Material manufacturing method

Abstract

To provide a material manufacturing method that improves prediction accuracy of a trained prediction model by setting appropriate weight using a trained clustering model and a classified cluster.SOLUTION: A prediction model creation method for material properties comprises steps of: obtaining a training data set; generating a trained clustering model using the training data set and a clustering model, and classifying the training data set into N clusters; calculating a distance between gravity centers of each cluster; calculating weight between the clusters using the distance between the gravity centers of the clusters and a parameter representing a feature of the training data set; and generating, for each cluster, a trained predictive model {Mi}1≤i≤N using the cluster and the weight.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This application claims priority to Basic Application No. 2021-159474, filed with the Japan Patent Office on September 29, 2021, the entire contents of which are incorporated herein by reference.

[0002] This disclosure relates to a method for creating a predictive model that takes clustering and weights into consideration, a predictive method, a predictive model creation apparatus, a predictive apparatus, a predictive model creation program, and a predictive program. [Background technology]

[0003] Traditionally, material design has been carried out through repeated prototyping based on the experience of material developers. In this case, a vast number of experiments are conducted to obtain the desired properties. Therefore, in recent years, attempts have been made to apply machine learning to material design. For example, by collecting the design conditions during prototyping and the evaluation results of the prototyped material (material property values, etc.), a model can be trained using this dataset, and the resulting trained model can be used to predict the property values ​​of the prototyped material under new design conditions. This can reduce the number of experiments required to obtain the desired properties to a minimum. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-187417 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] For example, Patent Document 1 discloses a method for predicting physical properties by clustering training data and using a first predetermined number of training data points near the representative vector in each cluster to predict physical property values, and a correction model that uses a second predetermined number of training data points near the representative vector to predict the inverse of the residuals for each base model. In predicting physical properties, the method involves searching for a base model and a correction model related to a representative vector close to the unknown input vector for an unknown input vector, calculating the predicted values ​​of the base model and the correction model, and calculating the predicted physical property value by summing the predicted values ​​of the base model and the correction model multiplied by a predetermined constant. However, in the case of the physical property prediction method disclosed in Patent Document 1, training data not included in the first and second predetermined numbers near the representative vector are not used to train the model, which may lead to a decrease in the accuracy of physical property prediction, and there is also a problem of being prone to overfitting.

[0006] This disclosure has been made with the above points in mind, and aims to provide a method for creating a predictive model that improves prediction accuracy by making full use of the training data. [Means for solving the problem]

