Method for predicting material properties and method for generating models

Abstract

A material property prediction method capable of accurately predicting material properties that vary with temperature is provided. The method includes a model setting step S11 in which a trained model 11 is set, which acquires through machine learning the correspondence between explanatory variables including information related to the material composition or manufacturing conditions of a target material and objective variables including information related to the material properties of the target material, and a prediction step S13 in which explanatory variables related to the target material whose material properties are to be predicted are input to the trained model 11 set in the model setting step S11, objective variables related to the information on the explanatory variables are output, and the material properties of the target material are predicted based on the objective variables. The explanatory variables further include information related to a material property evaluation temperature, which is the temperature at which the material properties included in the objective variables are measured, and the prediction step S13 performs an extrapolation prediction based on the material property evaluation temperature.

Description

Technical Field

[0001] The present disclosure relates to a method for predicting material properties, a program for predicting material properties, a device for predicting material properties, a model generation method, a program for model generation, and a model generation device.

Background Art

[0002] Conventionally, when designing a material composed of a plurality of compositions or a material manufactured by a combination of a plurality of manufacturing conditions, prototypes are repeatedly made while adjusting the composition and manufacturing conditions of the material based on the experience of the material developer, so as to obtain an optimal solution that can realize desired material properties. However, prototyping based on the experience of the material developer often requires repeating prototyping until an optimal design is obtained, which takes time and effort. In addition, condition exploration is often locally performed in the vicinity of the design conditions that the material developer has used in the past, and it is not suitable for exploring global optimal design conditions.

[0003] Therefore, in recent years, there has been an attempt to design materials using a past prototype-evaluation database or to predict material properties by applying machine learning to the database.

[0004] In addition, due to the increasing demand for weight reduction in automobiles, etc., the development of, for example, an aluminum alloy that exhibits high strength even at a high temperature of 150°C or higher is required. Generally, material properties tend to vary depending on the ambient temperature when measuring the properties. For this reason, it is desirable to be able to predict material properties under temperature conditions other than room temperature, such as at a high temperature higher than room temperature or at a low temperature lower than room temperature, using a prediction model based on machine learning.

[0005] However, the prediction accuracy of material properties by machine learning greatly depends on the number of measurement data. Generally, the number of measurement data of material properties at a high temperature that is time-consuming to measure is smaller than the number of measurement data of material properties at room temperature, and it is difficult to improve the prediction accuracy of the high-temperature property prediction model. That is, it is difficult to construct a highly reliable prediction model for designing a material with high strength at high temperature. The same applies to the case of low temperature.

[0006] To improve prediction accuracy even when the amount of data for generating such predictive models is small, there are methods such as constructing separate predictive models for multiple conditions with different output characteristics, constructing predictive models by transfer learning, where a trained model constructed by training on a surrogate output characteristic with a large amount of data is retrained on the target output characteristic for which there is insufficient data, or constructing predictive models by multitask learning, where a neural network or Gaussian process simultaneously learns the output characteristics of multiple different conditions (see Patent Document 1, etc.). [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 2020-95310 [Overview of the project] [Problems that the invention aims to solve]

[0008] However, the conventional methods described above for improving prediction accuracy when the amount of data is small may not yield sufficient prediction accuracy when using a large amount of data measured at room temperature and a small amount of data measured under high or low temperature conditions other than room temperature.

[0009] This disclosure aims to provide a material property prediction method, a material property prediction program, a material property prediction device, a model generation method, a model generation program, and a model generation device that can accurately predict material properties that fluctuate with temperature. [Means for solving the problem]

[0010] This disclosure comprises the following configuration.

[0011] [1] A method for predicting the material properties of a target material, A model setting step involves setting up a trained model in which the correspondence between explanatory variables, which include information relating to the material composition or manufacturing conditions of the target material, and a target variable, which includes information relating to the material properties of the target material, is obtained by machine learning. A prediction step in which explanatory variables relating to the target material whose material properties are to be predicted are input into the trained model set up in the model setting step, a target variable related to the information of the explanatory variables is output, and the material properties of the target material to be predicted are predicted based on the target variable, Includes, The explanatory variable further includes information relating to the material property evaluation temperature, which is the temperature at which the material properties included in the objective variable are measured. The prediction step involves performing an extrapolation prediction based on the material property evaluation temperature. Methods for predicting material properties.

[0012] [2] The model setting step involves obtaining the trained model using a set of explanatory variables and the objective variable, where the material property evaluation temperature is below a predetermined temperature, as training data. The prediction step involves inputting the explanatory variable, in which the material property evaluation temperature is higher than a predetermined temperature, into the trained model to predict the material properties. The material property prediction method described in [1].

[0013] [3] The trained model is a neural network. The material property prediction method described in [1].

[0014] [4] The material properties predicted in the prediction step are properties of a metallic material, polymer material, or glass material that change in an S-shape depending on the material property evaluation temperature, The activation function of the aforementioned neural network is an S-shaped function that includes a hyperbolic tangent function or a sigmoid function. [3] The material property prediction method described below.

[0015] [5] A method is used in which a kernel method is applied as the machine learning technique for the trained model. The material property prediction method described in [1].

[0016] [6] The material property predicted in the prediction step is a property that varies in an S shape according to the material property evaluation temperature in a metal material, a polymer material, or a glass material. An inverse sine function is used as the kernel function of the kernel method. The material property prediction method described in [5].

