Material design methods

Abstract

To improve prediction accuracy in a predication device using a trained model.SOLUTION: A prediction device has: a first trained model and a second trained model that respectively output first output data and second output data by receiving input data to be predicted; and an output unit that acquires the first and second output data, and outputs prediction data by calculating a weighted average value or by taking a weighted majority. The first trained model is configured to have higher prediction accuracy for input data in an interpolation region than that of the second trained model, and the second trained model is configured to have higher prediction accuracy for input data in an extrapolation region than that of the first trained model.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] This disclosure relates to Material design methods it.

Background Art

[0002] Conventionally, the design of materials has been carried out by repeating trial production based on the experience of material developers. On the other hand, attempts have been made to apply learning models in the design of materials. For example, by collecting the design conditions during trial production and the evaluation results of the trial-produced materials (such as the characteristic values of the materials) and training the model using them as a training dataset, it becomes possible to predict in advance the characteristic values of the materials to be trial-produced under new design conditions.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in general, while a certain degree of prediction accuracy can be obtained for the input data in the interpolation region for a trained model, it is difficult to obtain sufficient prediction accuracy for the input data in the extrapolation region.

[0005] This disclosure aims to improve the prediction accuracy in a prediction device using a trained model.

Means for Solving the Problems

[0006] The prediction device according to the first aspect of this disclosure is When the input data to be predicted is input, a first trained model and a second trained model output a first output data and a second output data, respectively. The system includes an output unit that obtains the first and second output data and outputs prediction data by calculating a weighted average or by taking a weighted majority vote, The first trained model is configured such that its prediction accuracy for the input data in the interpolation region is higher than that of the second trained model, and the second trained model is configured such that its prediction accuracy for the input data in the extrapolation region is higher than that of the first trained model.

[0007] A second aspect of this disclosure is a prediction device as described in the first aspect, The output unit calculates the weighted average value or takes the weighted majority vote under predetermined weights.

[0008] A third aspect of this disclosure is a prediction device as described in the second aspect, The predetermined weights mentioned above are: The prediction data, output from the output unit under multiple weights when input data for the validation dataset is provided, is determined based on the error between the prediction data and the ground truth data corresponding to the input data for the validation dataset.

[0009] A fourth aspect of this disclosure is a prediction device as described in the second aspect, The first trained model, the second trained model, and the predetermined weights are: For each of the multiple types of first pre-trained models and multiple types of second pre-trained models, the prediction data output from the output unit under multiple types of weights when input data from the validation dataset is input is determined based on the error between the prediction data and the ground truth data corresponding to the input data of the validation dataset.

[0010] A fifth aspect of this disclosure is a prediction device as described in the fourth aspect, The aforementioned multiple types of first trained models are trained with different hyperparameters and / or under different training methods. The aforementioned multiple types of second trained models are each set to different hyperparameters and / or trained using different learning methods.

[0011] A sixth aspect of this disclosure is a prediction device as described in the first aspect, The system further includes a determination unit that determines whether the input data to be predicted is input data in an interpolation region or input data in an extrapolation region. The output unit calculates the weighted average value or takes the weighted majority vote, based on the weights corresponding to the discrimination result by the discrimination unit.

[0012] A seventh aspect of this disclosure is a prediction device as described in the first aspect, The system further includes a discrimination unit that evaluates the strength of extrapolability of the input data to be predicted, The output unit calculates the weighted average value or takes the weighted majority vote, based on the evaluation results from the discrimination unit.

[0013] An eighth aspect of this disclosure is a prediction device as described in the seventh aspect, The aforementioned discrimination unit is The strength of extrapolability of the input data to be predicted is evaluated by using one or more of the following methods: an evaluation method based on the uncertainty of random forest prediction, an evaluation method based on the uncertainty of Bayesian estimation, an evaluation method based on kernel density estimation, or an evaluation method based on distance.

[0014] A ninth aspect of this disclosure is a prediction device as described in the sixth aspect, The weights corresponding to the aforementioned discrimination result include weights for the interpolation region and weights for the extrapolation region. The weights for the interpolation region are: When the input data in the interpolation region of the verification data set is input, it is determined based on the error between the prediction data respectively output under a plurality of types of weights from the output unit and the correct answer data corresponding to the input data in the interpolation region of the verification data set. The weights for the extrapolation region are determined based on the error between the prediction data respectively output under a plurality of types of weights from the output unit and the correct answer data corresponding to the input data in the extrapolation region of the verification data set when the input data in the extrapolation region of the verification data set is input.

[0015] The tenth aspect of the present disclosure is the prediction device according to the first aspect, where the first learned model has been learned under any one or more learning methods of decision tree, random forest, gradient boosting, bagging, adaboost, k-nearest neighbor method, and neural network, and the second learned model has been learned under any one or more learning methods of Gaussian process, kernel ridge, support vector machine, linear, partial least squares, lasso, linear ridge, elastic net, Bayesian ridge, and neural network.

[0016] The learning device according to the eleventh aspect of the present disclosure is a first learned model and a second learned model that output first output data and second output data respectively when input data of a verification data set is input, and an output unit that acquires the first and second output data and outputs respective prediction data by calculating a weighted average value or taking a weighted majority vote under a plurality of types of weights, and a determination unit that determines any one of the plurality of types of weights based on the error between the respective output prediction data and the correct answer data corresponding to the input data of the verification data set.

[0017] The twelfth aspect of the present disclosure is the learning device according to the eleventh aspect, The aforementioned determination unit, Based on the error between the predicted data output from the output unit by inputting the input data of the interpolation region of the validation dataset and the ground truth data corresponding to the input data of the interpolation region of the validation dataset, the weights for the interpolation region are determined. The weights for the extrapolation region are determined based on the error between the predicted data output from the output unit, which is obtained by inputting the input data for the extrapolation region of the validation dataset, and the correct data corresponding to the input data for the extrapolation region of the validation dataset.

[0018] A thirteenth aspect of this disclosure is a learning device as described in the eleventh aspect, The output unit is, By inputting the validation dataset's input data to multiple types of first pre-trained models and multiple types of second pre-trained models, the first and second output data output from the multiple types of first and second pre-trained models are obtained, and the respective predicted data is output by calculating the weighted average under the multiple types of weights, or by taking a weighted majority vote. The aforementioned determination unit, Based on the error between each of the outputted prediction data and the ground truth data corresponding to the input data of the validation dataset, A first trained model is selected from among the multiple types of first trained models. A second trained model is selected from among the multiple types of second trained models mentioned above. Determine one of the aforementioned multiple types of weights.

[0019] A fourteenth aspect of this disclosure is a learning device as described in the thirteenth aspect, The aforementioned multiple types of first trained models are trained with different hyperparameters and / or under different training methods. The aforementioned multiple types of second trained models are each set to different hyperparameters and / or trained using different learning methods.

[0020] A 15th aspect of this disclosure is a learning device as described in the 11th aspect, The first trained model is configured such that its prediction accuracy for the input data in the interpolation region is higher than that of the second trained model, and the second trained model is configured such that its prediction accuracy for the input data in the extrapolation region is higher than that of the first trained model.

[0021] A sixteenth aspect of this disclosure is a learning device as described in the fifteenth aspect, The first pre-trained model described above is trained using one or more of the following learning methods: decision trees, random forests, gradient boosting, bagging, AdaBoost, k-nearest neighbors, and neural networks. The second pre-trained model is trained using one or more of the following learning methods: Gaussian process, kernel ridge, support vector machine, linear, partial least squares, Lasso, linear ridge, elastic network, Bayesian ridge, or neural network.

[0022] The prediction method relating to the 17th aspect of this disclosure is: The process involves inputting the data to be predicted, causing the first trained model and the second trained model to output the first output data and the second output data, respectively. The process includes a step of obtaining the first and second output data and outputting prediction data by calculating a weighted average or by taking a weighted majority vote, The first trained model is configured such that its prediction accuracy for the input data in the interpolation region is higher than that of the second trained model, and the second trained model is configured such that its prediction accuracy for the input data in the extrapolation region is higher than that of the first trained model.

[0023] The learning method relating to the 18th aspect of this disclosure is: The process involves inputting data from a validation dataset, causing the first and second trained models to output first and second output data, respectively. The process involves obtaining the first and second output data, calculating a weighted average under multiple types of weights, or taking a weighted majority vote to output the respective prediction data, The process includes a step of determining one of the multiple types of weights based on the error between each of the outputted prediction data and the correct data corresponding to the input data of the verification dataset.

[0024] The prediction program relating to the 19th aspect of this disclosure is The process involves inputting the data to be predicted, causing the first trained model and the second trained model to output the first output data and the second output data, respectively. A prediction program that causes a computer to perform the steps of obtaining the first and second output data and outputting prediction data by calculating a weighted average or by taking a weighted majority vote. The first trained model is configured such that its prediction accuracy for the input data in the interpolation region is higher than that of the second trained model, and the second trained model is configured such that its prediction accuracy for the input data in the extrapolation region is higher than that of the first trained model.