[0007] The present invention has the following configuration. [1] A method for creating a predictive model of material properties, Steps to obtain a training dataset, The steps include generating a trained clustering model using the aforementioned training dataset and clustering model, and classifying the training dataset into N clusters, The steps include calculating the distance between the centroids of each cluster, A step of calculating the weights between the clusters using the distance between the centroids of the clusters and parameters representing the characteristics of the training dataset, For each of the aforementioned clusters, a trained predictive model {M} is created using the cluster and the aforementioned weights. i} 1≦i≦N A method for creating a predictive model of material properties, characterized by comprising the steps of generating and [2] A method for predicting material properties that is carried out following the method for creating a prediction model described in item 1 above, comprising the steps of obtaining prediction data, using the learned clustering model to identify that the prediction data belongs to cluster p among the learning data set clusters classified into N pieces, and using the learned prediction model M corresponding to the cluster p to obtain a predicted value with the prediction data as an input, wherein the method for predicting material properties is characterized by having these steps. p [3] The method for creating a prediction model for material properties described in item 1 above, in the step of constructing the clustering model described in item 1 above, using any one or more clustering methods selected from the K-means method, Nearest Neighbor method, hierarchical clustering method, Gaussian mixture method, DBSCAN method, t-SNE method, and self-organizing map method, wherein the method for creating a prediction model for material properties is characterized by this. [4] The method for creating a prediction model for material properties described in item 1 above, in the step of calculating the distance between the centroids of the clusters described in item 1 above, calculating the distance using any one or more combinations selected from the Euclidean distance method, Manhattan distance method, Mahalanobis distance method, Minkowski distance method, cosine distance method, single linkage method, complete linkage method, centroid method, group average method, ward method, Kullback-Leibler divergence, Jensen-Shannon divergence, Dynamic time warping, and Earth mover's distance, wherein the method for creating a prediction model for material properties is characterized by this. [5] The method for creating a prediction model for material properties described in item 1 above, using, as parameters representing the characteristics of the learning data set, any one or more parameters selected from systematic error, standard deviation, variance, coefficient of variation, quantile, kurtosis, and skewness related to the characteristic values of the learning data set, wherein the method for creating a prediction model for material properties is characterized by this. [6] The method for creating a prediction model of material properties according to item 1 above, In the step of calculating the weights, one or more weight functions selected from an exponential function type, a reciprocal type, and a power of reciprocal type are used. A method for creating a prediction model of material properties, characterized by this. [7] A device for creating a prediction model of material properties, By inputting a learning dataset, a learned clustering model is generated, and a clustering model that classifies the learning dataset into N clusters, A weight definition unit that calculates the distances between the centroids of the classified clusters, and calculates the weights between the clusters using the calculated distances between the centroids of the clusters and the parameters representing the characteristics of the learning dataset, For each cluster, a learned prediction model M is generated using the cluster and the weights i} 1≦i≦N A prediction model creation device, characterized by having the above. [8] A prediction device for material properties, By inputting prediction data, a learned clustering model created by the prediction model creation device according to item 7 above, which identifies that the prediction data belongs to cluster p among the N - classified clusters, A learned prediction model M created by the prediction model creation device according to item 7 above, which corresponds to the identified cluster p and obtains a predicted value with the prediction data as an input p And, An output unit that outputs the obtained predicted value. A prediction device for material properties, characterized by having the above. [9] A program for creating a prediction model of material properties, A step of acquiring a learning dataset, Using the learning dataset and a clustering model, a learned clustering model is generated, and the learning dataset is classified into N clusters, A step of calculating the distances between the centroids of the clusters, A step of calculating weights between the clusters using the distance between the centroids of the clusters and parameters representing the characteristics of the learning dataset; For each of the clusters, using the cluster and the weights, a trained prediction model {M i} 1≦i≦N is generated; A prediction model creation program for causing a computer to execute.

[10] A prediction program for material properties, A step of acquiring prediction data; Using the trained clustering model created by the prediction model creation program described in paragraph 9 above, a step of specifying that the prediction data belongs to cluster p among the learning dataset clusters classified into N; With the prediction data as input, corresponding to the specified cluster p, a step of obtaining a predicted value using the prediction model M described in paragraph 9 above; A prediction program for causing a computer to execute. p [Advantages of the Invention]

[0008] The prediction model created using the prediction model creation method of the present disclosure suppresses overfitting due to insufficient data volume by using all the training data without omission, and can improve the prediction accuracy by introducing weights that reflect the trend of the data. [Brief Description of the Drawings]

[0009] [Figure 1] It is a diagram showing an example of the functional configuration of a prediction model creation device in the learning phase and a prediction device in the prediction phase. [Figure 2] It is a diagram showing an example of the hardware configuration of a prediction model creation device and a prediction device. [Figure 3] It is a flowchart showing the flow of learning processing. [Figure 4] It is a flowchart showing the flow of prediction processing. [Figure 5]This is an illustrative diagram of an example of a requirements setting screen in the predictive model creation method according to this embodiment. [Figure 6] This flowchart shows the learning process flow for a standard random forest in the comparative example. [Figure 7] This flowchart shows the prediction process for a standard random forest in the comparative example. [Figure 8] This figure shows an example of prediction accuracy. [Modes for carrying out the invention]

[0010] 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.