[0017] [7] A material property prediction program for predicting the material properties of a target material, A model setting function that sets a learned model obtained by machine learning for the correspondence between an explanatory variable including information on the material composition or manufacturing conditions of the target material and an objective variable including information on the material properties of the target material, An explanatory variable related to the target material for which the material properties are to be predicted is input to the learned model set by the model setting function, and an objective variable related to the information of the explanatory variable is output, and the material properties of the target material to be predicted are predicted based on the objective variable. Implemented on a computer, The explanatory variable further includes information related to the material property evaluation temperature, which is the temperature at the time of measuring the material property included in the objective variable. The prediction function performs extrapolation prediction based on the material property evaluation temperature. Material property prediction program.

[0018] [8] A material property prediction device for predicting the material properties of a target material, A learned model obtained by machine learning for the correspondence between an explanatory variable including information on the material composition or manufacturing conditions of the target material and an objective variable including information on the material properties of the target material, An explanatory variable related to the target material for which the material properties are to be predicted is input to the learned model, and information on the objective variable related to the information of the explanatory variable is output, and the material properties of the target material to be predicted are predicted based on the objective variable. Comprising The explanatory variable further includes information related to a material property evaluation temperature, which is the temperature at the time of measuring the material property included in the objective variable. The prediction unit performs extrapolation prediction based on the material property evaluation temperature. Material property prediction device.

[0019] [9] A model generation method for generating a model for predicting the material properties of a target material, comprising: A teacher data creation step of obtaining a teacher data set including information related to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties are measured; Using the teacher data set created in the teacher data creation step, information related to the material composition or manufacturing conditions and the measurement conditions is used as an input to the model, and information related to the material properties is used as an output of the model, and machine learning is performed so that the input-output relationship of the model approaches the input-output relationship of the teacher data set to generate a learned model; a model learning step; including The measurement conditions include a material property evaluation temperature, which is the temperature at the time of measuring the material property included in the output of the model. The teacher data creation step creates the teacher data set so as to deviate from the range of the material property evaluation temperature of the input data to the learned model used for the prediction. Model generation method.

[0020]

[10] A model generation program for generating a model for predicting the material properties of a target material, comprising: A teacher data creation function for obtaining a teacher data set including information related to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties are measured; A model learning function generates a trained model by using the training dataset created by the training data creation function, taking information relating to the material composition or manufacturing conditions and the measurement conditions as inputs to the model, and information relating to the material properties as outputs to the model, and performing machine learning so that the input / output relationship of the model approaches the input / output relationship of the training dataset. To make this a reality on a computer, The measurement conditions include the material property evaluation temperature, which is the temperature at which the material properties included in the output of the model are measured. The training data creation function creates the training dataset such that the input data to the trained model used for prediction falls outside the range of material property evaluation temperatures. Model generation program.

[0021]

[11] A model generation apparatus for generating a model for predicting the material properties of a target material, A training data creation unit creates a training data set that includes information relating to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties were measured. The system includes a model learning unit that uses the training dataset created by the training data creation unit to perform machine learning on the model, taking information relating to the material composition or manufacturing conditions and the measurement conditions as inputs to the model and information relating to the material properties as outputs to the model, so that the input / output relationship of the model approaches the input / output relationship of the training dataset, thereby generating a trained model. The measurement conditions include the material property evaluation temperature, which is the temperature at which the material properties included in the output of the model are measured. The training data creation unit creates the training dataset such that the input data to the trained model used for prediction falls outside the range of material property evaluation temperatures. Model generation device. [Effects of the Invention]

[0022] According to this disclosure, it is possible to provide a material property prediction method, a material property prediction program, a material property prediction device, a model generation method, a model generation program, and a model generation device that can accurately predict material properties that fluctuate with temperature. [Brief explanation of the drawing]

[0023] [Figure 1] This is a functional block diagram of a material property prediction device according to an embodiment. [Figure 2] This is a functional block diagram of the model generation device according to the embodiment. [Figure 3] This is a hardware configuration diagram of a material property prediction device and a model generation device. [Figure 4] This figure shows the temperature dependence of the evaluation of tensile strength in aluminum alloys. [Figure 5] This figure shows the retention time dependence of tensile strength in aluminum alloys. [Figure 6] This figure shows the number of data points for each material property evaluation temperature for aluminum alloys, obtained from a publicly available database. [Figure 7] This is a flowchart of the material property prediction process according to the embodiment. [Figure 8] This is a flowchart of the model generation process according to the embodiment. [Figure 9] This figure shows the training dataset for the prediction model and the validation dataset used for prediction. [Figure 10] This figure shows the configuration of the prediction model used in Example 1. [Figure 11] This figure shows the configuration of the prediction model used in Comparative Example 1. [Figure 12] This figure shows the prediction results for Example 1 and Comparative Example 1. [Figure 13] This figure shows the interpolation prediction results for Examples 4 and 5. [Figure 14] This figure shows the interpolation prediction results for Examples 2 and 3. [Figure 15] This figure shows the extrapolated prediction results for Examples 4 and 5. [Figure 16] This figure shows the extrapolated prediction results for Examples 2 and 3. [Figure 17] These are enlarged views of the vicinity of the plotted locations in each figure of Figure 15. [Figure 18] These are enlarged views of the vicinity of the plotted locations in each figure of Figure 16. [Modes for carrying out the invention]

[0024] The embodiments will be described below with reference to the attached drawings. To facilitate understanding of the explanation, the same reference numerals are used for identical components in each drawing whenever possible, and redundant explanations are omitted.

[0025] Figure 1 is a functional block diagram of a material property prediction device 10 according to an embodiment. The material property prediction device 10 is a device for predicting the material properties of a target material, including a material consisting of multiple compositions or a material manufactured by a combination of multiple manufacturing conditions.