[0025] The learning program relating to the 20th aspect of this disclosure is: The process involves inputting data from a validation dataset, causing the first and second trained models to output first and second output data, respectively. The process involves obtaining the first and second output data, calculating a weighted average under multiple types of weights, or taking a weighted majority vote to output the respective prediction data, The computer is instructed to perform the following steps: determine one of the multiple types of weights based on the error between each of the outputted prediction data and the correct data corresponding to the input data of the verification dataset. [Effects of the Invention]

[0026] According to this disclosure, the prediction accuracy can be improved in a prediction device that uses a pre-trained model. [Brief explanation of the drawing]

[0027] [Figure 1] Figure 1 is the first diagram showing an example of the functional configuration of the learning device in the learning phase and the prediction device in the prediction phase. [Figure 2] Figure 2 shows an example of the hardware configuration of the learning device and prediction device. [Figure 3] Figure 3 is the first flowchart showing the flow of the learning and prediction processes. [Figure 4] Figure 4 is a second diagram showing an example of the functional configuration of the learning device in the learning phase and the prediction device in the prediction phase. [Figure 5] Figure 5 is a second flowchart showing the flow of the learning and prediction processes. [Figure 6] Figure 6 is the first diagram showing an example of the functional configuration of the learning device during the optimization phase. [Figure 7] Figure 7 is a second diagram showing an example of the functional configuration of the learning device during the optimization phase. [Figure 8] Figure 8 is a third flowchart illustrating the flow of the learning and prediction processes. [Figure 9] Figure 9 is the first flowchart showing the flow of the optimization process. [Figure 10] Figure 10 is a third diagram showing an example of the functional configuration of the learning device during the learning phase. [Figure 11] Figure 11 is a third diagram showing an example of the functional configuration of the learning device during the optimization phase. [Figure 12] Figure 12 is a fourth flowchart illustrating the flow of the learning and prediction processes. [Figure 13] Figure 13 is a second flowchart illustrating the optimization process flow. [Figure 14] Figure 14 is the fourth figure, showing an example of the functional configuration of the learning device during the learning phase. [Figure 15] Figure 15 is the fourth figure, showing an example of the functional configuration of the learning device during the optimization phase. [Figure 16] Figure 16 is a fifth flowchart illustrating the flow of the learning and prediction processes. [Figure 17] Figure 17 is a third flowchart illustrating the optimization process flow. [Figure 18] Figure 18 shows an example of prediction accuracy. [Modes for carrying out the invention]

[0028] Each embodiment will be described below with reference to the attached drawings. In this specification and the drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant descriptions will be omitted.

[0029] [First Embodiment] <Functional Configuration of Learning and Prediction Devices> First, the functional configurations of the learning device and prediction device according to the first embodiment will be described. The learning device according to the first embodiment will be described as an example of a learning device that performs learning using a learning dataset that includes the design conditions at the time of prototyping and the characteristic values ​​of the prototyped material. The prediction device according to the first embodiment will be described as an example of a prediction device that predicts the characteristic values ​​of a prototyped material under new design conditions.

[0030] However, the learning device and prediction device according to the first embodiment are not limited to the above-mentioned uses and may be used for purposes other than material design.

[0031] Figure 1 is the first diagram showing an example of the functional configuration of the learning device in the learning phase and the prediction device in the prediction phase. The learning device 120 has a learning program installed, and when this program is executed, the learning device 120 performs the following actions: • Interpolation prediction model 121_1, • Comparison / Modification Section 122_1, • Extrapolation prediction model 121_2, • Comparison / Modification Section 122_2, It functions as follows (see Figure 1, 1a).

[0032] The learning device 120 uses the training dataset 111 stored in the material data storage unit 110 to train the interpolation prediction model 121_1 and the extrapolation prediction model 121_2, and generates the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2.

[0033] As shown in Figure 1, 1a, the training dataset 111 includes "input data" and "ground truth data" as information items. In the example in Figure 1, 1a, "design condition 1" to "design condition n" are stored as "input data," and "characteristic value 1" to "characteristic value n" are stored as "ground truth data."

[0034] The interpolation prediction model 121_1 is a pre-trained model configured such that it generates a trained interpolation prediction model 131_1 with a higher prediction accuracy for the input data in the interpolation region than the trained extrapolation prediction model 131_2.

[0035] The interpolation prediction model 121_1 outputs output data when it receives "design condition 1" to "design condition n" stored in the "input data" of the training dataset 111.

[0036] The comparison / modification unit 122_1 updates the model parameters of the interpolation prediction model 121_1 according to the error between the output data output from the interpolation prediction model 121_1 and the "characteristic values ​​1" to "characteristic values ​​n" stored in the "ground truth data" of the training dataset 111.

[0037] As a result, the learning device 120 generates a trained interpolation prediction model 131_1 (the first trained model). The learning device 120 then applies the generated trained interpolation prediction model 131_1 (the second trained model) to the prediction device 130.

[0038] Furthermore, the interpolation prediction model 121_1 that the learning device 120 learns uses the following learning method: "Decision trees, random forests, gradient boosting, bagging, Adaboost, k-nearest neighbors, neural networks" The model is trained using one or more of the following learning methods. In other words, the learning device 120 uses a model as the interpolation prediction model 121_1, which is trained using a learning method suitable for the input data of the interpolation region.

[0039] Furthermore, when the learning device 120 trains the interpolation prediction model 121_1, the hyperparameters of the interpolation prediction model 121_1 are set to values ​​suitable for the input data of the interpolation region (hyperparameters for the interpolation prediction model).

[0040] On the other hand, extrapolation prediction model 121_2 is a pre-trained model configured to generate a trained extrapolation prediction model 131_2 with higher prediction accuracy for the input data in the extrapolation domain than the trained extrapolation prediction model 131_1.

[0041] The extrapolation prediction model 121_2 outputs output data when it receives "design condition 1" to "design condition n" stored in the "input data" of the training dataset 111.

[0042] The comparison / modification unit 122_2 updates the model parameters of the extrapolation prediction model 121_2 according to the error between the output data output from the extrapolation prediction model 121_2 and the "characteristic values ​​1" to "characteristic values ​​n" stored in the "ground truth data" of the training dataset 111.

[0043] As a result, the learning device 120 generates a trained extrapolated prediction model 131_2. The learning device 120 then applies the generated trained extrapolated prediction model 131_2 to the prediction device 130.

[0044] Furthermore, the extrapolation prediction model 121_2 that the learning device 120 learns uses the following learning method: "Gaussian processes, kernel ridges, support vector machines, linear, partial least squares, Lasso, linear ridges, elastic networks, Bayesian ridges, neural networks" The model is assumed to be trained under one or more of the following learning methods. In other words, the learning device 120 uses a model as the extrapolation prediction model 121_2, which is trained under a learning method suitable for the input data of the extrapolation domain.

[0045] Furthermore, when the learning device 120 trains the extrapolation prediction model 121_2, the hyperparameters of the extrapolation prediction model 121_2 are set to values ​​suitable for the input data of the extrapolation domain (hyperparameters for the extrapolation prediction model).

[0046] On the other hand, the prediction device 130 has a prediction program installed, and when this program is executed, the prediction device 130 will • Pre-trained interpolation prediction model 131_1 • Pre-trained extrapolation prediction model 131_2, Output section 132, It functions as follows (see Figure 1, 1b).

[0047] The trained interpolation prediction model 131_1 is generated when the learning device 120 trains the interpolation prediction model 121_1 using the training dataset 111. The trained interpolation prediction model 131_1 receives the input data to be predicted (design condition x), predicts a first characteristic value (first output data), and outputs it to the output unit 132.

[0048] The trained extrapolation prediction model 131_2 is generated when the learning device 120 trains the extrapolation prediction model 121_2 using the training dataset 111. The trained extrapolation prediction model 131_2 predicts a second characteristic value (second output data) when the design condition x is input and outputs it to the output unit 132.

[0049] The output unit 132 determines the characteristic value y for the design condition x based on the first characteristic value predicted by the trained interpolation prediction model 131_1 and the second characteristic value predicted by the trained extrapolation prediction model 131_2, and outputs it as prediction data.

[0050] The output unit 132 determines the characteristic value y by calculating the weighted average of the first characteristic value and the second characteristic value. Alternatively, the output unit 132 determines the characteristic value y by taking a weighted majority vote between the first characteristic value and the second characteristic value.

[0051] In this embodiment, the weights used by the output unit 132 when calculating the weighted average value, or the weights used when taking a weighted majority vote, are assumed to be optimized in advance.

[0052] For example, if the design condition x is a continuous value, the output unit 132 calculates the average value of the first characteristic value and the second characteristic value (i.e., it uses a weight of 0.5 for both the first and second characteristic values). Alternatively, the output unit 132 calculates a weighted average value of the first and second characteristic values ​​(for example, it uses a weight of 0.2 for the first characteristic value and a weight of 0.8 for the second characteristic value).

[0053] Furthermore, if the design condition x is a discrete value, the output unit 132 takes a majority vote of the first characteristic value and the second characteristic value (i.e., it uses a weight of 0.5 for the first characteristic value and a weight of 0.5 for the second characteristic value). Alternatively, the output unit 132 takes a weighted majority vote of the first characteristic value and the second characteristic value (for example, it uses a weight of 0.2 for the first characteristic value and a weight of 0.8 for the second characteristic value).

[0054] In this way, the prediction device 130 outputs predicted data with optimized weights using a trained interpolation prediction model suitable for the input data in the interpolation region and a trained extrapolation prediction model suitable for the input data in the extrapolation region.

[0055] As a result, the prediction device 130 can obtain a certain level of prediction accuracy for the input data in the interpolation region, and also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to this embodiment, the prediction accuracy can be improved in a prediction device using a trained prediction model.

[0056] <Hardware configuration of the learning device and prediction device> Next, the hardware configurations of the learning device 120 and the prediction device 130 will be described. Since the learning device 120 and the prediction device 130 have similar hardware configurations, their hardware configurations will be described together using Figure 2.