[0011] <Functional Configuration of Predictive Model Creation Device and Prediction Device> First, the functional configurations of the predictive model creation device and the predictive device will be described. The predictive model creation device will be described using an example that generates a predictive model using a training dataset that includes the design conditions at the time of prototyping and the characteristic values ​​of the prototyped material. The predictive device will be described using an example that predicts the characteristic values ​​of a prototyped material under new design conditions using the trained predictive model created by the predictive model creation device.

[0012] However, the predictive model creation apparatus and predictive apparatus according to the embodiment are not limited to the above-mentioned uses and may be used for purposes other than material design.

[0013] Figure 1 shows an example of the functional configuration of the predictive model creation device in the learning phase and the predictive device in the prediction phase. The predictive model creation device 120 has a learning program installed, and when this program is executed, the predictive model creation device 120 performs the following: Clustering model 121, • Weight definition section 122, • Predictive model 123, It functions as such (see Figure 1(a)).

[0014] The prediction model creation device 120 uses the training dataset 111 stored in the material data storage unit 110 to train the clustering model 121 and the prediction model 123, and generates the trained clustering model 131 and the trained prediction model 132.

[0015] As shown in Figure 1(a), the training dataset 111 includes "input data" and "correct answer data" as information items. In the example in Figure 1(a), "design condition 1" to "design condition n" are stored as "input data," and "characteristic value 1" to "characteristic value n" are stored as "correct answer data."

[0016] The clustering model 121 takes "design conditions 1" to "design conditions n" stored in the "input data" of the training dataset 111 as input, and outputs the training dataset clusters as output data. In other words, when the training dataset 111 is input, a trained clustering model 131 and the training dataset 111 classified into cluster i are generated.

[0017] The number of clusters generated by clustering model 121 is set to N.

[0018] The clustering model 121 that the predictive model creation device 120 learns uses learning methods such as "K-means method, Nearest Neighbor method, hierarchical clustering method, Gaussian mixture method, DBSCAN method, t-SNE method, self-organizing map method," Assume that the model is trained using one or more of the following learning methods.

[0019] More specifically, the clustering model 121 classifies the "design conditions 1" to "design conditions n" stored in the "input data" of the training dataset 111 into one of the clusters i (1 ≤ i ≤ N), and outputs the centroid coordinates of cluster i.

[0020] The weight definition unit 122 defines the weights {W} used in the prediction model 123, using the distance between the classified clusters and parameters representing the features of the training dataset 111. ij} 1≦i≦N,1≦j≦N Calculate.

[0021] Distance between classified clusters {l ij} 1≦i≦N,1≦j≦N This is expressed as the distance between the centroid coordinates mentioned above, and can be calculated in N(N-1) / 2 ways.

[0022] Furthermore, the methods used in the weight definition section 122 to calculate the distance between each cluster include, for example, "Euclidean distance, Manhattan distance, Mahalanobis distance, Minkowski distance, cosine distance, shortest distance method, longest distance method, centroid method, group average method, Ward method, Kullback-Leibler divergence, Jensen-Shannon divergence, Dynamic time warping, Earth mover's distance," The distance can be calculated using one or more combinations of the following:

[0023] The parameters used in the weight definition unit 122 to represent the features of the training dataset 111 can be defined using one or more of the following parameters from "systematic error, standard deviation, variance, coefficient of variation, quantile, kurtosis, and skewness" of "characteristic value 1" to "characteristic value n" stored in the "ground truth data".

[0024] The weights calculated using the distance between the classified clusters and parameters representing the features of the training dataset 111 are represented by a weight function, which can be, for example, "exponential, reciprocal, or power of reciprocal". It is defined using one or more of the following.

[0025] For example, the weight function W ij As for,

[0026]

number

[0027] Here, l ij τ is the distance between each cluster, τ is a parameter representing the features of the training dataset, and α is an arbitrary constant.