[0026] As shown in Figure 1, the material property prediction device 10 comprises a trained model 11 and a prediction unit 12.

[0027] The trained model 11 is a model that acquires the correspondence between explanatory variables containing information related to the material composition or manufacturing conditions of the target material and a target variable containing information related to the material properties of the target material through machine learning. Based on the input of explanatory variables related to the target material whose material properties are to be predicted, the trained model 11 outputs the target variable related to the input explanatory variables. It is preferable that the trained model is a neural network.

[0028] In this embodiment, the trained model 11 is a three-layer neural network having an input layer, an intermediate layer, and an output layer, as shown in Figure 1. Each node in the input layer of the trained model 11 is provided in the same number as the number of explanatory variables, and a numerical value for each item is input. In Figure 1, as information related to the material composition, a numerical value corresponding to the weight percentage (wt%) of additive elements such as Si and Fe is input to the corresponding node in the input layer. Also, as information related to the manufacturing conditions, a numerical value (e.g., 0 or 1) corresponding to whether or not artificial aging is performed is input to the corresponding node in the input layer. Furthermore, each node in the output layer of the trained model 11 is provided in the same number as the number of objective variables, and a numerical value for each item is output. In Figure 1, as information related to material properties, a numerical value corresponding to tensile strength, 0.2% yield strength, etc. is output from the corresponding node in the output layer.

[0029] Furthermore, as shown in Figure 1, the trained model 11 may have a configuration in which bias nodes are connected to each node of the hidden layer and each node of the output layer. Alternatively, a deep neural network with two or more hidden layers may be used. However, increasing the number of hidden layers to two or more increases the number of parameters, i.e., the number of networks, which tends to overfit the training data and decrease the prediction accuracy on the test data. For this reason, it is preferable to have only one hidden layer.

[0030] Although not shown in Figure 1, when predicting the material properties of aluminum alloys, the explanatory variables can include additive elements such as Si, Fe, Cu, Mn, Mg, Cr, Ni, Zn, Ti, V, Pb, Sn, B, Bi, Zr, Li, B, Ga, and P as material composition. The explanatory variables can also include manufacturing conditions such as heat treatment temperatures and times for annealing, solution treatment, artificial aging, natural aging, stabilization, high-temperature processing, and cold working, as well as processing conditions related to extrusion, forging, drawing, and rolling. The objective variables can include material properties such as tensile strength, 0.2% yield strength, elongation, coefficient of thermal expansion, Young's modulus, fatigue strength, hardness, Poisson's ratio, creep properties, shear strength, specific heat, thermal conductivity, electrical conductivity, and density.

[0031] In this embodiment in particular, the explanatory variables further include information relating to the "material property evaluation temperature," which is the temperature at which the material properties included in the objective variable are measured, and the "evaluation temperature retention time," which is the time the material property evaluation temperature is maintained until the material properties are measured. In other words, the input information of the trained model 11 includes information relating to the material property evaluation temperature and the evaluation temperature retention time, and numerical values ​​corresponding to the material property evaluation temperature and the evaluation temperature retention time are input to the corresponding nodes in the input layer.

[0032] Furthermore, any configuration that includes at least one of the "material property evaluation temperature" and the "evaluation temperature holding time" as explanatory variables is acceptable; a configuration that includes only the "material property evaluation temperature" as an explanatory variable, or a configuration that includes only the "evaluation temperature holding time" as an explanatory variable, is also acceptable.

[0033] The prediction unit 12 inputs explanatory variables related to the target material whose material properties are to be predicted into the trained model 11, outputs an objective variable related to the input explanatory variables, and predicts the material properties of the target material based on the output objective variable.

[0034] The trained model 11 used in the material property prediction device 10 can be generated by the model generation device 20 shown in Figure 2. Figure 2 is a functional block diagram of the model generation device 20 according to this embodiment.

[0035] As shown in Figure 2, the model generation device 20 includes a training data creation unit 21, a model learning unit 22, and a prediction model 11A. The prediction model 11A is the model before machine learning is completed for the trained model 11 shown in Figure 1, and is a model for predicting the material properties of the target material. The configuration of the prediction model 11A is the same as that of the trained model 11 described with reference to Figure 1.

[0036] The training data creation unit 21 creates a training dataset that includes information relating to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties were measured.

[0037] The model learning unit 22 uses the training dataset created by the training data creation unit 21 to train the prediction model 11A and generate a trained model. The model learning unit 22 takes information related to material composition or manufacturing conditions and measurement conditions from the training data as input to the prediction model 11A, and information related to material properties as output to the prediction model 11A, and performs machine learning so that the input / output relationship of the prediction model 11A approaches the input / output relationship of the training dataset, thereby generating a trained model 11.

[0038] The model learning unit 22 can train the prediction model 11A using well-known machine learning methods such as backpropagation, batch gradient descent, stochastic gradient descent, mini-batch gradient descent, and variational Bayesian methods, when the prediction model 11A is a neural network.

[0039] Figure 3 is a hardware configuration diagram of the material property prediction device 10 and the model generation device 20. As shown in Figure 3, the material property prediction device 10 and the model generation device 20 can be physically configured as a computer system including a CPU (Central Processing Unit) 101, a GPU (Graphics Processing Unit) 108, main memory such as RAM (Random Access Memory) 102 and ROM (Read Only Memory) 103, input devices such as a keyboard and mouse 104, output devices such as a display 105, a communication module 106 which is a data transmission and reception device such as a network card, and auxiliary storage devices such as a hard disk 107.

