" further includes a "final value" as an information item. The "final value" stores the total amount of change obtained by adding the amount of change assigned to each step for each step to the initial value.

[0071] <Distribution of Training Data> Next, a description will be given of the distribution of time-series data of various simulation patterns that is generated by the training data generation device 120 by executing the training data generation process shown in Fig. 5 and stored in the "input data" of the training data 600. Fig. 7 is a diagram showing an example of the distribution of the input data of the training data generated by the training data generation process.

[0072] In Fig. 7(a), the horizontal axis represents the amount of change that can occur in the manufacturing conditions per step, and the vertical axis represents the frequency of allocation of the amount of change allocated to each step. For comparison, the example in Fig. 7(a) shows overlapping data for various simulation patterns generated based on the random walk theory, and data for various simulation patterns generated by the learning data generation process shown in Fig. 5.

[0073] As shown in Fig. 7(a), in the case of time-series data of various simulation patterns generated based on the random walk theory, the frequency of allocation of the change amount assigned to each step is uniform. On the other hand, in the case of time-series data of various simulation patterns generated by the learning data generation process shown in Fig. 5, the frequency of allocation of the change amount assigned to each step is not uniform, and -1 or 1 is assigned more frequently than ±0.

[0074] 7(b), the horizontal axis represents the total amount of change, and the vertical axis represents the frequency of the total amount of change. For comparison, the example in FIG. 7(b) also shows overlapping time-series data of various simulation patterns generated based on the random walk theory, and time-series data of various simulation patterns generated by the learning data generation process shown in FIG.

[0075] As shown in Figure 7(b), in the case of time-series data of various simulation patterns generated based on the random walk theory, the total change amount was most frequently ±0, and the total change amount was least frequently -5 or +5. In other words, as the number of simulation patterns increased, newly generated simulation patterns became biased toward the average value of the simulation patterns generated up to that point.

[0076] On the other hand, in the case of the time-series data of various simulation patterns generated by the learning data generation process shown in FIG. 5, the frequency of the total amount of change is uniform.

[0077] In this way, it can be said that the time series data of various simulation patterns generated by the learning data generation process shown in Figure 5 is less biased, that is, it is time series data that is evenly dispersed.

[0078] <Specific example of a path> Next, a specific example of a path generated by the learning data generation device 120 by executing the learning data generation process shown in Fig. 5 will be described. Fig. 8 is a diagram showing an example of a path generated by the learning data generation process. In Fig. 8, the horizontal axis represents steps, and the vertical axis represents the total amount of change up to each step.

[0079] 8(a) shows time-series data of various simulation patterns generated based on the random walk theory, while FIG. 8(b) shows time-series data of various simulation patterns generated by the learning data generation process shown in FIG.

[0080] 8(a) and 8(b), in the case of time-series data of various simulation patterns generated based on the random walk theory, paths in which the total amount of change passes near ±0 are generated with a high frequency. On the other hand, in the case of time-series data of various simulation patterns generated by the learning data generation process shown in FIG. 5, the total amount of change is not biased toward paths in which the total amount of change passes near ±0, but is generated with an even distribution.

[0081] <Regarding Prediction Accuracy of Trained Model> Next, a description will be given of the prediction accuracy of a trained model when a predetermined model is trained using training data generated by the training data generation device 120 as a result of executing the training data generation process, and a surrogate model, which is a trained model, is constructed. Note that, in the following, a four-layer perceptron, which is a neural network constructed using the neural network library keras, is used as the predetermined model.

[0082] FIG. 9 is a first diagram illustrating an example of the prediction accuracy of a trained model trained using the training data generated by the training data generation process.

[0083] 9 (a-1) to (a-3) show graphs obtained by: generating 500 pieces of time-series data with manufacturing condition IDs of 1 to 500 from training data 600; using the generated 500 pieces of time-series data to train a predetermined model and construct a trained model; and calculating, for the constructed trained model, predicted data when time-series data with correct answer = -5.0 is input as input data, predicted data when time-series data with correct answer = 0 is input as input data, and predicted data when time-series data with correct answer = 5.0 is input as input data. Note that the left side of each graph shows, as a comparative example, predicted data predicted by a trained model when 500 pieces of time-series data are generated based on random walk theory.