[0028] The prediction model 123 is generated by taking the values ​​obtained by multiplying the explanatory variables contained in the training dataset clusters output by the clustering model 121 by the weights calculated in the weight definition unit 122 as input, and training it to output characteristic values ​​as output data, with the target variable (characteristic value) corresponding to the explanatory variables (design conditions) used as input.

[0029] Furthermore, the predictive model that the predictive model creation device 120 learns uses the following learning method: "Random forest, decision tree, gradient boosting, Adaboost, bagging, linear, partial least squares, lasso, linear ridge, elastic net", You may use any one or more combinations of the following.

[0030] Furthermore, when the prediction model creation device 120 trains the prediction model 123, it applies the following logic to the N clusters classified by the clustering model 121: i} 1≦i≦N It is assumed that this will learn. That is, for cluster i, the weight W ij The training was performed using the applied model, and the trained predictive model 132{M i} 1≦i≦N These are each generated.

[0031] One example of a weighted learning method is to input the weights as parameters within the fit function of the random forest regression algorithm stored in scikit-learn.

[0032] As a result, the prediction model creation device 120 generates a trained clustering model 131 and a trained prediction model 132. The prediction model creation device 120 then applies the generated trained clustering model 131 and trained prediction model 132 to the prediction device 130.

[0033] 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 clustering model 131, • 132 pre-trained predictive models, Output section 133, It functions as such (see Figure 1(b)).

[0034] The trained clustering model 131 is generated when the predictive model creation device 120 trains the clustering model 121 using the "design conditions 1" to "design conditions n" stored in the "input data" of the training dataset 111.

[0035] Furthermore, the trained clustering model 131, upon inputting prediction data (design condition x), identifies that the training dataset 111 belongs to cluster p among the N clusters into which it was classified.

[0036] The trained prediction model 132 is generated when the prediction model creation device 120 trains the prediction model 123 using the N clusters into which the training dataset 111 was classified, and the weights calculated by the weight definition unit 122.

[0037] Furthermore, the trained prediction model 132 receives the design condition x and the cluster membership category p output by the trained clustering model as input, and then the trained prediction model 132M corresponding to membership category p is created. p The characteristic value y is predicted using this method, and the output unit 133 outputs the predicted characteristic value as prediction data.

[0038] As a result, the prediction device 130 can obtain sufficient prediction accuracy by using a trained model that has been trained using the cluster to which the design condition x belongs and the weights corresponding to that cluster to predict the characteristic value. In other words, according to this embodiment, the prediction accuracy can be improved in a prediction device that uses a trained prediction model. <Hardware configuration of the predictive model creation device and predictive device> Next, the hardware configurations of the prediction model creation device 120 and the prediction device 130 will be described. Since the prediction model creation device 120 and the prediction device 130 have similar hardware configurations, their hardware configurations will be described together using Figure 2.

[0039] 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.

[0040] 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.

[0041] 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.

[0042] The auxiliary storage device 203 stores various programs and various data used when those programs are executed by the processor 201.

[0043] 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.

[0044] 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.

[0045] 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.

[0046] <Learning Process Flow> Next, we will explain the learning process flow. Figure 3 is a flowchart of the learning process flow.

[0047] In step S301, the predictive model creation device 120 acquires the training dataset 111.

[0048] In step S302, the prediction model creation device 120 uses the acquired training dataset 111 to train the clustering model 121, generating a trained clustering model 131, and obtaining the centroid coordinates between each cluster and the training dataset clusters classified into N clusters.

[0049] In step S303, the weight definition unit 122 defines the distance between the centroids of each cluster {l} for the training dataset cluster i. ij} 1≦i≦N,1≦j≦N Calculate.

[0050] In step S304, the weight definition unit 122 defines the weights {W} to be used in the prediction model 123 using the distance between clusters and parameters representing the features of the training dataset 111. ij} 1≦i≦N,1≦j≦N Calculate.