[0040] Each function of the material property prediction device 10 shown in Figure 1 is realized by loading predetermined computer software (material property prediction program) onto hardware such as the CPU 101 and RAM 102, thereby operating the communication module 106, input device 104, and output device 105 under the control of the CPU 101, and reading and writing data to the RAM 102 and auxiliary storage device 107. In other words, by executing the material property prediction program of this embodiment on a computer, the material property prediction device 10 functions as the trained model 11 and prediction unit 12 shown in Figure 1.

[0041] Similarly, each function of the model generation device 20 shown in Figure 2 is realized by loading predetermined computer software (model generation program) onto hardware such as the CPU 101 and RAM 102, thereby operating the communication module 106, input device 104, and output device 105 under the control of the CPU 101, and reading and writing data to the RAM 102 and auxiliary storage device 107. In other words, by executing the model generation program of this embodiment on a computer, the model generation device 20 functions as the training data creation unit 21, the model learning unit 22, and the prediction model 11A shown in Figure 2.

[0042] The material property prediction program and model generation program according to this embodiment are stored, for example, in a storage device provided by a computer. Alternatively, part or all of the material property prediction program and model generation program may be transmitted via a transmission medium such as a communication line and received and recorded (including installation) by a communication module 106 provided by the computer. Furthermore, part or all of the material property prediction program and model generation program may be stored on a portable storage medium such as a CD-ROM, DVD-ROM, or flash memory, and then recorded (including installation) into the computer.

[0043] Similarly, the trained model 11 generated by the model generation device 20 may be stored on a storage medium and carried around, transmitted via a transmission medium, or recorded in a computer.

[0044] Here, it is preferable to use an activation function in the nodes of the intermediate layer of the trained model 11 that has properties common to the evaluation temperature dependence and retention time dependence of the material properties of the objective function. Here, aluminum alloy is given as the target material, and tensile strength is used as an example of the material properties of the objective function. Figure 4 shows the evaluation temperature dependence of tensile strength in aluminum alloy. Figure 5 shows the retention time dependence of tensile strength in aluminum alloy.

[0045] In Figure 4, the horizontal axis represents the material property evaluation temperature (°C), and the vertical axis represents the tensile strength (MPa). Figure 4 plots the tensile strength values ​​measured after the same holding time (10,000 hours) at multiple material property evaluation temperatures, illustrating the characteristic of tensile strength fluctuating according to the material property evaluation temperature (evaluation temperature dependence). As shown in Figure 4, tensile strength tends to decrease as the temperature increases across the entire temperature range, and the rate of change of this decreasing trend becomes relatively large and steep in the range between approximately 100°C and approximately 250°C. Furthermore, this decreasing trend curve has an inflection point at approximately 200°C. In other words, as shown in Figure 4, tensile strength is a characteristic that fluctuates in an S-shape according to the material property evaluation temperature.

[0046] In Figure 5, the horizontal axis represents the evaluation temperature holding time (hr (hours)), and the vertical axis represents the tensile strength (MPa). The horizontal axis in Figure 5 is shown on a logarithmic scale. In Figure 5, the tensile strength values ​​measured after multiple evaluation temperature holding times at the same evaluation temperature (205°C) are plotted, illustrating the characteristic of tensile strength fluctuating according to the evaluation temperature holding time (holding time dependence). As shown in Figure 5, the tensile strength generally tends to decrease as the time increases, and this decreasing curve has an inflection point between 1000 and 10000 hours. In other words, as shown in Figure 5, the tensile strength has an S-shaped characteristic of fluctuating according to the evaluation temperature holding time.

[0047] Thus, if the material properties of the target material predicted by the prediction unit 12 are properties that change in an S-shape depending on the material property evaluation temperature or the evaluation temperature holding time, it is preferable that an S-shaped function including a hyperbolic tangent function or a sigmoid function be used as the activation function of the neural network of the trained model 11.

[0048] Furthermore, a linear (identity function) activation function may be used for the hidden layers of the neural network. It is also preferable to use a linear (identity function) activation function for the output layer.

[0049] Here, the effects of this embodiment will be explained. As shown in Figures 4 and 5, the tensile strength of the material remains stable at a nearly constant value at room temperature (approximately 15-40°C), but tends to fluctuate significantly as the evaluation temperature increases and as the time of holding the evaluation temperature increases. In recent years, due to the increasing need for weight reduction in automobiles, there has been a demand for the development of aluminum alloys that exhibit high strength even at high temperatures of 150°C or higher, for example. It is desirable to be able to accurately predict material properties even under high-temperature conditions where material properties fluctuate significantly.

[0050] However, the accuracy of predicting material properties using machine learning is highly dependent on the number of measurement data points. Generally, the number of measurement data points for material properties at high temperatures, which are time-consuming to measure, is smaller than the number of measurement data points for material properties at room temperature. Figure 6 shows an example of this.

[0051] Figure 6 shows the number of data points for each material property evaluation temperature for a set of material property measurement data for aluminum alloys obtained from a public database. In Figure 6, for 246 data points of 6000 series aluminum alloys with material property evaluation temperatures ranging from -80°C to 371°C, the bar graph shows the number of data points included in each temperature range when the evaluation temperature is divided into 50°C intervals. The data set shown in Figure 6 uses material data collected from public databases: the US MatWeb database (http: / / www.matweb.com / index.aspx) and the Japan Aluminum Association database (http: / / metal.matdb.jp / JAA-DB / ). As can be seen in the example shown in Figure 6, as mentioned above, the number of data points in the temperature range including room temperature from 0 to 50°C is extremely large, while the number of data points on the higher and lower temperature sides of room temperature are generally small.