[0057] Figure 2 shows an example of the hardware configuration of the learning device and the prediction device. As shown in Figure 2, the learning device 120 and the prediction device 130 each have a processor 201, memory 202, auxiliary storage device 203, I / F (Interface) device 204, communication device 205, and drive device 206. The hardware of the learning device 120 and the prediction device 130 are interconnected via a bus 207.

[0058] The processor 201 has various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 201 reads various programs (for example, training programs, prediction programs, etc.) into memory 202 and executes them.

[0059] Memory 202 has main memory devices such as ROM (Read Only Memory) and RAM (Random Access Memory). The processor 201 and memory 202 form a so-called computer, and the computer realizes various functions by having the processor 201 execute various programs read from memory 202.

[0060] The auxiliary storage device 203 stores various programs and various data used when those programs are executed by the processor 201.

[0061] The I / F device 204 is a connection device that connects to an external device (not shown). The communication device 205 is a communication device for communicating with an external device (e.g., a material data storage unit 110) via a network.

[0062] The drive device 206 is a device for setting the recording medium 210. The recording medium 210 here includes media that record information optically, electrically, or magnetically, such as CD-ROMs, flexible disks, and magneto-optical disks. The recording medium 210 may also include semiconductor memory that records information electrically, such as ROMs and flash memory.

[0063] The various programs to be installed on the auxiliary storage device 203 are installed, for example, when the distributed recording medium 210 is set in the drive device 206 and the various programs recorded on the recording medium 210 are read by the drive device 206. Alternatively, the various programs to be installed on the auxiliary storage device 203 may be installed by downloading them from the network via the communication device 205.

[0064] <Flow of learning and prediction processes> Next, we will explain the flow of the learning process and the prediction process. Figure 3 is the first flowchart showing the flow of the learning process and the prediction process.

[0065] In step S301, the learning device 120 acquires the training dataset 111.

[0066] In step S302, the learning device 120 uses the acquired training dataset 111 to train the interpolation prediction model 121_1 and the extrapolation prediction model 121_2, and generates the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2.

[0067] In step S303, the prediction device 130 inputs the input data to be predicted (design condition x) into the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2.

[0068] In step S304, the prediction device 130 acquires the first characteristic value and the second characteristic value predicted by the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2, respectively.

[0069] In step S305, the prediction device 130 calculates a weighted average of the acquired first characteristic value and second characteristic value, or determines the characteristic value by taking a weighted majority vote.

[0070] In step S306, the prediction device 130 outputs the determined characteristic value as prediction data for the input data (design condition x) to be predicted.

[0071] <Summary> As is clear from the above description, the prediction device 130 according to the first embodiment is It has a pre-trained interpolation prediction model suitable for the input data in the interpolation region, and a pre-trained extrapolation prediction model suitable for the input data in the extrapolation region. The predicted data is output by calculating the weighted average of the first characteristic value predicted by the trained interpolation prediction model and the second characteristic value predicted by the trained extrapolation prediction model, under optimized weights, or by taking a weighted majority vote.

[0072] As a result, the prediction device 130 according to the first embodiment can obtain a certain degree of prediction accuracy for the input data in the interpolation region, and can also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to the first embodiment, the prediction accuracy can be improved in a prediction device using a trained prediction model.

[0073] [Second Embodiment] In the first embodiment described above, the weights used when calculating the weighted average, or the weights used when taking a weighted majority vote, were pre-optimized. In contrast, in the second embodiment, the weights used when calculating the weighted average, or the weights used when taking a weighted majority vote, are pre-optimized for each input data to be predicted, and the prediction device switches between different weights depending on the input data to be predicted. The second embodiment will now be described, focusing on the differences from the first embodiment.

[0074] <Functional Configuration of the Prediction Device> First, the functional configuration of the prediction device according to the second embodiment will be explained using Figure 4. Figure 4 is a second diagram showing an example of the functional configuration of the learning device in the learning phase and the prediction device in the prediction phase. The difference from Figure 1 is that in Figure 4, the functional configuration of the prediction device 400 includes an interpolation / extrapolation discrimination unit 410 (see 4b in Figure 4).

[0075] The interpolation / extrapolation discrimination unit 410 determines whether the input data to be predicted (design condition x) is input data for the interpolation region or input data for the extrapolation region. The interpolation / extrapolation discrimination unit 410 also sets weights (weights for the interpolation region, weights for the extrapolation region) according to the result of the discrimination in the output unit 132.

[0076] In the interpolation / extrapolation discrimination unit 410, for example, if it determines that the input data to be predicted (design condition x) is input data for the interpolation region, it sets the weights for the interpolation region as follows: the weight of the first characteristic value = 0.8 and the weight of the second characteristic value = 0.2 in the output unit 132. Alternatively, the interpolation / extrapolation discrimination unit 410 may set the weights for the interpolation region as follows: the weight of the first characteristic value = 1.0 and the weight of the second characteristic value = 0.0 in the output unit 132.

[0077] Furthermore, if the interpolation / extrapolation discrimination unit 410 determines, for example, that the input data to be predicted (design condition x) is input data for the extrapolation region, it sets the weights for the extrapolation region as follows: the weight of the first characteristic value = 0.3 and the weight of the second characteristic value = 0.7 in the output unit 132. Alternatively, the interpolation / extrapolation discrimination unit 410 may set the weights for the extrapolation region as follows: the weight of the first characteristic value = 0.0 and the weight of the second characteristic value = 1.0 in the output unit 132.

[0078] The weights set in the output unit 132 (weights for the interpolation region and weights for the extrapolation region) are arbitrary. Furthermore, the method used by the interpolation / extrapolation discrimination unit 410 to distinguish between input data is also arbitrary.

[0079] For example, the interpolation / extrapolation discrimination unit 410 may train a one-class support vector machine using the training dataset 111, and then make the discrimination by inputting the input data to be predicted into the trained one-class support vector machine. In this case, if the interpolation / extrapolation discrimination unit 410 determines that the input data to be predicted is an outlier, it determines that the input data to be predicted is input data in the extrapolation region. Also, if the interpolation / extrapolation discrimination unit 410 does not determine that the input data to be predicted is an outlier, it determines that the input data to be predicted is input data in the interpolation region.

[0080] Alternatively, the interpolation / extrapolation discrimination unit 410 may use the local outlier factor method to pre-define the interpolation region from the training dataset 111, thereby determining whether the input data to be predicted is input data within the interpolation region.

[0081] Alternatively, the interpolation / extrapolation discrimination unit 410 may use a Gaussian mixture model to pre-define the interpolation region from the training dataset 111, thereby determining whether the input data to be predicted is input data within the interpolation region.

[0082] Alternatively, the interpolation / extrapolation discrimination unit 410 may use an isolation forest to pre-define the interpolation region from the training dataset 111, thereby determining whether the input data to be predicted is input data within the interpolation region.

[0083] <Flow of learning and prediction processes> Next, we will explain the flow of the learning process and prediction process. Figure 5 is a second flowchart showing the flow of the learning process and prediction process. The difference from the first flowchart explained using Figure 3 is step S501.

[0084] In step S501, the prediction device 400 determines whether the input data to be predicted (design condition x) is input data for the interpolation region or input data for the extrapolation region. The prediction device 400 also sets weights (weights for the interpolation region, weights for the extrapolation region) according to the determination result.

[0085] <Summary> As is clear from the above description, the prediction device 400 according to the second embodiment has the functions of the prediction device 130 according to the first embodiment, It has a function to determine whether the input data to be predicted is input data for the interpolation region or input data for the extrapolation region, and to set weights (weights for the interpolation region, weights for the extrapolation region) according to the determination result. • It has a function to calculate the weighted average of the first and second characteristic values, or to take a weighted majority vote, and to output prediction data using weights corresponding to the discrimination result.

[0086] As a result, the prediction device 130 according to the second embodiment can obtain a certain degree of prediction accuracy for the input data in the interpolation region, and can also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to the second embodiment, the prediction accuracy can be improved in a prediction device using a trained prediction model.

[0087] [Third Embodiment] In the second embodiment described above, it was explained that the system determines whether the input data to be predicted is input data for the interpolation region or input data for the extrapolation region, and outputs predicted data using weights (weights for the interpolation region, weights for the extrapolation region) according to the determination result.

[0088] In contrast, the third embodiment evaluates the strength of extrapolability (continuous value) of the input data to be predicted, and outputs predicted data with weights corresponding to the evaluation results. The third embodiment will be described below, focusing on the differences from the second embodiment described above.

[0089] <Functional Configuration of the Prediction Device> In the prediction device 400 according to the third embodiment, the interpolation / extrapolation discrimination unit 410 in Figure 4 evaluates the strength of extrapolatability of the input data to be predicted, instead of determining whether the input data to be predicted is input data for the interpolation region or input data for the extrapolation region. Also, in the prediction device 400 according to the third embodiment, the interpolation / extrapolation discrimination unit 410 in Figure 4 sets weights according to the evaluation result (weights for the interpolation region, weights for the extrapolation region), instead of setting weights according to the discrimination result, to the output unit 132.

[0090] Specifically, in the case of the prediction device 400 according to the third embodiment, the interpolation / extrapolation discrimination unit 410 continuously changes the weights based, for example, on the strength of extrapolability of the input data (design condition x) to be predicted. Continuously changing the weights means continuously changing them in increments of 0.1, for example, from (weight of the first characteristic value = 1.0, weight of the second characteristic value = 0.0) to (weight of the first characteristic value = 0.0, weight of the second characteristic value = 1.0) according to the evaluation result.