[0051] In step S305, the prediction model creation device 120 determines whether or not it has calculated weights for all clusters in the training dataset 111, which has been classified into N clusters. If it determines in step S305 that there are clusters for which weights have not been calculated (i.e., if the answer in step S305 is NO), the process returns to step S304.

[0052] On the other hand, if it is determined in step S305 that there are no clusters for which weights have not been calculated (i.e., the answer in step S305 is YES), the process proceeds to step S306.

[0053] In step S306, the prediction model creation device 120 uses the generated training dataset cluster and the corresponding weight combination to train the prediction model 123 and generate the trained prediction model 132.

[0054] In step S307, the prediction model creation device 120 determines whether or not it has trained the prediction model 123 for all clusters of the training dataset 111, which has been classified into N clusters. In step S306, if it is determined that there are clusters for which a trained prediction model 132 has not been generated (if the answer in step S307 is NO), the process returns to step S306.

[0055] On the other hand, if it is determined in step S307 that there are no training dataset clusters that have not generated a trained prediction model 132 (i.e., if the answer in step S007 is YES), the training process is terminated. <Forecasting Process Flow> Next, we will explain the flow of the prediction process. Figure 4 is a flowchart showing the flow of the prediction process.

[0056] In step S401, the prediction device 130 acquires prediction data (design condition x).

[0057] In step S402, the prediction device 130 inputs the acquired prediction data into the trained clustering model 131 and identifies that it belongs to cluster p among the training dataset clusters.

[0058] In step S403, the prediction device 130 uses the trained prediction model 132M corresponding to the identified cluster p. p The system obtains the data and uses the obtained prediction data as input to predict characteristic values.

[0059] In step S404, the prediction device 130 outputs the predicted characteristic values ​​as predicted data for the input data (design condition x) to be predicted.

[0060] The screen 500 in Figure 5 shows a GUI (Graphical User Interface) for generating the clustering model 121, where the user selects whether to manually set the number of clusters N or automatically set it using the elbow method, and selects the type of parameters that represent the characteristics of the training dataset 111. The user selects the optimal number of clusters from the screen and selects parameters such as systematic error, standard deviation, variance, coefficient of variation, quartiles, kurtosis, and skewness. For example, Figure 5 shows an example where automatic setting using the elbow method is selected as the method for setting the number of clusters, and systematic error is selected as the parameter that represents the characteristics of the training dataset 111. When the "Predict" button is pressed in the state shown in Figure 5, the clustering model 121 and weight definition unit 122 of the prediction model creation device 120, and the prediction model 123 generate the trained clustering model 131 and the trained prediction model 132 according to the procedure in the flowchart of Figure 3.

[0061] <Summary> As is clear from the above description, the prediction device 130 according to this embodiment is It includes a pre-trained clustering model 131 for clustering input data and a pre-trained prediction model 132 corresponding to cluster p. The characteristic values ​​predicted by the pre-trained prediction model 132, under appropriate weights, are output as prediction data.

[0062] As a result, according to the prediction device 130 of this embodiment, the prediction accuracy can be improved in the prediction device 130 using the trained prediction model 132.

[0063] [Examples] A specific example of the prediction method of the present invention will be described using a known dataset. Note that the property prediction method according to the present invention is applicable not only to the field of materials science but also to other fields.

[0064] For the purposes of describing the embodiment, it is assumed that the material data storage unit 110 stores, for example, a dataset of 506 data points on Boston housing prices, which is publicly available in scikit-learn's Toy datasets (https: / / scikit-learn.org / stable / datasets / toy_dataset.html).

[0065] When creating a prediction model and performing prediction processing using the Boston housing price dataset, the process is carried out, for example, by following the steps below.

[0066] [Learning Procedure] (1) Step 1 The learning procedure is explained below.