[0052] Thus, if there are differences in the amount of data that can be obtained depending on the temperature range of the material property evaluation temperature, that is, if there are differences in the amount of data that can be used as training data for machine learning, it is difficult to obtain highly accurate predictive models across the entire range of material property evaluation temperatures using conventional machine learning methods, and it is particularly difficult to improve the prediction accuracy of property prediction models on the high-temperature and low-temperature sides.

[0053] Therefore, in this embodiment, as shown in Figure 1, the trained model 11 is configured to add information related to the material property evaluation temperature and the evaluation temperature holding time as explanatory variables, which are input information.

[0054] In this embodiment, where material properties are the objective variable, the material property evaluation temperature and evaluation temperature retention time are information that would normally be included in the model's output. However, in this embodiment, such material property evaluation temperature and evaluation temperature retention time information is deliberately used as input information for the model. This makes it possible to have the model acquire the correspondence between the material property evaluation temperature and evaluation temperature retention time in the training data and the material properties, enabling the trained model 11 to accurately predict the material properties corresponding to the material property evaluation temperature and evaluation temperature retention time. As a result, the material property prediction device 10 according to this embodiment can accurately predict material properties that fluctuate with temperature.

[0055] By improving the accuracy of material property predictions in this way, the number of prototypes required to design a material with the desired properties can be reduced. Furthermore, by utilizing the generalization ability of the prediction model, material properties can be predicted even under design conditions not listed in the database.

[0056] Furthermore, in this embodiment, if the trained model 11 is a neural network, and the material properties of the target material predicted by the prediction unit 12 are properties that change in an S-shape depending on the material property evaluation temperature or the evaluation temperature holding time, then an S-shaped function including a hyperbolic tangent function or a sigmoid function is used as the activation function of the neural network of the trained model 11.

[0057] This configuration ensures that the activation function shares common characteristics with the temperature dependence and retention time dependence of the objective function's material properties. This improves the function approximation capability of the trained model 11, allowing for more accurate prediction of temperature-dependent material properties. As shown in Figure 4, metallic materials are generally known to exhibit a rapid decrease in strength at a certain temperature. Therefore, using S-shaped activation functions such as the hyperbolic tangent function, sigmoid function, error function, and inverse tangent function, which reproduce this behavior, is considered to contribute to improved prediction accuracy at high temperatures. Consequently, the effect of using an S-shaped activation function to improve prediction accuracy is particularly pronounced in predicting material properties at high temperatures, where the number of measurement data points is generally small or nonexistent.

[0058] Furthermore, the configuration using an S-shaped activation function can also be applied to predict the properties of polymer materials and glass materials whose properties change rapidly at the glass transition temperature, thereby achieving the same improvement in prediction accuracy as in this embodiment. In other words, the target materials for predicting material properties in this embodiment may be polymer materials and glass materials in addition to metals.

[0059] Figure 7 is a flowchart of the material property prediction process according to the embodiment. Each process in the flowchart shown in Figure 7 is performed by the material property prediction device 10.

[0060] In step S11 (model setup step), a pre-trained model 11 is set. The pre-trained model 11 can be one that has been generated in advance by, for example, the model generation device 20, or one that has been obtained by other methods.

[0061] In step S12, the explanatory variables to be used for prediction are set.

[0062] In step S13 (prediction step), the prediction unit 12 inputs the explanatory variables set in step S12 into the trained model 11 set in step S11, and outputs the target variable corresponding to these explanatory variables.

[0063] In step S14, the prediction unit 12 calculates material properties from the target variable output in step S13, according to the explanatory variables set in step S12.

[0064] Figure 8 is a flowchart of the model generation process according to the embodiment. Each process in the flowchart shown in Figure 8 is performed by the model generation device 20.

[0065] In step S21 (training data creation step), the training data creation unit 21 creates training data to be used for model training. The training data may be created by obtaining desired information from an existing public database, for example, as illustrated in Figure 6, or by using data that has actually been measured.

[0066] In step S22 (model learning step), the model learning unit 22 learns the predictive model 11A using the training data created in step S21.

[0067] In step S23, the model learning unit 22 completes the training of the prediction model 11A, and the trained model 11 is completed.

[0068] In this embodiment, a neural network is used as an example of a model for predicting material properties. However, any model that can acquire the input-output relationship between explanatory variables and the target variable through machine learning may be used instead of a neural network. For example, a configuration that predicts high-temperature material properties using a Gaussian process may also be used. In this case, it is preferable to use an inverse sine function as the kernel function. In addition to Gaussian processes, other machine learning methods that apply kernel methods may include kernel ridges and support vector machines. [Examples]

[0069] Next, we will specifically describe embodiments of the present invention.

[0070] <Influence of Model Input Information> Example 1 and Comparative Example 1 were set up as described below, and the effect of adding material property evaluation temperature and evaluation temperature holding time to the input information of the prediction model on the prediction accuracy of material properties was investigated.

[0071] [Example 1] Figure 9 shows the validation datasets used for training and prediction of the prediction model. As shown in Figure 9, 246 validation datasets were created using data obtained from the public database shown in Figure 6, with each dataset associated with material composition (Si, Fe, etc.), manufacturing conditions (annealing, artificial aging, etc.), measurement conditions (material property evaluation temperature and evaluation temperature holding time), and material properties (tensile strength, 0.2% yield stress). As shown in Figure 9, the validation datasets include data for nine different material property evaluation temperatures: room temperature, -80°C, -28°C, 100°C, 149°C, 204°C, 260°C, 316°C, and 371°C.