[0091] The method used by the interpolation / extrapolation discrimination unit 410 to evaluate the strength of extrapolatability of the input data is arbitrary. One example is the evaluation method using kernel density estimation. Specifically, the interpolation / extrapolation discrimination unit 410 first constructs a kernel density estimation model using the training dataset 111 and estimates the density of the input data included in the training dataset 111. Next, the interpolation / extrapolation discrimination unit 410 uses the constructed kernel density estimation model to estimate the density of the input data to be predicted (design condition x). Then, the interpolation / extrapolation discrimination unit 410 evaluates the strength of extrapolatability to the input data to be predicted (design condition x) by comparing the density of the input data included in the training dataset 111 with the density of the input data to be predicted (design condition x).

[0092] Alternatively, another example is an evaluation method based on distance. Specifically, the interpolation / extrapolation discrimination unit 410 first extracts α input data points from the training dataset 111 that are close in distance to the input data to be predicted (design condition x). Here, α is a value determined by the number of input data points included in the training dataset 111. Next, the interpolation / extrapolation discrimination unit 410 calculates the average distance between the extracted α input data points and the input data to be predicted (design condition x). Then, the interpolation / extrapolation discrimination unit 410 evaluates the strength of extrapolation from the calculated average distance.

[0093] Alternatively, another example is an evaluation method based on the uncertainty of random forest prediction. Specifically, the interpolation / extrapolation discrimination unit 410 first constructs a prediction model using a random forest with the training dataset 111, and calculates the standard deviation of the distribution of estimated values ​​for each tree when the input data to be predicted (design condition x) is input. Then, the interpolation / extrapolation discrimination unit 410 evaluates the strength of extrapolation from the calculated standard deviation.

[0094] Alternatively, another example is an evaluation method based on the uncertainty of Bayesian estimation. Specifically, the interpolation / extrapolation discriminant unit 410 uses the training dataset 111 to construct prediction models using variational Bayesian methods, Markov chain Monte Carlo methods to create Bayesian neural networks, or nonparametric Bayesian Gaussian processes. Next, the interpolation / extrapolation discriminant unit 410 calculates the standard deviation of the distribution of estimated values ​​when the input data to be predicted (design condition x) is input to the constructed Bayesian neural network or Gaussian process prediction model. Then, the interpolation / extrapolation discriminant unit 410 evaluates the strength of extrapolatability from the calculated standard deviation.

[0095] <Summary> As is clear from the above description, the prediction device 400 according to the third embodiment has the functions of the prediction device 130 according to the first embodiment, • It has a function to evaluate the strength of extrapolability of the input data to be predicted and set weights according to the evaluation results. • It has a function to calculate the weighted average of the first and second characteristic values, or to take a weighted majority vote, and to output prediction data using weights according to the evaluation results.

[0096] As a result, the prediction device 400 according to the third embodiment can obtain a certain degree of prediction accuracy for the input data in the interpolation region, and can also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to the third embodiment, the prediction accuracy of a prediction device using a trained prediction model can be improved.

[0097] [Fourth Embodiment] In the first and second embodiments described above, when the prediction devices 130 and 400 perform prediction processing, the output unit 132 is set with pre-optimized weights (or pre-optimized weights corresponding to the discrimination result or evaluation result of the input data to be predicted). In contrast, the fourth embodiment will describe a method for optimizing the weights set in the output unit 132.

[0098] <Functional Configuration of the Learning Device in the Optimization Phase, Part 1> First, we will describe the functional configuration of the learning device in the optimization phase, which optimizes the weights set in the output unit 132 (Figure 1) of the prediction device 130. Figure 6 is the first diagram showing an example of the functional configuration of the learning device in the optimization phase. As shown in Figure 6, the learning device 620 in the optimization phase is • Pre-trained interpolation prediction model 131_1 • Pre-trained extrapolation prediction model 131_2, Output section 621, • Weight change section 622, ·Error calculation unit 623, • Decision section 624, It functions as such.

[0099] The learning device 620 uses the validation dataset 610 stored in the material data storage unit 110 to optimize the weights set in the output unit 132 of the prediction device 130.

[0100] As shown in Figure 6, the validation dataset 610 includes "input data" and "ground truth data" as information items. The example in Figure 6 shows the case where "design conditions n+1" to "design conditions n+m" are stored as "input data" and "characteristic values ​​n+1" to "characteristic values ​​n+m" are stored as "ground truth data". In this way, when optimizing the weights, the learning device 620 uses a different validation dataset 610 than the training dataset 111 used when training the interpolation prediction model 121_1 and the extrapolation prediction model 121_2.

[0101] The trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2 are the same as the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2 described with reference to Figure 1 in the first embodiment described above.

[0102] However, in the fourth embodiment, the pre-trained interpolation prediction model 131_1 and the pre-trained extrapolation prediction model 131_2 are sequentially input with the "design conditions n+1" to "design conditions n+m" stored in the "input data" of the validation dataset 610. As a result, the pre-trained interpolation prediction model 131_1 and the pre-trained extrapolation prediction model 131_2 sequentially predict the first characteristic value and the second characteristic value.

[0103] The output unit 621 sequentially outputs prediction data based on the first characteristic value and the second characteristic value, using the weights modified by the weight modification unit 622. For example, the output unit 621 sequentially outputs multiple prediction data sets with multiple types of weights for the first and second characteristic values ​​predicted when "design condition n+1" is input to the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2. Similarly, the output unit 621 sequentially outputs multiple prediction data sets with multiple types of weights for the first and second characteristic values ​​predicted when "design condition n+2" is input to the trained interpolation prediction model 131_1 and the trained extrapolation prediction model 131_2.

[0104] The weight changing unit 622 sets the weights used by the output unit 621 when it sequentially outputs the predicted data. The weight changing unit 622 sets multiple types of weights, changing them in increments of 0.1, for example, between (weight of the first characteristic value = 1.0, weight of the second characteristic value = 0.0) and (weight of the first characteristic value = 0.0, weight of the second characteristic value = 1.0).

[0105] The error calculation unit 623 calculates the error between the multiple prediction data sequentially output from the output unit 621 and one of the "characteristic values ​​n+1" to "characteristic values ​​n+m" stored in the "ground truth data" of the verification dataset 610, and outputs it to the determination unit 624.

[0106] The determination unit 624 determines the optimal weight by referring to the table 630 in which the calculated errors are stored.

[0107] In Figure 6, Table 630 shows a list of errors calculated by the error calculation unit 623. In Table 630, • Weight A = (Weight of the first characteristic value = 1.0, Weight of the second characteristic value = 0.0) • Weight B = (Weight of the first characteristic value = 0.9, Weight of the second characteristic value = 0.1), • Weight C = (Weight of the first characteristic value = 0.8, Weight of the second characteristic value = 0.2), ... That is the case.

[0108] Also, in Table 630, The error A_n+1 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error B_n+1 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error C_n+1 refers to the error between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. ... The error A_n+2 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error B_n+2 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error C_n+2 refers to the error between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. ... The error A_n+m refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m. The error B_n+m refers to the difference between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m. The error C_n+m refers to the difference between the predicted data output by the output unit 621 under the weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m.

[0109] Also, in Table 820, • Error index A is a statistical value (e.g., mean) of the error A_n+1 to error A_n+m. • Error index B is a statistical value (e.g., mean) of the error B_n+1 to error B_n+m. The error index C is a statistical value (e.g., mean) of the error C_n+1 to error C_n+m. It refers to.

[0110] The determination unit 624 identifies the minimum value from, for example, error index A, error index B, error index C, ... and determines the corresponding weight as the optimal weight. The determination unit 624 also sets the determined weight in the output unit 132 of the prediction device 130.

[0111] This allows the prediction device 130 to perform prediction processing with optimized weights.

[0112] <Functional Configuration of the Learning Device in the Optimization Phase, Part 2> Next, we will describe the functional configuration of the learning device in the optimization phase, which optimizes the weights set in the output unit 132 (Figure 4) of the prediction device 400 according to the discrimination result of the input data to be predicted. Figure 7 is a second diagram showing an example of the functional configuration of the learning device in the optimization phase. The difference from the functional configuration of the learning device 620 shown in Figure 6 is that the function of the determination unit 711 of the learning device 710 is different from the function of the determination unit 624 of the learning device 620.

[0113] The determination unit 711 determines the optimal weights for the interpolation region and the optimal weights for the extrapolation region by referring to the table 720 in Figure 7.

[0114] In Table 720, the hatched errors (e.g., error A_n+2, error B_n+2, error C_n+2, ...) indicate the errors corresponding to the input data in the extrapolation region.

[0115] However, it is assumed that the input data of the verification dataset 610 (design conditions n+1, n+2, ...n+m) is determined in advance, for example, by the interpolation / extrapolation discrimination unit 410, to determine whether it is input data for the interpolation region or input data for the extrapolation region. Note that the interpolation / extrapolation discrimination unit 410 referred to here is the same as the interpolation / extrapolation discrimination unit 410 in Figure 4.