[0067] The Boston housing price dataset was randomly split into a training dataset and a prediction dataset in a 75% / 25% ratio. The explanatory variables used from the Boston housing price dataset were CRIM (crime rate by neighborhood), ZN (proportion of large homes), INDUS (proportion of non-retail businesses by neighborhood), CHAS (proportion of river proximity), NOX (NOx concentration in 0.1 ppm units), RM (average number of rooms per home), AGE (proportion of older homes), DIS (distance to major facilities), RAD (accessibility to major highways), TAX (property tax rate), PTRATIO (student-to-teacher ratio by neighborhood), B (proportion of Black residents by neighborhood), and LSTAT (proportion of low-income residents). The dependent variable was MEDV (median housing price in thousands of dollars).

[0068] (2) Step 2 Using the training dataset obtained in Step 1, we trained the clustering model using the K-Means method, a clustering algorithm included in scikit-learn, to obtain a trained clustering model.

[0069] (3) Step 3 Using the trained clustering model learned in Step 2, the training dataset was input, resulting in N clusters of the training dataset. Here, using the elbow method, two clusters were obtained.

[0070] (4) Step 4 For the clusters of the training dataset classified in step 3, the distance between the centroids of each cluster {l ij} 1≦i≦N,1≦j≦N We calculated N(N-1) / 2 possibilities. Here, we used the Euclidean distance as the distance between the centroids of each cluster.

[0071] (5) Step 5 The distance between clusters calculated in step 4 {l ij} 1≦i≦N,1≦j≦N Using the parameters that represent the characteristics of the training dataset, the weights between the clusters {W ij} 1≦i≦N,1≦j≦N The following was calculated. Here, the standard deviation of the MEDV of the training dataset was used as a parameter representing the characteristics of the training dataset. In addition, the weight function expressed by equation (1) below was used as the weight between clusters. Note that α = 1.0 was used as an arbitrary constant.

[0072]

number

[0073] (6) Step 6 The training dataset clusters generated in step 2 and the weights for each cluster generated in step 5 are used with the random forest regression algorithm stored in scikit-learn as the prediction model, and the prediction model M is applied to each cluster. i We trained the model and obtained two pre-trained predictive models. Here, as a method for applying weights to the training, we input the weights as parameters within the `fit` function of the random forest regression algorithm stored in scikit-learn.

[0074] [Prediction Procedure] (7) Step 7 The prediction procedure will be explained below.

[0075] Prediction data was obtained from the prediction dataset acquired in Step 1. Next, using the trained clustering model learned in Step 2, it was determined whether the prediction data belonged to cluster p among the clusters described in Step 3.

[0076] (8) Step 8 Using the prediction data as input, the trained prediction model M corresponding to the cluster p to which the prediction data belongs, created in step 6, is generated. p The characteristic values ​​were predicted using [a specific method / tool], and the predicted characteristic values ​​were output as predicted data.

[0077] The same process was used to output the prediction data for each remaining prediction data point in the prediction dataset.

[0078] (9) Step 9 The prediction accuracy of the prediction method of the present invention was determined. The prediction accuracy is defined by R as shown in equation (2) below. 2 The value was used as the evaluation metric. 2 The closer the value is to 1, the higher the prediction accuracy.

[0079]

number

[0080] [Comparative Example] On the other hand, as a comparative example, except that clustering in step 2 is not performed and the weights in step 5 are not used, the prediction for creating a predictive model is performed in the same way as in the example, as shown in the flowcharts of Figures 6 and 7, and R 2 The value was calculated.

[0081] The prediction accuracy of the example is R 2 = 0.879 was obtained. On the other hand, the prediction accuracy of the comparative example was R 2 We obtained = 0.868.

[0082] As shown in Figure 8, the prediction accuracy of the random forest regression model that incorporates cluster classification and appropriate weights, as in this example, is higher than that of the comparative example, compared to the case of a standard random forest regression model.

[0083] By classifying the prediction data into appropriate clusters and constructing a model that considers appropriate weights for each cluster, we can make predictions with greater accuracy than in the comparative example.