[0072] Figure 10 shows the configuration of the prediction model 11A used in Example 1. As shown in Figure 10, the prediction model 11A of this embodiment, described with reference to Figure 2, was created, which is a three-layer neural network whose input information includes material property evaluation temperature and evaluation temperature retention time. The explanatory variables input to the input layer of the model include "material composition," "manufacturing conditions," and "measurement conditions" from the verification dataset shown in Figure 9. The number of nodes in the input layer of prediction model 11A is 21. The target variables output from the output layer of the model include two "material properties" (tensile strength and 0.2% yield strength) from the verification dataset shown in Figure 9. The number of nodes in the output layer of prediction model 11A is 2. A hyperbolic tangent function was applied to the activation function of the intermediate layer of prediction model 11A, and a linear function was applied to the activation function of the output layer.

[0073] Using the validation dataset shown in Figure 9, predictions were made for the material property evaluation temperature of the prediction model 11A shown in Figure 10. Of the 246 validation datasets, 63 Category 1 data sets measured material properties at all nine material property evaluation temperatures (room temperature, -80°C, -28°C, 100°C, 149°C, 204°C, 260°C, 316°C, and 371°C) with a temperature holding time of 10,000 hours. The remaining 183 Category 2 data sets either measured material properties at temperatures other than the nine mentioned above, measured material properties at only some of the nine material property evaluation temperatures, or had temperature holding times other than 10,000 hours. Of these, 27 Category 1 data sets were used as prediction data, and a total of 219 data sets (the remaining 36 Category 1 data sets and 183 Category 2 data sets) were used as training data. Then, the prediction model 11A was trained using the training data, and predictions were made using the trained model 11 after training was completed, using the "material composition," "manufacturing conditions," and "measurement conditions" of the prediction data. The batch gradient descent (L-BFGS) method implemented in the Python scikit-learn library was used to train the prediction model 11A. The prediction accuracy was verified by comparing the material properties (predicted values) predicted by the trained model 11 with the values ​​in the "material properties" column of the prediction data (measured values). Prediction accuracy was evaluated using two indicators: mean squared error and mean absolute error.

[0074] [Comparative Example 1] Figure 11 shows the configuration of the prediction model 30 used in Comparative Example 1. As shown in Figure 11, a model was created that excludes the material property evaluation temperature and evaluation temperature holding time from the input information of the prediction model 11A of this embodiment, which was described with reference to Figure 2, i.e., the three-layer neural network. That is, the explanatory variables input to the input layer of the prediction model 30 include only the "material composition" and "manufacturing conditions" of the verification dataset shown in Figure 9, and do not include "measurement conditions". In other words, the number of nodes in the input layer of the prediction model 30 is 19. The target variables output from the output layer of the prediction model 30 include the two "material properties" (tensile strength and 0.2% yield strength) of the verification dataset shown in Figure 9, for nine different material property evaluation temperatures. In other words, the number of nodes in the output layer of the prediction model 30 is 2 × 9 = 18. A hyperbolic tangent function was applied as the activation function of the intermediate layer, and a linear function was applied as the activation function of the output layer. In the neural network structure of Comparative Example 1, learning and prediction can only be used when all (9) sets of output node values ​​are available for each set of input values. Therefore, only Category 1 data can be used for learning and prediction. Here, the same 27 sets of Category 1 data used in Example 1 are used for prediction, and the remaining 36 sets of Category 1 data are used for learning.

[0075] Under the same conditions as in Example 1, predictions for material property evaluation temperatures were made using the verification dataset shown in Figure 9, with the prediction model 30 shown in Figure 11. To verify the prediction accuracy, the output values ​​(predicted values) of two nodes corresponding to the relevant material property evaluation temperature out of the 18 outputs of the prediction model 30 were compared with the values ​​in the "Material Properties" column (measured values) of the prediction data, referring to the "Material Property Evaluation Temperature" in the prediction data.

[0076] Figure 12 shows the prediction results for Example 1 and Comparative Example 1. Figures 12(A) and (B) show the predicted tensile strength and 0.2% yield stress for Example 1, respectively. Figures 12(C) and (D) show the predicted tensile strength and 0.2% yield stress for Comparative Example 1, respectively. In Figures 12(A) to (D), the horizontal axis represents the values ​​(measured values) in the "Material Properties" column of the prediction data, and the vertical axis represents the material properties (predicted values) predicted by the trained model. In Figures 12(A) to (D), the predicted values ​​corresponding to the measured values ​​are plotted for each of the prediction data. The straight line x shown in each figure represents the characteristics when the measured value and the predicted value are the same, and the closer the plot is to this line x, the higher the prediction accuracy.

[0077] Comparing Figures 12(A) and (C), it can be seen that, regarding the tensile strength among the material properties, Example 1 shows less plot variation than Comparative Example 1, and the points are clustered closer to the line x. The mean squared error (RMSE) was 45.3 in Comparative Example 1, while it was 24.5 in Example 1. The mean absolute error (MAE) was 29.6 in Comparative Example 1, while it was 17.3 in Example 1.

[0078] Similarly, comparing Figures 12(B) and (D), it can be seen that, regarding the 0.2% yield strength among the material properties, Example 1 shows less plot variability than Comparative Example 1, and the points are clustered closer to the line x. The mean squared error (RMSE) was 47.6 in Comparative Example 1, while it was 29.3 in Example 1. The mean absolute error (MAE) was 33.7 in Comparative Example 1, while it was 20.8 in Example 1.