[0116] Also, in Table 720, • Error index A1 is the statistical value (e.g., mean) of the errors corresponding to the input data in the interpolation region from errors A_n+1 to A_n+m (errors without hatching (e.g., errors A_n+1, A_n+m)). • Error index A2 is the statistical value (e.g., mean) of the errors corresponding to the input data in the extrapolation region (hatched errors (e.g., error A_n+2)) from errors A_n+1 to A_n+m. The error index B1 is the statistical value (e.g., mean) of the errors corresponding to the input data in the interpolation region from errors B_n+1 to B_n+m (errors without hatching (e.g., errors B_n+1, B_n+m)). The error index B2 is the statistical value (e.g., mean) of the errors corresponding to the input data in the extrapolation region (hatched errors (e.g., error B_n+2)) from errors B_n+1 to B_n+m. The error index C1 is the statistical value (e.g., mean) of the errors corresponding to the input data in the interpolation region from errors C_n+1 to C_n+m (errors without hatching (e.g., errors C_n+1, C_n+m)). The error index C2 is the statistical value (e.g., mean) of the errors corresponding to the input data in the extrapolation region (hatched errors (e.g., error C_n+2)) from errors C_n+1 to C_n+m. It refers to.

[0117] Therefore, the determination unit 711 identifies the minimum value from among error indicators A1, B1, C1, ... and determines the corresponding weight as the optimal weight for the interpolation region. The determination unit 711 also notifies the prediction device 400 that the determined optimal weight for the interpolation region should be set in the output unit 132 of the prediction device 400.

[0118] Similarly, the determination unit 711 identifies the minimum value from error index A2, error index B2, error index C2, ... and determines the corresponding weight as the optimal weight for the extrapolation region. The determination unit 711 also notifies the prediction device 400 that the determined optimal weight for the extrapolation region should be set in the output unit 132 of the prediction device 400.

[0119] <Flow of learning and prediction processes> Next, we will explain the flow of the learning process and prediction process. Figure 8 is a third flowchart showing the flow of the learning process and prediction process. The difference from the first flowchart explained using Figure 1 is step S801.

[0120] In step S801, the learning device 120 performs an optimization process to optimize the weights (or the weights for each input data to be predicted) set in the output unit 132 of the prediction device 130 (or the output unit 132 of the prediction device 400). The details of the optimization process (step S801) will be explained below.

[0121] <Optimization Process Flow> Figure 9 is a first flowchart showing the flow of the optimization process. In step S901, the learning device 120 acquires a validation dataset.

[0122] In step S902, the learning device 120 sets a default weight from among several types of weights.

[0123] In step S903, the learning device 120 inputs the input data of the validation dataset into the trained interpolation prediction model and the trained extrapolation prediction model, respectively, to obtain the first characteristic value and the second characteristic value.

[0124] In step S904, the learning device 120 outputs predicted data based on the set weights, using the acquired first and second characteristic values. The learning device 120 also calculates the error between the predicted data and the corresponding ground truth data from the validation dataset.

[0125] In step S905, the learning device 120 determines whether or not all of the multiple types of weights have been set. If it is determined in step S905 that there are weights that have not been set (i.e., the answer in step S905 is NO), the process proceeds to step S906.

[0126] In step S906, the learning device 120 sets the next weight that has not been set, and returns to step S904.

[0127] On the other hand, if it is determined in step S905 that all weights have been set (i.e., the answer in step S905 is YES), the process proceeds to step S907.

[0128] In step S907, the learning device 120 determines whether all input data from the validation dataset has been input to the trained interpolation prediction model and the trained extrapolation prediction model, respectively. If it is determined in step S907 that there is input data that has not been input (i.e., the answer in step S907 is NO), the device proceeds to step S908.

[0129] In step S908, the learning device 120 processes the next input data from the validation dataset and returns to step S903.

[0130] On the other hand, if it is determined in step S907 that all input data has been entered (i.e., the answer in step S907 is YES), the process proceeds to step S909.

[0131] In step S909, the learning device 120 calculates an error index for each set weight (or for each set weight and for each interpolation / extrapolation). The learning device 120 also determines the weight that minimizes the calculated error index as the optimal weight (or the optimal weight for the interpolation region, or the optimal weight for the extrapolation region).

[0132] <Summary> As is clear from the above description, the learning devices 620 and 710 according to the fourth embodiment are It has a pre-trained interpolation prediction model suitable for the input data in the interpolation region, and a pre-trained extrapolation prediction model suitable for the input data in the extrapolation region. The input data from the validation dataset is input into the trained interpolation prediction model and the trained extrapolation prediction model, respectively, and the first and second characteristic values ​​are predicted. Based on the predicted first and second characteristic values, the predicted data is output under multiple types of weights, and the error with the ground truth data is calculated to determine the error index for each weight. Identify the weights corresponding to the smallest error index (or the smallest error index for each input data in the interpolation region / extrapolation region) and determine them as the optimal weights (or the optimal weights for the interpolation region and the optimal weights for the extrapolation region).

[0133] As a result, the prediction device 400 according to the fourth embodiment can obtain a certain degree of prediction accuracy for the input data in the interpolation region, and also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to the fourth embodiment, the prediction accuracy can be improved in a prediction device using a trained prediction model.

[0134] [Fifth Embodiment] In the fourth embodiment described above, the case in which weights are optimized using a trained interpolation prediction model and a trained extrapolation prediction model that have been trained with specific hyperparameters was explained.

[0135] In contrast, the fifth embodiment describes a case where hyperparameters are optimized in conjunction with weight optimization. The fifth embodiment will be described below, focusing on the differences from the first and fourth embodiments described above.

[0136] <Functional configuration of the learning device in the learning phase> First, the functional configuration of the learning device in the learning phase according to the fifth embodiment will be described. Figure 10 is a third figure showing an example of the functional configuration of the learning device in the learning phase. The difference from the functional configuration described using Figure 1a is that the learning device 1000 has a hyperparameter changing unit 1010.

[0137] The hyperparameter modification unit 1010 modifies the hyperparameters for the interpolation prediction model set in the interpolation prediction model 121_1. The hyperparameter modification unit 1010 also modifies the hyperparameters for the extrapolation prediction model set in the extrapolation prediction model 121_2.

[0138] As a result, the learning device 1000 trains the interpolation prediction model 121_1 and the extrapolation prediction model 121_2 using the training dataset 111 each time the hyperparameters are changed. Consequently, the learning device 1000 generates multiple trained interpolation prediction models and multiple trained extrapolation prediction models.

[0139] <Functional configuration of the learning device in the optimization phase> Next, the functional configuration in the optimization phase of the learning device according to the fifth embodiment will be described. Figure 11 is a third figure showing an example of the functional configuration of the learning device in the optimization phase. As shown in Figure 11, in the optimization phase, the learning device 1100 is • Pre-trained interpolation and prediction models 131_1_1, 131_1_2, 131_1_3, ... • Pre-trained extrapolation prediction models 131_2_1, 131_2_2, 131_2_3, ... Output section 621, • Weight change section 622, ·Error calculation unit 623, • Decision unit 1101, It functions as such.

[0140] The learning device 1100 uses the validation dataset 610 stored in the material data storage unit 110 to optimize the trained interpolation prediction model and trained extrapolation prediction model applied to the prediction device 130, and also optimizes the weights set in the output unit 132.

[0141] The trained interpolation prediction models 131_1_1, 131_1_2, 131_1_3, ... are multiple trained interpolation prediction models generated during the training phase by training with the training dataset 111 each time the hyperparameters are changed.

[0142] The trained extrapolation prediction models 131_2_1, 131_2_2, 131_2_3, ... are multiple trained extrapolation prediction models generated by training using the training dataset 111 each time the hyperparameters are changed during the training phase.

[0143] The pre-trained interpolation prediction models 131_1_1, 131_1_2, 131_1_3, ... and the pre-trained extrapolation prediction models 131_2_1, 131_2_2, 131_2_3, ... are sequentially inputted from the validation dataset 610. Furthermore, the pre-trained interpolation prediction models 131_1_1, 131_1_2, 131_1_3, ... and the pre-trained extrapolation prediction models 131_2_1, 131_2_2, 131_2_3, ... each sequentially predict multiple first and second characteristic values.

[0144] The output unit 621 outputs prediction data based on the first and second characteristic values, using the weights modified by the weight modification unit 622. For example, the output unit 621 sequentially outputs multiple prediction data sets with multiple types of weights for the first and second characteristic values ​​predicted when "design condition n+1" is input to the trained interpolation prediction model 131_1_1 and the trained extrapolation prediction model 131_2_1. Similarly, the output unit 621 sequentially outputs multiple prediction data sets with multiple types of weights for the first and second characteristic values ​​predicted when "design condition n+2" is input to the trained interpolation prediction model 131_1_2 and the trained extrapolation prediction model 131_2_2.

[0145] The weight changing unit 622 sets the weights used by the output unit 621 when it sequentially outputs the predicted data. The weight changing unit 622 sets multiple types of weights, changing them in increments of 0.1, for example, between (weight of the first characteristic value = 1.0, weight of the second characteristic value = 0.0) and (weight of the first characteristic value = 0.0, weight of the second characteristic value = 1.0).

[0146] The error calculation unit 623 calculates the error between the multiple prediction data sequentially output from the output unit 621 and one of the "characteristic values ​​n+1" to "characteristic values ​​n+m" stored in the "ground truth data" of the verification dataset 610, and outputs it to the determination unit 1101.

[0147] The determination unit 1101 determines the optimal hyperparameters and optimal weights by referring to tables 1111, 1112, 1113, ... which store the calculated errors.