[0084] 9 (b-1) to (b-3) show graphs obtained by: generating 1,000 pieces of time-series data for manufacturing condition IDs 1 to 1,000 in the training data 600; using the generated 1,000 pieces of time-series data to train a predetermined model and construct a trained model; and calculating, for the constructed trained model, predicted data when time-series data for which correct answer data = -5.0 is input as input data, predicted data when time-series data for which correct answer data = 0 is input as input data, and predicted data when time-series data for which correct answer data = 5.0 is input as input data. Note that the left side of each graph shows, as a comparative example, predicted data predicted by the trained model when 1,000 pieces of time-series data are generated based on the random walk theory.

[0085] 9 (c-1) to (c-3) show graphs obtained by: generating 5,000 pieces of time-series data for manufacturing condition IDs = 1 to 5,000 in the training data 600; using the generated 5,000 pieces of time-series data to train a predetermined model and construct a trained model; and calculating, for the constructed trained model, predicted data when time-series data for which correct answer data = -5.0 is input as input data, predicted data when time-series data for which correct answer data = 0 is input as input data, and predicted data when time-series data for which correct answer data = 5.0 is input as input data. Note that the left side of each graph shows, as a comparative example, predicted data predicted by the trained model when 5,000 pieces of time-series data are generated based on the random walk theory.

[0086] 9 (d-1) to (d-3) show graphs obtained by: generating 10,000 pieces of time-series data for manufacturing condition IDs = 1 to 10,000 in the training data 600; using the generated 10,000 pieces of time-series data to train a predetermined model and construct a trained model; and calculating, for the constructed trained model, predicted data when time-series data for which correct answer data = -5.0 is input as input data, predicted data when time-series data for which correct answer data = 0 is input as input data, and predicted data when time-series data for which correct answer data = 5.0 is input as input data. Note that the left side of each graph shows, as a comparative example, predicted data predicted by the trained model when 10,000 pieces of time-series data are generated based on the random walk theory.

[0087] In this way, when comparing a trained model constructed using training data generated by the training data generation process shown in Figure 5 with a trained model constructed using training data generated based on random walk theory, the following can be found: - the trained model constructed using training data generated by the training data generation process shown in Figure 5 can achieve higher prediction accuracy when predicting values ​​far from the average, regardless of the number of training data; - as the number of training data increases, the prediction accuracy when predicting values ​​close to the average becomes similar.

[0088] In other words, it was shown that reducing the bias in the time-series data of the simulation pattern can improve the prediction accuracy.

[0089] <Summary> As is clear from the above explanation, the training data generation device 120 according to the first embodiment is a device that generates training data for constructing a surrogate model, and randomly determines the initial value of the input data and the final value of the input data. Generates a path by randomly assigning to each step the amount of change that can occur at each step in the path from the determined initial value to the final value. Generates training data that includes the determined initial value and final value, and the amount of change assigned to each step of the generated path.

[0090] In this way, the training data generation device 120 according to the first embodiment is configured to randomly determine the initial value and the final value and then randomly generate a path. As a result, the training data generation device 120 according to the first embodiment can generate training data that includes time-series data of a simulation pattern that is less biased than training data generated based on the random walk theory.

[0091] In other words, according to the first embodiment, it is possible to reduce bias in simulation patterns when generating learning data for constructing a surrogate model.

[0092] [Second Embodiment] In the above-described first embodiment, a case has been described in which the final value is used as the correct answer data of the training data when verifying the prediction accuracy of the trained model. That is, the prediction accuracy of the trained model is verified on the premise that the training data is a case in which the correct answer data has a linear relationship with the input data. In contrast, in the second embodiment, the prediction accuracy of the trained model is verified on the premise that the training data is a case in which the correct answer data has a nonlinear relationship with the input data.

[0093] 10 is a second diagram showing a specific example of training data. The difference from the training data 600 shown in FIG. 6 is that instead of storing the final value as the correct answer data, the cubed value of the final value is stored.

[0094] FIG. 11 is a second diagram illustrating an example of the prediction accuracy of a trained model trained using the training data generated by the training data generation process.

[0095] 11 (a-1) to (a-3) show graphs obtained by: generating 5,000 pieces of time-series data with manufacturing condition IDs of 1 to 5,000 from training data 1,000; using the generated 5,000 pieces of time-series data to train a predetermined model and construct a trained model; and calculating, for the constructed trained model, predicted data when time-series data with correct answer data = -125 is input as input data, predicted data when time-series data with correct answer data = 0 is input as input data, and predicted data when time-series data with correct answer data = 125 is input as input data. Note that the left side of each graph shows, as a comparative example, predicted data predicted by the trained model when 5,000 pieces of time-series data are generated based on the random walk theory.