[0079] The results shown in Figure 12 demonstrate that adding the material property evaluation temperature and the evaluation temperature holding time to the input information (explanatory variables) of the prediction model improves the accuracy of predicting material properties.

[0080] <Influence of activation function> Examples 2 and 3 and Examples 4 and 5 were set up as described below, and the effect of changing the activation function of the prediction model on the prediction accuracy of material properties was investigated.

[0081] [Example 2] Using the validation dataset shown in Figure 9, interpolation predictions were performed on the material property evaluation temperature of the prediction model 11A shown in Figure 10. Specifically, the 246 validation datasets were randomly divided into 75% (185 sets) of training data and 25% (61 sets) of prediction data. The hyperbolic tangent function was applied as the activation function.

[0082] Furthermore, using the validation dataset shown in Figure 9, extrapolation predictions were performed for the material property evaluation temperature of the prediction model 11A shown in Figure 10. Specifically, of the 246 validation datasets, 185 sets (75%) with material property evaluation temperatures of 150°C or less were used as training data, and 61 sets (25%) with material property evaluation temperatures higher than 150°C were used as prediction data. In this specification, data randomly divided into training data and prediction data is referred to as "interpolation prediction," while data with temperatures of 150°C or less is used as training data and data with temperatures above 150°C is referred to as "extrapolation prediction."

[0083] [Example 3] Except for applying the sigmoid function as the activation function, interpolation and extrapolation predictions were performed in the same manner as in Example 2.

[0084] [Example 4] Except for applying a linear function to the activation function, interpolation and extrapolation predictions were performed in the same manner as in Example 2.

[0085] [Example 5] Interpolation and extrapolation predictions were performed in the same manner as in Example 2, except that a normalized linear unit function was applied as the activation function.

[0086] Figure 13 shows the interpolation prediction results for Examples 4 and 5. Figure 14 shows the interpolation prediction results for Examples 2 and 3. The outlines of Figures 13 and 14 are the same as those of Figure 12. From Figures 13 and 14, regarding the tensile strength among the material properties, the mean squared error (RMSE) was 49.5 for Example 4, 26.8 for Example 5, 24.3 for Example 2, and 24.8 for Example 3. In addition, the mean absolute error (MAE) was 36.4 for Example 4, 18.5 for Example 5, 15.8 for Example 2, and 16.6 for Example 3.

[0087] Furthermore, regarding the 0.2% yield strength among the material properties, the mean squared error (RMSE) was 46.4 in Example 4, 28.2 in Example 5, 26.4 in Example 2, and 26.2 in Example 3. In addition, the mean absolute error (MAE) was 38.4 in Example 4, 18.5 in Example 5, 17.1 in Example 2, and 17.7 in Example 3.

[0088] As shown in Figures 13 and 14, regarding interpolation prediction, Example 4 shows greater plot variability and is further from the line x compared to the others for both tensile strength and 0.2% yield strength, indicating relatively inferior prediction accuracy.

[0089] Figure 15 shows the extrapolated prediction results for Examples 4 and 5. Figure 16 shows the extrapolated prediction results for Examples 2 and 3. Figure 17 is a magnified view of the vicinity of the plotted locations in each figure of Figure 15. Figure 18 is a magnified view of the vicinity of the plotted locations in each figure of Figure 16. The overview of Figures 15 to 18 is the same as in Figure 12.

[0090] As shown in Figures 15 to 18, regarding the tensile strength among the material properties, the mean squared error (RMSE) was 131.5 in Example 4, 41.3 in Example 5, 26.9 in Example 2, and 27.1 in Example 3. In addition, the mean absolute error (MAE) was 117.0 in Example 4, 35.3 in Example 5, 21.8 in Example 2, and 20.8 in Example 3.

[0091] Furthermore, regarding the 0.2% yield strength among the material properties, the mean squared error (RMSE) was 133.3 in Example 4, 44.1 in Example 5, 25.3 in Example 2, and 24.6 in Example 3. In addition, the mean absolute error (MAE) was 117.6 in Example 4, 37.9 in Example 5, 20.9 in Example 2, and 20.4 in Example 3.

[0092] As shown in Figures 15 to 18, the extrapolated predictions show that, for both tensile strength and 0.2% yield strength, Examples 2 and 3 have less plot variability and are closer to the line x compared to Examples 4 and 5, indicating better prediction accuracy. Therefore, considering the results of both interpolation and extrapolation, Examples 2 and 3 can be said to have relatively higher prediction accuracy than Examples 4 and 5.

[0093] Thus, as shown in Figures 13 to 18, the results demonstrate that applying an S-shaped function, including a hyperbolic tangent function or a sigmoid function, as the activation function of the prediction model improves the accuracy of predicting material properties.

[0094] The embodiments have been described above with reference to specific examples. However, this disclosure is not limited to these specific examples. Modifications made to these specific examples by those skilled in the art are also included within the scope of this disclosure, as long as they retain the features of this disclosure. The elements, their arrangement, conditions, shapes, etc., of each of the aforementioned specific examples are not limited to those illustrated and can be modified as appropriate. The elements of each of the aforementioned specific examples can be combined in different ways as appropriate, as long as no technical inconsistencies arise.