[0148] In Figure 11, Table 1111 shows a list of errors calculated by the error calculation unit 623. In Table 1111, • Hyperparameter 1 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_1 and the hyperparameters for the extrapolation prediction model set in the trained extrapolation prediction model 131_2_1. This refers to... Also, in Table 1112, • Hyperparameter 2 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_2 and the hyperparameters for the extrapolation prediction model set in the trained interpolation prediction model 131_1_2. This refers to... Also, in Table 1113, • Hyperparameter 3 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_3 and the hyperparameters for the extrapolation prediction model set in the trained interpolation prediction model 131_1_3. It refers to.

[0149] Also, in Table 1111, • Weight A = (Weight of the first characteristic value = 1.0, Weight of the second characteristic value = 0.0) • Weight B = (Weight of the first characteristic value = 0.9, Weight of the second characteristic value = 0.1), • Weight C = (Weight of the first characteristic value = 0.8, Weight of the second characteristic value = 0.2), ... That is the case.

[0150] Also, in Table 1111, The error A_n+1 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error B_n+1 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error C_n+1 refers to the error between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. ... The error A_n+2 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error B_n+2 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error C_n+2 refers to the error between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. ... The error A_n+m refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m. The error B_n+m refers to the difference between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m. The error C_n+m refers to the difference between the predicted data output by the output unit 621 under the weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+m, and the correct data, which is the characteristic value n+m.

[0151] Also, in Table 1111, • Error index A is a statistical value (e.g., mean) of the error A_n+1 to error A_n+m. • Error index B is a statistical value (e.g., mean) of the error B_n+1 to error B_n+m. The error index C is a statistical value (e.g., mean) of the error C_n+1 to error C_n+m. It refers to.

[0152] The determination unit 1101 identifies the minimum value from among error indicators A, B, C, etc. for each of the hyperparameters 1, 2, 3, etc. The determination unit 1101 also identifies the minimum value from among the minimum values ​​identified for each hyperparameter (i.e., from among the minimum values ​​identified in table 1111, the minimum values ​​identified in table 1112, etc.).

[0153] Furthermore, the determination unit 1101 determines the hyperparameters and weights corresponding to the identified minimum value as the optimal hyperparameters and optimal weights. In addition, the determination unit 1101 notifies the prediction device 130 of the optimal trained interpolation prediction model and the optimal trained extrapolation prediction model, which are generated by setting each combination pointed to by the determined hyperparameters and performing training, along with the determined optimal weights.

[0154] As a result, the prediction device 130 can perform prediction processing using an optimized pre-trained interpolation prediction model, a pre-trained extrapolation prediction model, and optimized weights.

[0155] <Flow of learning and prediction processes> Next, we will explain the flow of the learning process and prediction process. Figure 12 is a fourth flowchart showing the flow of the learning process and prediction process. The difference from the first flowchart explained using Figure 1 is in steps S1201, S1202-S1204.

[0156] In step S1201, the learning device 1000 sets the default hyperparameters from among the multiple hyperparameters for the interpolation prediction model to the interpolation prediction model. The learning device 1000 also sets the default hyperparameters from among the multiple hyperparameters for the extrapolation prediction model to the extrapolation prediction model.

[0157] In step S1202, the learning device 1000 determines whether it has performed training with all of the multiple hyperparameters for the interpolation prediction model set in the interpolation prediction model 121_1. The learning device 1000 also determines whether it has performed training with all of the multiple hyperparameters for the extrapolation prediction model set in the extrapolation prediction model 121_2.

[0158] If it is determined in step S1202 that there are hyperparameters that have not been set (i.e., the answer is NO in step S1202), the process proceeds to step S1203.

[0159] In step S1203, the learning device 1000 sets the following hyperparameters for the interpolation prediction model and the following hyperparameters for the extrapolation prediction model, and then returns to step S301.

[0160] On the other hand, if it is determined in step S1202 that all hyperparameters have been set (i.e., the answer in step S1202 is YES), the process proceeds to step S1204.

[0161] In step S1204, the learning device 1100 performs an optimization process to optimize the hyperparameters and weights. The details of the optimization process (step S1204) for optimizing the hyperparameters and weights are described below.

[0162] <Optimization Process Flow> Figure 13 is a second flowchart showing the flow of the optimization process. The difference from the optimization process shown in Figure 9 is in steps S1301, S1302-S1304.

[0163] In step S1301, the learning device 1100 sets a default combination of a trained interpolation prediction model and a trained extrapolation prediction model from among the multiple trained interpolation prediction models and multiple trained extrapolation prediction models generated in the learning phase.

[0164] In step S1302, the learning device 1100 determines whether or not it has performed the processing in steps S902 to S908 for all combinations of the multiple trained interpolation prediction models and multiple trained extrapolation prediction models generated in the learning phase.

[0165] If, in step S1302, it is determined that there are combinations of trained interpolation and extrapolation models that have not undergone the processing in steps S902 to S908 (i.e., the answer in step S1302 is NO), then the process proceeds to step S1303.

[0166] In step S1303, the learning device 1100 sets the next combination of a trained interpolation prediction model and a next trained extrapolation prediction model, and returns to step S902.

[0167] On the other hand, if it is determined in step S1302 that the processes in steps S902 to S908 have been executed for all combinations (i.e., the answer in step S1302 is YES), then the process proceeds to step S1304.

[0168] In step S1304, the learning device 1100 determines the optimal combination of hyperparameters and optimal weights based on the error index.

[0169] <Summary> As is clear from the above description, the learning devices 1000 and 1100 according to the fifth embodiment are By setting multiple hyperparameters for each interpolation prediction model and training the model, multiple trained interpolation prediction models are generated. By setting multiple hyperparameters for each extrapolation prediction model and training the extrapolation prediction model, multiple trained extrapolation prediction models are generated. The input data from the validation dataset is input into all combinations of multiple trained interpolation and extrapolation models, and the first and second characteristic values ​​are predicted. Based on the predicted first and second characteristic values, the predicted data is output under multiple types of weights, and the error with the ground truth data is calculated. This allows for the calculation of an error index for each of the multiple types of weights for each combination of hyperparameters. The system determines the combination of hyperparameters and weights corresponding to the smallest error index, notifies the prediction device of the optimal trained interpolation prediction model and the optimal trained extrapolation prediction model, and also notifies the prediction device of the optimal weights.

[0170] As a result, the prediction device 130 according to the fifth embodiment can obtain a certain degree of prediction accuracy for the input data in the interpolation region, and can also obtain sufficient prediction accuracy for the input data in the extrapolation region. In other words, according to the fifth embodiment, the prediction accuracy can be improved in a prediction device using a trained prediction model.

[0171] [Sixth Embodiment] In the fifth embodiment described above, the case where the interpolation prediction model and the extrapolation prediction model are optimized assuming that they are each trained under a specific learning method was explained. In contrast, the sixth embodiment explains the case where the learning method used in the interpolation prediction model and the extrapolation prediction model, the set hyperparameters, and the weights are optimized.

[0172] <Functional configuration of the learning device in the learning phase> First, the functional configuration in the learning phase of the learning device according to the sixth embodiment will be described. Figure 14 is the fourth figure showing an example of the functional configuration of the learning device in the learning phase. The difference from the functional configuration described using Figure 10 is that, in the case of the learning device 1400, • Multiple interpolation prediction models 121_1_1, 121_1_2, 121_1_3, ..., which are trained using different learning methods, • Multiple extrapolation prediction models 121_2_1, 121_2_2, 121_2_3, ..., which are trained using different learning methods, It has the characteristic of having.

[0173] In the learning device 1400, each time the hyperparameters are changed, the interpolation prediction model 121_1_1 and the extrapolation prediction model 121_2_1 are trained using the training dataset 111. As a result, the learning device 1400 obtains multiple trained interpolation prediction models and multiple trained models from the interpolation prediction model 121_1_1 and the extrapolation prediction model 121_2_1. outside Generate an interpolation prediction model.

[0174] Next, the learning device 1400 trains the interpolation prediction model 121_1_2 and the extrapolation prediction model 121_2_2 using the training dataset 111 each time the hyperparameters are changed. As a result, the learning device 1400 obtains multiple trained interpolation prediction models and multiple trained outside Generate an interpolation prediction model.

[0175] Next, the learning device 1400 trains the interpolation prediction model 121_1_3 and the extrapolation prediction model 121_2_3 using the training dataset 111 each time the hyperparameters are changed. As a result, the learning device 1400 obtains multiple trained interpolation prediction models and multiple trained outside Generate an interpolation prediction model.

[0176] Note that in Figure 14, for the sake of simplicity, three interpolation and three extrapolation models with different learning methods are shown; however, the number of interpolation and extrapolation models with different learning methods is not limited to three.

[0177] <Functional configuration of the learning device in the optimization phase> Next, the functional configuration in the optimization phase of the learning device according to the sixth embodiment will be described. Figure 15 is the fourth figure showing an example of the functional configuration of the learning device in the optimization phase. As shown in Figure 15, in the optimization phase, the learning device 1500 is • Pre-trained interpolation and prediction models 131_1_1, 131_1_2, 131_1_3, ... • Pre-trained extrapolation prediction models 131_2_1, 131_2_2, 131_2_3, ... Output section 621, • Weight change section 622, ·Error calculation unit 623, • Decision section 1501, It functions as such.

[0178] The learning device 1500 uses the validation dataset 610 stored in the material data storage unit 110 to optimize the trained interpolation prediction model and trained extrapolation prediction model applied to the prediction device 130, and also optimizes the weights set in the output unit 132.

[0179] The trained interpolation prediction models 131_1_1 to 131_1_3 are multiple trained interpolation prediction models generated during the training phase by training interpolation prediction model 121_1_1 each time the hyperparameters for the interpolation prediction model are changed.