[0096] As described above, according to the second embodiment, even when the training data is such that the correct answer data has a nonlinear relationship with the input data, a trained model with high prediction accuracy can be constructed.

[0097] [Other embodiments] In the first embodiment, the manufacturing conditions of the manufacturing apparatus 10 are used as input data for the learning data, but the input data for the learning data is not limited to the manufacturing conditions of the manufacturing apparatus 10 and may be other physical quantities input to the simulation apparatus 110.

[0098] In the first embodiment, the number of steps, the amount of change that the manufacturing conditions can take per step, the maximum and minimum values ​​that the initial value can take, and the maximum and minimum values ​​that the final value can take are input by the user. However, any of these values ​​may be preset in the training data generation device 120.

[0099] In the first embodiment, the range of initial values ​​is input by inputting the maximum and minimum values ​​that the initial values ​​can take, but the method for inputting the range of initial values ​​is not limited to this.Similarly, in the first embodiment, the range of final values ​​is input by inputting the maximum and minimum values ​​that the final values ​​can take, but the method for inputting the range of final values ​​is not limited to this.

[0100] Although the first embodiment does not mention details of the method for randomly determining the initial value and the final value, for example, the initial value and the final value may be randomly determined using uniform random numbers. Alternatively, the initial value and the final value may be randomly determined using random numbers according to a predetermined generation distribution. In either case, the initial value and the final value are determined so as to vary.

[0101] The present invention is not limited to the configurations described in the above embodiments, but may be combined with other elements, etc. These aspects can be changed without departing from the spirit of the present invention, and can be appropriately determined depending on the application form.

[0102] This application claims priority based on Japanese Patent Application No. 2023-176970, filed on October 12, 2023, the entire contents of which are incorporated herein by reference.

[0103] 10: Manufacturing device 100: Learning system 110: Simulation device 120: Learning data generation device 121: Generation unit 130: Learning device 131: Learning unit 411: Number of steps setting unit 412: Change amount setting unit 413: Initial value range setting unit 414: Final value range setting unit 420: Determination unit 430: Path generation unit 440: Difference calculation unit 450: Extraction unit 460: Simulation result acquisition unit 470: Learning data generation unit 600: Learning data 1000: Learning data

Claims

1. A training data generation device that generates training data for constructing a surrogate model, comprising: a determination unit that randomly determines an initial value of input data and a final value of the input data; a path generating unit that generates a path by randomly allocating to each step a change amount that can be taken by each step in the path from the determined initial value to the final value; a learning data generating unit that generates learning data including the determined initial value and final value and a change amount assigned to each step of the generated path; A learning data generation device having the above configuration.

2. further comprising an extraction unit that extracts, from the generated paths, a path in which a total change amount obtained by adding up the change amounts assigned to each step is equal to a difference value between the determined initial value and final value; The learning data generation unit The learning data generating device according to claim 1 , wherein the learning data generating device generates learning data including the determined initial value and final value and a variation amount assigned to each step of the extracted path.

3. a first setting unit configured to receive a range of an initial value of the input data and a range of a final value of the input data, The device for generating training data according to claim 1 , wherein the determination unit randomly determines the initial value of the input data and the final value of the input data within a range accepted by the first setting unit.

4. further comprising a second setting unit that accepts the number of steps; the path generation unit randomly assigns a possible change amount for each step to each step of the number of steps accepted by the second setting unit; The learning data generating device according to claim 1 .

5. a third setting unit that receives a possible change amount for each step; The learning data generating device according to claim 1 , wherein the path generating unit randomly assigns the amount of change received by the third setting unit to each step.

6. an acquisition unit that acquires a simulation result when the determined initial value and final value and a change amount assigned to each step of the extracted path are input as input data to a simulation device; The learning data generation unit 3. The learning data generating device according to claim 2, wherein the device generates learning data including the determined initial and final values, a variation assigned to each step of the extracted path, and the simulation result.

7. a computer of a training data generation device that generates training data for constructing a surrogate model, a step of randomly determining an initial value of input data and a final value of the input data; generating a path by randomly assigning to each step a change amount that can be taken by each step in the path from the determined initial value to the final value; generating learning data including the determined initial and final values ​​and a change amount assigned to each step of the generated path; A method for generating training data.

8. A computer of a training data generation device that generates training data for constructing a surrogate model, a step of randomly determining an initial value of input data and a final value of the input data; generating a path by randomly assigning to each step a change amount that can be taken by each step in the path from the determined initial value to the final value; generating learning data including the determined initial and final values ​​and a change amount assigned to each step of the generated path; A learning data generation program for executing the above.