[0095] This international application claims priority based on Japanese Patent Application No. 2021-043191, filed on 17 March 2021, and the entire contents of No. 2021-043191 are incorporated herein by reference. [Explanation of Symbols]

[0096] 10. Material property prediction device 11. Pre-trained models 11A Predictive Model 12 Prediction Section 20 Model Generators 21 Training Data Creation Department 22 Model Learning Department

Claims

1. A method for predicting the material properties of a target material, A model setting step involves setting up a trained model in which the correspondence between explanatory variables, which include information relating to the material composition or manufacturing conditions of the target material, and a target variable, which includes information relating to the material properties of the target material, is obtained by machine learning. A prediction step in which explanatory variables relating to the target material whose material properties are to be predicted are input into the trained model set up in the model setting step, a target variable related to the information of the explanatory variables is output, and the material properties of the target material to be predicted are predicted based on the target variable, Includes, The explanatory variable further includes information relating to the material property evaluation temperature, which is the temperature at which the material properties included in the objective variable are measured. The prediction step involves performing an extrapolation prediction based on the material property evaluation temperature. Methods for predicting material properties.

2. The model setting step involves obtaining the trained model using a set of explanatory variables and the target variable, where the material property evaluation temperature is below a predetermined temperature, as training data. The prediction step involves inputting the explanatory variable, in which the material property evaluation temperature is higher than a predetermined temperature, into the trained model to predict the material properties. The method for predicting material properties according to claim 1.

3. The aforementioned trained model is a neural network. The method for predicting material properties according to claim 1.

4. The material properties predicted in the prediction step are those of a metallic material, polymer material, or glass material that change in an S-shape depending on the material property evaluation temperature. The activation function of the aforementioned neural network is an S-shaped function that includes a hyperbolic tangent function or a sigmoid function. The method for predicting material properties according to claim 3.

5. A method that applies a kernel method as the machine learning technique for the aforementioned trained model is used. The method for predicting material properties according to claim 1.

6. The material properties predicted in the prediction step are those of a metallic material, polymer material, or glass material that change in an S-shape depending on the material property evaluation temperature. The kernel function of the aforementioned kernel method is an inverse sine function. The method for predicting material properties according to claim 5.

7. A material property prediction program for predicting the material properties of a target material, A model setting function that sets a trained model obtained by machine learning, which sets the correspondence between explanatory variables, which include information related to the material composition or manufacturing conditions of the target material, and a target variable, which includes information related to the material properties of the target material. A prediction function that inputs explanatory variables relating to the target material whose material properties are to be predicted into the trained model set by the model setting function, outputs an objective variable related to the information of the explanatory variables, and predicts the material properties of the target material based on the objective variable, To make this a reality on a computer, The explanatory variable further includes information relating to the material property evaluation temperature, which is the temperature at which the material properties included in the objective variable are measured. The prediction function performs extrapolation prediction based on the material property evaluation temperature. 、 Material property prediction program.

8. A material property prediction device for predicting the material properties of a target material, A trained model obtained by machine learning that shows the correspondence between explanatory variables, which include information relating to the material composition or manufacturing conditions of the target material, and a target variable, which includes information relating to the material properties of the target material. A prediction unit inputs explanatory variables relating to a target material whose material properties are to be predicted into the pre-trained model, outputs information of a target variable related to the information of the explanatory variables, and predicts the material properties of the target material based on the target variable. Equipped with, The explanatory variable further includes information relating to the material property evaluation temperature, which is the temperature at which the material properties included in the objective variable are measured. The prediction unit performs extrapolation prediction based on the material property evaluation temperature. Material property prediction device.

9. A model generation method for generating a model for predicting the material properties of a target material, A training data creation step involves obtaining a training data set that includes information relating to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties were measured. A model learning step in which machine learning is performed using the training dataset created in the training data creation step, with the information relating to the material composition or manufacturing conditions and the measurement conditions as input to the model, and the information relating to the material properties as output to the model, so that the input / output relationship of the model approaches the input / output relationship of the training dataset, thereby generating a trained model; Includes, The measurement conditions include the material property evaluation temperature, which is the temperature at which the material properties included in the output of the model are measured. The training data creation step involves creating the training dataset such that the input data to the trained model used for prediction falls outside the range of material property evaluation temperatures. Model generation method.

10. A model generation program for generating a model to predict the material properties of a target material, A training data creation function that acquires a training data set containing information relating to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties were measured. A model learning function generates a trained model by using the training dataset created by the training data creation function, taking information relating to the material composition or manufacturing conditions and the measurement conditions as inputs to the model, and information relating to the material properties as outputs to the model, and performing machine learning so that the input / output relationship of the model approaches the input / output relationship of the training dataset. To make this a reality on a computer, The measurement conditions include the material property evaluation temperature, which is the temperature at which the material properties included in the output of the model are measured. The training data creation function creates the training dataset such that the input data to the trained model used for prediction falls outside the range of material property evaluation temperatures. Model generation program.

11. A model generation device for generating a model for predicting the material properties of a target material, A training data creation unit creates a training data set that includes information relating to the material composition or manufacturing conditions of the target material, the material properties, and the measurement conditions when the material properties were measured. A model learning unit generates a trained model by using the training dataset created by the training data creation unit, taking information relating to the material composition or manufacturing conditions and the measurement conditions as input to the model, and information relating to the material properties as output to the model, and performing machine learning so that the input / output relationship of the model approaches the input / output relationship of the training dataset. Equipped with, The measurement conditions include the material property evaluation temperature, which is the temperature at which the material properties included in the output of the model are measured. The training data creation unit creates the training dataset such that the input data to the trained model used for prediction falls outside the range of material property evaluation temperatures. Model generation device.

Images