[0084] [Other embodiments] In each of the embodiments described above, the predictive model creation device and the predictive device were described as separate devices. However, the predictive model creation device and the predictive device may be configured as a single integrated device.

[0085] In the above embodiment, the distance between centroids was calculated using Euclidean distance, and no other specific examples were mentioned. However, the method for calculating the distance between centroids may also be, for example, the Manhattan distance method, Mahalanobis distance method, Minkowski distance method, cosine distance method, shortest distance method, longest distance method, centroid method, group average method, Ward method, Kullback-Leibler divergence, Jensen-Shannon divergence, Dynamic time warping, Earth mover's distance, etc.

[0086] Furthermore, in the above embodiment, the K-Means method and the random forest regression algorithm were used, and specific examples of other learning methods were not mentioned. However, the learning methods used when training the clustering model may also be, for example, the Nearest Neighbor method, hierarchical clustering method, Gaussian mixture method, DBSCAN method, t-SNE method, self-organizing map method, etc.

[0087] On the other hand, the learning methods used when training a predictive model may include, for example, decision trees, gradient boosting, AdaBoost, bagging, linear, partial least squares, lasso, linear ridge, and elastic networks.

[0088] In one embodiment of the present invention, the design conditions of a material whose properties have been predicted by the prediction method of the present invention can also be used in manufacturing. For example, a material manufacturing apparatus can acquire information on the design conditions of a material whose properties have been predicted by the prediction device 130 from the prediction device 130, and use the acquired design condition information to manufacture the material.

[0089] 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. [Explanation of Symbols]

[0090] 111: Training dataset 120: Predictive model creation device 121: Clustering Models 122: Weight Definition Section 123: Predictive Model 130: Prediction device 131: Pre-trained clustering models 132: Pre-trained predictive model 133: Output section

Claims

1. A method by which a computer performs an action, Steps to obtain a training dataset, The steps include generating a trained clustering model using the aforementioned training dataset and clustering model, and classifying the training dataset into N clusters, The steps include calculating the distance between the centroids of each cluster, A step of calculating the weights between the clusters using the distance between the centroids of the clusters and parameters representing the characteristics of the training dataset, For each of the aforementioned clusters, a trained predictive model {M} is created using the cluster and the aforementioned weights. ｉ } １≦ｉ≦Ｎ The steps to generate, Steps to obtain data for prediction, The steps include using the trained clustering model to identify that the prediction data belongs to cluster p among the N classified training dataset clusters, The trained prediction model M corresponding to cluster p takes the aforementioned prediction data as input. ｐ The steps include: obtaining predicted values ​​for the material properties using [this method], A step of obtaining information on the design conditions of the material whose characteristics were predicted, Using the acquired design condition information, the step of manufacturing the material and A method characterized by having the following:

2. The method according to claim 1, A method characterized in that, in the step of constructing the clustering model described in claim 1, one or more clustering methods from among the K-means method, Nearest Neighbor method, hierarchical clustering method, Gaussian mixture method, DBSCAN method, t-SNE method, and self-organizing map method are used.

3. The method according to claim 1, A method for calculating the distance between the centroids of a cluster according to claim 1, characterized in that the distance is calculated using one or more combinations of the following methods: Euclidean distance, Manhattan distance, Mahalanobis distance, Minkowski distance, cosine distance, shortest distance method, longest distance method, centroid method, group average method, Ward method, Kullback-Leibler divergence, Jensen-Shannon divergence, Dynamic time warping, and Earth mover's distance.

4. The method according to claim 1, A method characterized by using one or more parameters from among systematic error, standard deviation, variance, coefficient of variation, quantile, kurtosis, and skewness related to the characteristic values ​​of the training dataset as parameters that represent the characteristics of the training dataset.

5. The method according to claim 1, A method characterized in that, in the step of calculating the weights, one or more weight functions from among exponential, reciprocal, and power-of-reciprocal types are used.

Images