[0180] Furthermore, the trained interpolation prediction models 131_1_4 to 131_1_6 are multiple trained interpolation prediction models generated during the training phase by training interpolation prediction model 121_1_2 each time the hyperparameters for the interpolation prediction model are changed.

[0181] Furthermore, the trained interpolation prediction models 131_1_7 to 131_1_9 are multiple trained interpolation prediction models generated during the training phase by training interpolation prediction model 121_1_3 each time the hyperparameters for the interpolation prediction model are changed.

[0182] On the other hand, the trained extrapolation prediction models 131_2_1 to 131_2_3 are multiple trained extrapolation prediction models that are generated when extrapolation prediction model 121_2_1 is trained each time the hyperparameters for the extrapolation prediction model are changed during the training phase.

[0183] Furthermore, the trained extrapolation prediction models 131_2_4 to 131_2_6 are multiple trained extrapolation prediction models generated during the training phase by training extrapolation prediction model 121_2_2 each time the hyperparameters for the extrapolation prediction model are changed.

[0184] Furthermore, the trained extrapolation prediction models 131_2_7 to 131_2_9 are multiple trained extrapolation prediction models generated during the training phase by training extrapolation prediction model 121_2_3 each time the hyperparameters for the extrapolation prediction model are changed.

[0185] The pre-trained interpolation prediction model 131_1_1 and the pre-trained extrapolation prediction model 131_2_1 are sequentially fed the input data from the validation dataset 610. As a result, the pre-trained interpolation prediction model 131_1_1 and the pre-trained extrapolation prediction model 131_2_1 sequentially predict multiple first and second characteristic values, respectively.

[0186] The output section 621 to the error calculation section 623 are the same as those in Figure 11, so their explanation is omitted here.

[0187] The decision unit 1501 determines the optimal learning method, optimal hyperparameters, and optimal weights by referring to tables 1511 to 1519.

[0188] In Figure 15, Tables 1511 to 1519 show a list of errors calculated by the error calculation unit 623. In Tables 1511 to 1519, • Learning method 1 is a combination of the learning method used when training the interpolation prediction model 121_1_1 and the learning method used when training the extrapolation prediction model 121_2_1. • Learning method 2 is a combination of the learning method used when training the interpolation prediction model 121_1_2 and the learning method used when training the extrapolation prediction model 121_2_2. • Learning method 3 is a combination of the learning method used when training the interpolation prediction model 121_1_3 and the learning method used when training the extrapolation prediction model 121_2_3. It refers to.

[0189] Also, in tables 1511, 1512, and 1513, • Hyperparameter 1 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_1 and the hyperparameters for the extrapolation prediction model set in the trained extrapolation prediction model 131_2_1. • Hyperparameter 2 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_2 and the hyperparameters for the extrapolation prediction model set in the trained extrapolation prediction model 131_2_2. • Hyperparameter 3 is a combination of the hyperparameters for the interpolation prediction model set in the trained interpolation prediction model 131_1_3 and the hyperparameters for the extrapolation prediction model set in the trained extrapolation prediction model 131_2_3. It refers to.

[0190] Also, in Table 1511, • Weight A = (Weight of the first characteristic value = 1.0, Weight of the second characteristic value = 0.0) • Weight B = (Weight of the first characteristic value = 0.9, Weight of the second characteristic value = 0.1), • Weight C = (Weight of the first characteristic value = 0.8, Weight of the second characteristic value = 0.2), ... That is the case.

[0191] Also, in Table 1511, The error A_n+1 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error B_n+1 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. The error C_n+1 refers to the error between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​obtained by inputting the design condition n+1, and the correct data, which is the characteristic value n+1. ... The error A_n+2 refers to the error between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error B_n+2 refers to the error between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​obtained by inputting the design condition n+2, and the correct data, which is the characteristic value n+2. The error C_n+2 refers to the difference between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​predicted by the input design condition n+2, and the correct data, which is the characteristic value n+2. ... The error A_n+m refers to the difference between the predicted data output by the output unit 621 under weight A, based on the first and second characteristic values ​​predicted by the input design condition n+m, and the correct data, which is the characteristic value n+m. The error B_n+m refers to the difference between the predicted data output by the output unit 621 under weight B, based on the first and second characteristic values ​​predicted by the input design condition n+m, and the correct data, which is the characteristic value n+m. The error C_n+m refers to the difference between the predicted data output by the output unit 621 under weight C, based on the first and second characteristic values ​​predicted by the input design condition n+m, and the correct data, which is the characteristic value n+m.

[0192] Also, in Table 1511, • Error index A is a statistical value (e.g., mean) of the error A_n+1 to error A_n+m. • Error index B is a statistical value (e.g., mean) of the error B_n+1 to error B_n+m. The error index C is a statistical value (e.g., mean) of the error C_n+1 to error C_n+m. It refers to.

[0193] The decision unit 1501 identifies the minimum value from among error indices A, B, C, ... for learning methods 1, 2, and 3, and hyperparameters 1, 2, and 3. Furthermore, the decision unit 1501 identifies a further minimum value from among the minimum values ​​identified for each learning method and each hyperparameter. As a result, the decision unit 1501 determines the corresponding learning method, corresponding hyperparameter, and corresponding weight as the optimal learning method, optimal hyperparameter, and optimal weight.

[0194] Furthermore, the decision unit 1501 notifies the prediction device 130 of the trained interpolation prediction model and the trained extrapolation prediction model, which are generated by setting the combination indicated by the determined hyperparameters under the combination indicated by the determined learning method and performing learning. The decision unit 1501 also notifies the prediction device 130 of the determined weights.

[0195] As a result, the prediction device 130 can perform prediction processing using an optimized pre-trained interpolation prediction model, a pre-trained extrapolation prediction model, and optimized weights.

[0196] <Flow of learning and prediction processes> Next, we will explain the flow of the learning process and prediction process. Figure 16 is the fifth flowchart showing the flow of the learning process and prediction process. The difference from the fourth flowchart explained using Figure 12 is in steps S1601, S1602-S1604.

[0197] In step S1601, the learning device 1400 sets an interpolation prediction model that is trained under a default learning method from among multiple interpolation prediction models that are trained under different learning methods. The learning device 1500 also sets an extrapolation prediction model that is trained under a default learning method from among multiple extrapolation prediction models that are trained under different learning methods.

[0198] In step S1602, the learning device 1400 trains the interpolation prediction model under all of the pre-prepared learning methods and determines whether or not it has generated a trained interpolation prediction model. The learning device 1400 also trains the extrapolation prediction model under all of the pre-prepared learning methods and determines whether or not it has generated a trained extrapolation prediction model.

[0199] If it is determined in step S1602 that there is a learning method that has not been trained (i.e., the answer is NO in step S1602), proceed to step S1603.

[0200] In step S1603, the learning device 1400 sets up an interpolation prediction model or an extrapolation prediction model that will be trained under the next learning method, and returns to step S1201.

[0201] On the other hand, if it is determined in step S1602 that learning has been performed under all pre-prepared learning methods (YES in step S1602), then the process proceeds to step S1604.

[0202] In step S1604, the learning device 1500 performs an optimization process to optimize the learning method, hyperparameters, and weights. The details of the optimization process (step S1604) for optimizing the learning method, hyperparameters, and weights are described below.

[0203] <Optimization Process Flow> Figure 17 is a third flowchart showing the flow of the optimization process. The difference from the optimization process shown in Figure 13 is step S1701.

[0204] In step S1701, the learning device 1500 determines the optimal combination of learning methods, the optimal combination of hyperparameters, and the optimal weights based on the error index.

[0205] <Summary> As is clear from the above description, the learning device 1500 according to the sixth embodiment · For a plurality of interpolation prediction models that are learned under different learning methods, by setting a plurality of hyperparameters for each interpolation prediction model and performing learning, a plurality of learned interpolation prediction models are generated. · For a plurality of extrapolation prediction models that are learned under different learning methods, by setting a plurality of hyperparameters for each extrapolation prediction model and performing learning, a plurality of learned extrapolation prediction models are generated. · Input the input data of the verification dataset into all combinations of the plurality of learned interpolation prediction models and the plurality of learned extrapolation prediction models, and predict the first characteristic value and the second characteristic value. · Based on the predicted first characteristic value and second characteristic value, output prediction data under a plurality of types of weights respectively, and calculate the error from the correct data. Thereby, for each combination of learning methods and combination of hyperparameters, an error index for each of the plurality of types of weights is calculated. · Determine the combination of learning methods, combination of hyperparameters, and weights corresponding to the minimum error index, notify the prediction device of the optimal learned interpolation prediction model and learned extrapolation prediction model, and notify the prediction device of the optimal weights.

[0206] According to this, according to the prediction device 130 according to the sixth embodiment, a certain degree of prediction accuracy can be obtained for the input data in the interpolation region, and sufficient prediction accuracy can also be obtained for the input data in the extrapolation region. That is, according to the sixth embodiment, the prediction accuracy of the prediction device using the learned prediction model can be improved.

[0207] [Examples] Next, a specific example of the fourth embodiment among the embodiments described above will be explained. For the purposes of this explanation, it will be assumed that the material data storage unit 110 stores, for example, a solubility dataset for 1311 molecules that is publicly available on ALOGPS (http: / / www.vcclab.org / lab / alogps / ).

[0208] When performing learning and prediction processing using the solubility dataset, according to the second embodiment, the processing is carried out, for example, by the following procedure.

[0209] (1) Step 1 The molecular structures described in SMILES format in the solubility dataset are converted into 187-dimensional feature vectors using RDKit's rdkit.Chem.Descriptors.

[0210] (2) Step 2 The solubility dataset, transformed into 187-dimensional feature vectors, is randomly split into training, validation, and prediction datasets in the proportions of 56.25%, 18.75%, and 25%.

[0211] (3) Step 3 We will train a random forest regression model using scikit-learn, which is an interpolation prediction model, with the training dataset. We will also train a Gaussian process regression model, which is an extrapolation prediction model, using the same training dataset.

[0212] (4) Step 4 The trained random forest regression model acquired in step 3 is used as the interpolation / extrapolation discrimination unit 410, and the standard deviation of each predicted value is calculated by inputting each input data from the validation dataset. If the calculated standard deviation is less than 0.6, the corresponding input data is determined to be input data in the interpolation region. If the calculated standard deviation is 0.6 or greater, the corresponding input data is determined to be input data in the extrapolation region. The threshold = 0.6 is the median of the standard deviations of each predicted value.

[0213] (5) Step 5-1 For each input data point in the validation dataset, the input data from the interpolation region is input into a trained interpolation prediction model (a trained random forest regression model) to predict the first characteristic value. Similarly, for each input data point in the validation dataset, the input data from the interpolation region is input into a trained extrapolation prediction model (a trained Gaussian process regression model) to predict the second characteristic value. The predicted first and second characteristic values ​​are compared with the ground truth data in the validation dataset to optimize the "weight of the first characteristic value:weight of the second characteristic value". As a result, in this embodiment, a weight ratio of 0.55:0.45 was obtained for the input data in the interpolation region.

[0214] (6) Step 5-2 For each input data in the validation dataset, the input data from the extrapolation region is input into a trained extrapolation prediction model (a trained random forest regression model) to predict the first characteristic value. Similarly, for each input data in the validation dataset, the input data from the extrapolation region is input into a trained extrapolation prediction model (a trained Gaussian process regression model) to predict the second characteristic value. The predicted first and second characteristic values ​​are compared with the ground truth data in the validation dataset to optimize the "weight of the first characteristic value:weight of the second characteristic value". As a result, in this embodiment, a weight ratio of 0.40:0.60 was obtained for the input data in the extrapolation region.

[0215] (7) Step 6 For each input data point in the prediction dataset, perform the same processing as in step 4 above to determine whether it is input data for the interpolation region or input data for the extrapolation region.

[0216] (8) Step 7 Each input data point from the prediction dataset is input into a trained interpolation prediction model (a trained random forest regression model) to predict the first characteristic value. Additionally, each input data point from the prediction dataset is input into a trained extrapolation prediction model (a trained Gaussian process regression model) to predict the second characteristic value. The predicted first and second characteristic values ​​are then weighted according to the discrimination result in step 6, and the predicted data is output. For example, if the data is determined to be an interpolation input, a weighted average of the weights of the first and second characteristic values ​​(1st characteristic value:2nd characteristic value = 0.55:0.45) is output as the predicted data. If the data is determined to be an extrapolation input, a weighted average of the weights of the first and second characteristic values ​​(1st characteristic value:2nd characteristic value = 0.40:0.60) is output as the predicted data.

[0217] (9) Step 8 The prediction accuracy of the prediction data calculated for each input data in the prediction dataset is defined by the square of the correlation coefficient, R. 2 The evaluation is performed as follows: In this process, the prediction accuracy of the predicted data is compared with that of the predicted data when all input data of the prediction dataset is input into a trained interpolation prediction model (a trained random forest regression model) (Comparative Example 1). Furthermore, the prediction accuracy of the predicted data is compared with that of the predicted data when all input data of the prediction dataset is input into a trained extrapolation prediction model (a trained Gaussian process regression model) (Comparative Example 2).

[0218] Figure 18 shows an example of prediction accuracy. As shown in Figure 18, it can be seen that the prediction accuracy of this embodiment is higher than that of Comparative Examples 1 and 2, whether the input data is from the interpolation region or the extrapolation region.

[0219] In this way, by calculating a weighted average value based on optimal weights, the prediction error from the random forest regression model and the prediction error from the Gaussian process regression model can be offset.

[0220] [Other embodiments] In each of the above embodiments, the learning device and the prediction device have been described as separate devices. However, the learning device and the prediction device may be configured by an integrated device. Also, in the above-described third to sixth embodiments, the learning device in the learning phase and the learning device in the optimization phase have been described as separate devices. However, the learning device in the learning phase and the learning device in the optimization phase may be configured by an integrated device.

[0221] In addition, in the sixth embodiment above, no specific example of the learning method was mentioned. However, the learning method used when learning the interpolation prediction model is preferably a decision tree-based ensemble method such as a decision tree, random forest, gradient boosting, bagging, or AdaBoost. This is because decision tree-based ensemble methods tend to be prone to overfitting and can achieve high prediction accuracy for input data in the interpolation region.

[0222] Alternatively, the learning method used when learning the interpolation prediction model may be the k-nearest neighbor method, which is strongly influenced by the learning dataset.

[0223] Alternatively, the learning method used when learning the interpolation prediction model may be a neural network. In particular, when using a neural network with two or more hidden layers, there is a tendency to be prone to overfitting, and high prediction accuracy can be achieved for input data in the interpolation region.

[0224] On the other hand, the learning method used when learning the extrapolation prediction model is preferably, for example, a Gaussian process. This is because Gaussian processes tend to be less prone to overfitting and can achieve relatively high prediction accuracy for input data in the extrapolation region.

[0225] Alternatively, when training an extrapolation prediction model, it is preferable to use kernel-based learning methods such as kernel ridges or support vector machines, in addition to Gaussian processes. This is because kernel-based learning methods can achieve high prediction accuracy for input data in the extrapolation domain by appropriately setting the kernel function.

[0226] Alternatively, the learning method used when training an extrapolation prediction model may be a linear learning method such as linear, partial least squares, lasso, linear ridge, elastic net, or Bayesian ridge. Linear learning methods are effective in cases where the material properties in the extrapolation domain are expected to exhibit linear behavior, or where the behavior of the material properties can be linearized by preprocessing the input data of the training dataset.

[0227] Alternatively, the learning method used when training an extrapolation prediction model may be a neural network. In particular, using a neural network with two or fewer hidden layers tends to reduce overfitting, and furthermore, by appropriately setting the activation function, high prediction accuracy can be achieved for input data in the extrapolation region.

[0228] Furthermore, although the above embodiments did not mention how to use the prediction data output by the prediction device, the prediction data can be used, for example, to determine the design conditions when manufacturing the target material. Also, when the material is manufactured by the material manufacturing device under the determined design conditions, the determined design conditions and the measured characteristic values ​​can be added to the training dataset by measuring the characteristic values ​​of the manufactured material. Moreover, by configuring the system to retrain the interpolation prediction model or the extrapolation prediction model using this training dataset, the development cycle of the prediction model can be repeated.

[0229] It should be noted that the present invention is not limited to the configurations shown in the above embodiments, including combinations with other elements. These aspects can be modified without departing from the spirit of the present invention and can be appropriately determined according to their application.

[0230] This application claims priority based on Japanese Patent Application No. 2021-083921, filed on 18 May 2021, which is incorporated herein by reference to the entire contents of the said Japanese Patent Application. [Explanation of Symbols]

[0231] 111: Training dataset 120: Learning device 121_1: Interpolation Prediction Model 121_2: Extrapolation Prediction Model 130: Prediction device 131_1: Pre-trained interpolation and prediction model 131_2: Pre-trained extrapolation predictive model 132: Output section 400: Prediction device 410: Interpolation / extrapolation discriminator 610: Validation dataset 620: Learning device 621: Output section 622: Weight change section 623:Error calculation section 624: Decision Section 710: Learning device 711: Decision-making section 1000: Learning device 1010: Hyperparameter modification section 1100: Learning device 1101: Decision Section 1400: Learning device 1500: Learning device 1501: Decision Section

Claims

1. The process involves inputting the data to be predicted, causing the first trained model and the second trained model to output the first output data and the second output data, respectively. A step of determining or evaluating whether the input data to be predicted is input data in the interpolation region or input data in the extrapolation region, or the strength of the extrapolatability of the input data to be predicted, The process of outputting predicted data by setting weights for the first output data and the second output data output by the first trained model and the second trained model, respectively, according to the result of the discrimination or evaluation, and calculating a weighted average value under the weights, or by taking a weighted majority vote under the weights, A process of determining the design conditions for manufacturing the material using the aforementioned prediction data, The computer executes this, The process of outputting the prediction data involves setting a higher weight for the first output data with respect to the input data of the interpolation region, and setting a higher weight for the second output data with respect to the input data of the extrapolation region. Material design methods.

2. The evaluation process described above is: A method for designing a material according to claim 1, wherein the strength of extrapolability of the input data to be predicted is evaluated by using one or more of the following: an evaluation method based on the uncertainty of random forest prediction, an evaluation method based on the uncertainty of Bayesian estimation, an evaluation method based on kernel density estimation, and an evaluation method based on distance.

3. The first pre-trained model described above is trained using one or more of the following learning methods: decision trees, random forests, gradient boosting, bagging, AdaBoost, k-nearest neighbors, and neural networks. The material design method according to claim 1, wherein the second trained model is trained using one or more learning methods selected from Gaussian processes, kernel ridges, support vector machines, linear, partial least squares, Lasso, linear ridges, elastic networks, Bayesian ridges, and neural networks.

Images